Posts

AI Brand Visibility: You’re Tracking It Wrong

/0 Comments/in /by

Key Takeaways

  • Most AI brand visibility tracking today replicates keyword tracking logic, using prompts instead of search terms. The underlying assumption is the same, and that’s the problem.
  • Traditional search engines are deterministic: the same query tends to return similar results. LLMs are probabilistic: the same prompt can produce a wide range of valid answers.
  • Measuring a probabilistic system with deterministic tools produces data that looks clean but doesn’t reflect how the system actually behaves.
  • The prompts most brands are tracking (‘Best CRM in 2026,’ ‘Top accounting software’) describe a user who doesn’t exist, someone with no context, no history, and no specific intent. This is a known gap in current AI SEO measurement approaches.
  • Fixing this requires a different measurement philosophy, not just better prompts.

Have you started tracking your brand in ChatGPT, Perplexity, or Google AI Overviews? Good. You’re thinking about the right problem.

Here’s the harder question: what are you actually measuring?

Most teams doing AI brand visibility tracking today have taken a familiar mental model and applied it to an unfamiliar system. Prompts have become the new keywords. Visibility scores have become the new rankings. Tracking platforms have emerged to show how often your brand appears in AI responses over time. On the surface, it looks like a natural evolution of the work you’ve already been doing.

It isn’t.

The tools built for traditional search were designed for a deterministic system, one where the same query reliably returns the same results. Large language models (LLMs) don’t work that way. They’re probabilistic: the same prompt can produce a range of valid answers, shaped by phrasing, context, model version, and more. Applying rank-tracking logic to a system that doesn’t produce ranks is the core mismatch, and it’s quietly corrupting the data most teams are reporting on.

This post breaks down exactly what’s going wrong and what a better approach looks like. It’s the first in a three-part series on AI visibility measurement. Part two introduces a structured framework for building prompts that actually reflect how your buyers use AI. Part three covers what the resulting data reveals about your content strategy.

The Tool The Industry Reached For (and Why It Doesn’t Fit)

The industry’s current approach to AI visibility measurement wasn’t irrational. It was fast. When a new channel emerges, teams reach for the tools and frameworks they already understand, and in digital marketing, that means rankings, share of voice, and tracked keywords. The logic was simple: prompts are the new search queries, so treat them the same way.

The problem is that search engines and LLMs are fundamentally different types of systems.

Traditional search is deterministic. Submit the same query to Google twice and you’ll get a broadly similar set of results. Position may shift slightly, but the system is stable enough that rank tracking works. That predictability is the entire foundation of AI keyword research and traditional SEO measurement.

LLMs are probabilistic. Run the same prompt multiple times and you’ll get a distribution of responses, not a fixed answer. The model generates each response based on statistical associations, not a retrievable index. There is no ‘rank one’ to hold.

The table below illustrates the mismatch. Applying rank-tracking logic to a probabilistic system doesn’t give you a less accurate version of the right answer. It gives you a fundamentally different kind of measurement entirely.

  Traditional Search LLM Ecosystem
System Type Deterministic Probabilistic
Behavior Predictable / Stable Variable / Generative
Core Metric Rank (Position) Presence (Likelihood)
Same query = same result? Broadly yes Not necessarily

This isn’t a minor calibration issue. It’s structural. If you’re reporting on AI visibility using methods designed for predictable, stable systems, you’re building strategy on a foundation that doesn’t reflect how LLMs actually work.

The User Who Doesn’t Exist

The second flaw in current AI visibility tracking is less obvious but equally important.

Most prompt tracking today relies on generic, decontextualized inputs:

  1. ‘Best CRM in 2026’
  2. ‘Top accounting software’
  3. ‘Best project management tool for small teams’

These prompts are clean, scalable, and easy to standardize. They look exactly like the keywords we’ve always tracked.

They also don’t resemble how real people use AI tools.

Real users carry context. They have prior conversations, professional constraints, specific goals, and levels of knowledge that shape what they’re actually asking. A prompt like ‘Best CRM in 2026’ represents an abstract, anonymous user with no history, no constraints, and no intent beyond the words in the query.

A graphic breaking down the differences between abstract users and how actual users use LLMs.

So when you measure AI visibility using these prompts, you’re measuring how the model responds to a hypothetical person who rarely shows up in real decision-making moments. That’s directionally useful at best.

Real audit work bears this out. In one analysis, a brand showed strong visibility for broad category queries, the kind that show up well in standard tracking. But when prompts were shaped around the specific contexts their buyers actually operate in, visibility dropped to zero in the topics most directly connected to purchase decisions. The tracking looked healthy. The actual picture wasn’t.

Generic prompts measure AI visibility for a user who rarely exists. If you want to know how your brand appears to real buyers, you need inputs that reflect real buyer contexts.

The Scaling Trap

The instinctive response to ‘generic prompts aren’t representative’ is volume. If one prompt isn’t enough, run a thousand variations. Add synonyms, modifiers, intent signals, geographic qualifiers. Cover the space more thoroughly.

This logic leads directly into what we call the scaling trap.

Every topic branches into multiple phrasings, intents, personas, and contextual modifiers. The number of prompts required to meaningfully approximate reality grows exponentially. A topic with five main phrasings, three intent signals, and four persona types generates 60 prompt combinations before you’ve added geographic variation or industry context. Scale that across a full content strategy and you’re looking at tens of thousands of prompts, run repeatedly, across multiple models, on a recurring basis.

A graphic explaining the volume fallacy and how prompts properly reflect reality.

Two problems follow. The first is practical: the cost of running this at scale is significant, and it compounds across every client account and every reporting cycle. The second is more fundamental: even after all of that, there’s no guarantee the resulting dataset is meaningfully more representative of actual user behavior. You’ve scaled the volume without fixing the flaw in the input logic.

More prompts don’t fix a representativeness problem. They just make the flawed measurement more expensive.

What Good Measurement Actually Requires

If the problem is that prompts lack context, and brute-force volume doesn’t solve that, the answer is to improve the quality of the input rather than the quantity.

Good measurement of a probabilistic system requires asking a different question entirely. The old question was: ‘Where do we rank?’ The right question is: ‘How reliably does our brand appear when the conditions that actually matter are present?’

That shift has real implications. A brand that appears 85 percent of the time when the right persona and intent conditions are met has a genuinely strong position, even if its average visibility across generic prompts looks modest. A brand that appears 50 percent of the time on generic queries but near zero percent in high-intent, decision-stage contexts has a problem that average tracking completely obscures.

Visibility, measured correctly, is a probability distribution across specific user contexts, not a single score. Getting to that measurement requires inputs that reflect those contexts: structured prompts built around real user personas, specific intent stages, and the actual questions buyers ask when they’re close to a decision.

That’s the foundation of a better approach to AI visibility measurement. The next post in this series walks through exactly how to build it.

Image related to AI Brand Visibility: You’re Tracking It Wrong

In the next post, I’ll walk through the framework we use at NP Digital to build prompts that reflect how real buyers actually engage with AI and what the data looks like when you do it right.

Why This Matters Now

AI-driven search has moved from a future consideration to a present reality, faster than most marketing teams anticipated.

ChatGPT now has over 700 million users, with exponential growth going on. That’s not a niche research tool. That’s a primary discovery channel for a significant and growing share of your buyers.

Image related to AI Brand Visibility: You’re Tracking It Wrong

Google AI Overviews now appear on roughly 48 percent of tracked queries, up 58 percent year over year according to BrightEdge data. In B2B technology, that figure reaches 82 percent of queries. If your buyers research software, services, or professional categories, AI is already shaping what they find before they ever reach your site.

The competitive dynamics are shifting accordingly. Brands that appear consistently in AI responses for the right queries, at the right intent stages, are building an advantage that compounds over time. Brands that don’t appear, or that appear for the wrong queries, are losing ground in the consideration phase before a sales conversation ever starts.

Every week you’re tracking AI visibility with flawed inputs is a week you’re making content and strategy decisions based on data that doesn’t reflect how your buyers actually use AI. The window to get ahead of this is open now.

FAQs

Why should I track AI brand visibility?

Your buyers are already using AI tools to research options, compare solutions, and form opinions about your category. Tracking AI brand visibility tells you whether your brand is present in those moments or invisible. Unlike traditional search, where a low ranking is visible and actionable, AI invisibility is silent, so you won’t know it’s happening unless you measure it.

What Is AI visibility?

AI visibility refers to how often and how favorably your brand appears in responses generated by AI tools like ChatGPT, Perplexity, Google Gemini, and Google AI Overviews. Strong AI visibility means your brand is being surfaced when users ask questions relevant to your product or service.

What are the top AI visibility solutions?

The most widely used platforms for tracking visibility include Writesonic and Profound, alongside a growing number of specialist tools. Each uses a defined prompt set to measure how often your brand appears across major AI platforms. The quality of your prompt set determines the quality of what you can learn — which is exactly the problem I want to address with this series.

Conclusion

Marketers aren’t doing something foolish by tracking AI visibility. They’re doing something natural: applying the tools and mental models they already know to a new channel. The problem is that those tools were built for a deterministic world, and LLMs don’t operate that way.

The mismatch matters. It means the data most teams are reporting on is structurally limited, not wrong exactly, but not representative of what’s actually happening when your buyers use AI to research your category.

The fix starts with a different question. Stop asking where you rank. Start asking how reliably you appear when it actually matters.

In the next post in this series, I walk through a framework built specifically for that question. This is a structured approach to prompt construction that reflects real buyer contexts and makes probabilistic measurement genuinely useful.

Read more at Read More

Bing Webmaster Tools updates AI reporting with Intents, Topics, Citation Share and Compare

/0 Comments/in , /by

Microsoft is now officially releasing a preview of the new AI performance report within Bing Webmaster Tools that now includes Intents, Topics, Citation Share, and Compare. We saw Microsoft demo these features in late April but now it is actually starting to roll out to users.

As a reminder, Bing officially rolled out its AI performance report in February. Google didn’t roll out its AI reporting in Search Console until June, and it seemed forced.

What is new. “These new capabilities build on that foundation by helping publishers better understand why their content is being surfaced, which broader subject areas they are gaining visibility in, how their presence evolves relative to other cited sources, and how citation patterns change over time,” wrote Krishna Madhavan from Microsoft.

Intent: The new Intents feature in Bing Webmaster Tools now classifies the grounding queries in the AI Performance Report in broader categories, such as Informational, Commercial, Navigational, Learn and Solve, Research, Creation, Local, and more. This in a sense helps you understand the intent behind the prompt or query. “This helps publishers move beyond simply seeing which queries triggered citations and begin understanding the broader query context our systems associate with those citation appearances,” Krishna Madhavan wrote.

The example provided was that an e-commerce publisher may discover strong visibility in comparison-oriented or shopping-focused AI experiences, while an educational publisher may find that their content is frequently surfaced in research or learning-oriented interactions. These insights can help publishers better align content structure and depth with the types of experiences where AI systems are surfacing their content.

Topics: The Topics in the AI performance reports group related grounding queries into broader thematic clusters. AI systems reason across concepts and themes rather than isolated keywords, Microsoft explained. So by having topics, it will help publishers understand visibility in the same thematic structure that modern AI systems use to organize information.

So for example, queries such as “solar panels,” “solar energy efficiency,” and “residential solar installation,” for example, may all map into a broader topic cluster like Solar Energy.  “This creates a more natural way to analyze AI visibility. Content teams and publishers often think in terms of themes, editorial areas, and audience interests rather than isolated keywords. Topics help bridge that gap by turning grounding query data into a more thematic view of AI engagement,” Microsoft wrote.

One note, “during the preview phase, some labels may still be broad – especially for highly specialized or niche domains – but the system is already beginning to reveal meaningful thematic patterns,” Microsoft wrote.

Citations. Microsoft also added citation share, which shows how much of the citation space your site receives for a specific grounding query. Citation share is calculated as the percentage of citations attributed to your site out of all citations shown across all sites for that same grounding query. “This helps publishers understand not just whether they were cited, but how much visibility they received within the full set of cited sources for that query,” Microsoft explained.

Microsoft added these points:

  • “This can provide a more directional view into how visibility is evolving over time. Publishers may begin to identify areas where their content has strong and growing representation in AI-generated experiences, as well as areas where visibility may be more fragmented across many sources.”
  • “Importantly, Citation Share is designed as an observational metric – not a ranking system or a competitive scoreboard. It does not expose competitor domains, represent traffic share, or assign quality scores to content.”
  • “AI citation ecosystems are inherently dynamic. Citation patterns can shift due to changes in user behavior, evolving models, freshness signals, partner refresh cycles, and broader changes across the web itself.”

Compare. With all of that, you can also compare the changes over time. The compare feature allows you to overlay a previous time period directly onto the current reporting view. 

“Compare is designed to help publishers observe changes over time. Citation activity can be influenced by many factors including evolving AI models, competing content, freshness signals, and shifts in user demand,” Microsoft wrote.

Here is what it looks like:

Why we care. While we still do not have click and click-through rate data, Microsoft keeps adding more and more to its AI performance reports.

I am hopful that one day we will get click data, but I am still not expecting to see that from Google or Microsoft any time soon.

Read more at Read More

Google penalties: Why prevention is cheaper than recovery

/0 Comments/in , /by
Google penalties- Why prevention is cheaper than recovery

Google penalties, also known as manual spam actions, are among the few events in search that can disrupt an otherwise healthy online business overnight. 

For companies heavily dependent on organic traffic, the consequences often extend far beyond lost rankings. Revenue drops, customer acquisition costs rise, expansion plans stall, and the effects can linger long after the original policy violations have been remedied.

With a steady 90% market share, Google remains the primary traffic source for many publishers, ecommerce platforms, retailers, travel brands, affiliates, and lead generation businesses. 

Direct traffic rarely compensates for a major visibility loss, and Bing seldom offsets the difference. As a result, a manual spam action carries serious operational implications, not merely SEO risks.

Manual actions aren’t algorithm updates

One point still misunderstood throughout the industry deserves clarification. Manual spam actions differ from algorithmic updates. They aren’t fluctuations caused by changes in relevance calculations or ranking system adjustments. 

Google’s manual penalties involve direct enforcement after suspected violations against Google Search Essentials, formerly Google Webmaster Guidelines, have been identified and confirmed. The distinction matters because the response required is completely different.

A website affected by changing ranking systems requires analysis, adaptation, and recrawling. A website affected by a manual spam action requires remediation and applying for reconsideration. Those are separate situations entirely.

Google doesn’t issue manual spam actions casually. The process involves internal senior employee review cycles. Suspected violations must be investigated and confirmed first.

Google states clearly that manual actions are the consequence of proven policy transgressions. Despite frequent cries of foul, false positives are exceptionally rare. Once a manual action appears in Google Search Console, the enforcement is already in the production pipeline.

The operational problem is that many businesses fail to recognize how much unresolved policy exposure their web platforms have accumulated over time.

Be the brand customers find first.

Track, grow, and measure your visibility across Google, AI search, social, local, and every channel that influences buying decisions.

Start your free trial

How penalties develop

The initial steps that ultimately lead to a manual penalty and a website’s drop in search visibility often begin inconspicuously, gradually eroding policy compliance. 

  • An ecommerce business launches an aggressive link acquisition campaign during an early growth phase. Over the years, PageRank-passing spam links accumulate unchecked until eventually nobody remembers where thousands of exact-match backlinks originate. 
  • A publisher enters into commercial partnerships involving sponsored content or affiliate sections, which gradually become structurally integrated into the editorial architecture of the website. 
  • A SaaS company creates large numbers of low-quality location pages while expanding into new markets. 
  • A lead generation business scales supplemental SEO content through low-cost LLM production systems with limited editorial oversight because that appears to be what most competitors are doing.

The underlying patterns are remarkably similar across industries. In many cases, organic search visibility initially improves and may even generate measurable revenue gains attributable to the SEO initiative. 

The short-term results reinforce the perception that the approach is working. However, as time passes, nobody revisits whether those earlier decisions remain aligned with evolving search quality standards and webmaster policies.

Why historical violations still matter

One reason manual spam actions create so much disruption is that policy violations often persist quietly for years before review. Many organizations incorrectly assume that questionable SEO tactics of the past lose their relevance over time. 

Yet Google Search systems don’t forget historical footprints. Archived URLs remain crawlable. Legacy sections continue contributing content quality signals long after internal ownership was abandoned. 

Most persistently, backlink patterns remain visible for decades. Large numbers of websites remain affected by backlinks generated through manipulative campaigns dating back many years. 

Paid placements, article syndication networks, private blog networks, commercial keyword-heavy guest posting campaigns, expired domain backlinks, directory spam, and widget distribution schemes that once formed part of mainstream SEO activity are today’s liabilities. 

Some of these practices continue to operate more or less openly for years, while enforcement may appear erratic or inconsistent. When left unaddressed, they represent an incalculable risk to the website publisher.

Get the newsletter search marketers rely on.

See terms.

Time doesn’t eliminate the risk

This becomes particularly important during acquisitions. Businesses purchasing established domains frequently inherit unresolved compliance exposure alongside rankings and traffic. Google evaluates the website’s condition, not which employee, agency, or previous owner introduced the violations.

Traffic growth alone doesn’t confirm compliance health. A domain generating millions of clicks may still carry unresolved risks tied to old link schemes, expired sponsorship arrangements, deceptive user-agent cloaking, manipulative redirects, or scaled low-quality content sections. Those issues often go unnoticed until they’re brought to the surface by a Google manual spam action notification.

A common sign of a manual spam action: rapid loss of visibility
A common sign of an algorithmic adjustment: Gradual loss of visibility

Reputation abuse and publisher liability

The mechanics behind reputation abuse are straightforward. A trusted brand with an established web platform allows third parties to publish unrelated, often unsupervised content under the same domain name. In many cases, publishers integrated discount coupon sections, casino reviews, affiliate content, or commercially motivated informational pages directly into existing editorial systems. 

The problem frequently worsened significantly because the content wasn’t properly segmented. The consequence is that the distinction between trusted editorial work and commercially motivated material became blurred.

Once confronted with a site-wide penalty, affected publishers experience broad visibility declines across the entire platform, not merely within the originally offending sections of the website. The damage to a brand that lends its reputation to a disreputable third party is often substantial. 

Recovery efforts frequently prove time-consuming and costly. Removing isolated pages rarely resolves the problem. Many organizations require broader structural changes, including archive cleanup, internal link reviews, crawl management adjustments, sponsorship governance reforms, the removal of spammy redirects, stronger editorial oversight, and stricter technical segmentation.

In short, recovering from such a penalty takes time, costs significant amounts of money, and is often a painful process.

A common sign of a manual spam action: Rapid loss of visibility
A common sign of a manual spam action: Rapid loss of visibility

The risks of scaled content

Google increasingly scrutinizes large-scale publishing systems that produce repetitive, low-value content without a unique selling proposition. 

The issue isn’t maintaining many websites simultaneously. Large website portfolios have thrived in Google Search for years and continue to do so. The underlying problem involves quality control, editorial oversight, originality, and informational value.

  • Affiliate networks produce near-identical product comparison pages across thousands of long-tail keywords. 
  • Local SEO operations deploy templated service pages across hundreds of regions with minimal differentiation. 
  • AI-assisted workflows publish large numbers of informational pages without factual oversight or genuine expertise to support them. 
  • Travel websites generate mass-produced destination pages through repetitive, generic content systems.

Most organizations don’t cross into problematic territory intentionally. The transition usually occurs gradually, often unbeknownst to the decision-makers who rely on outdated or misleading recommendations. 

The resulting manual spam action in Google Search Console, followed by a sharp decline in rankings, frequently occurs after a prolonged period of spam signal accumulation rather than during the apparent growth phase.

Incomplete remediation prolongs penalties

Many site owners approach reconsideration requests as if they were negotiating with Google. That puts them at a significant disadvantage from the outset. 

The reconsideration process exists for one purpose only: to demonstrate that the website has been restored to full compliance with Google’s guidelines. It’s important to note that Google expects complete compliance before lifting a manual spam action.

This means the requirement extends beyond the specific violation highlighted in Google Search Console. A site owner who addresses only one known spam issue while leaving unrelated policy violations unresolved elsewhere on the website will typically face rejection.

A common testing approach, such as a publisher removing some problematic sponsored content while retaining similar affiliate arrangements elsewhere, will fail. Likewise, a business that disavows recent manipulative backlinks while ignoring historical paid link schemes is unlikely to convince Google of its genuine commitment to complying with Google’s policies going forward.

Similarly, a website network that cleans up one property while continuing identical publishing practices across related domains signals incomplete remediation rather than meaningful operational reform. As a result, it stands little chance of regaining Google’s trust.

Why repeated rejections make recovery harder

Effective website recovery requires a comprehensive review rather than selective cleanup. Technical infrastructure, content quality, sponsorship structures, redirect behavior, link acquisition history, indexing patterns, archive sections, and ownership transparency all require examination during serious compliance recovery efforts.

The Google Search team expects compelling documentation detailing what has changed and how future violations will be prevented. Temporary cosmetic adjustments rarely persuade reviewers to lift a manual spam action.

Making matters worse, each rejection typically requires an even more comprehensive review and cleanup effort. At the same time, every reconsideration request that Google deems disingenuous further erodes Google’s trust in the publisher.

The cost of uncertainty

There’s no guaranteed turnaround time for reconsideration processing. Some reviews are completed within days. Others take weeks or months.

At the same time, large websites with extensive SEO legacies accumulated over many years often require longer assessment periods due to the substantial volumes of data that must be crawled and analyzed before changes can be evaluated.

For businesses that rely primarily on Google traffic, this uncertainty creates a potentially existential threat.

  • An ecommerce business approaching a peak seasonal period with an unresolved manual spam action can face cash flow problems quickly.
  • Publishers dependent on advertising revenue experience ranking losses that translate directly into declining commercial performance.
  • Lead generation businesses often encounter immediate pipeline contraction once visibility declines significantly.

The operational risk becomes even greater when companies fail to build a strong brand capable of partially offsetting organic traffic declines through direct navigation or alternative revenue-generating channels. In this context, paid traffic is a poor substitute due to its associated costs.

In short, some online businesses can’t afford to be penalized in the first place.

Penalties can cripple operations

The issue extends beyond SEO performance. Search visibility directly affects commercial expansion, investor confidence, company valuation, partnership negotiations, and revenue stability.

Penalty expiration represents another commonly misunderstood aspect. Google manual spam actions may expire after prolonged periods, often years. However, this is rarely a viable strategy for an affected business. 

Waiting passively through an extended period of declining visibility seldom aligns with commercial realities. More importantly, expiration alone doesn’t guarantee recovery or renewed growth, as the penalty could be reapplied not too long after it expired.

Google’s search systems continue evaluating overall site quality independently of manual enforcement status. A website carrying unresolved spam signals across its content, technical infrastructure, or off-page profile may continue to struggle long after the manual action itself has been lifted.

Compliance requires ongoing oversight

Compliance reviews can’t be considered optional or a luxury. Organizations heavily dependent on organic Google visibility require ongoing operational review cycles focused specifically on comprehensive policy compliance.

These reviews shouldn’t be conducted internally. Even the most talented in-house SEO teams are often hard-pressed to diligently identify shortcomings that may reflect on their own work or that of their colleagues. Policy compliance requires external expertise, sufficient authority, and a proven track record.

Purely technical SEO audits, while indispensable, are insufficient if commercial partnerships bypass oversight. Editorial standards alone won’t suffice if historical link manipulation remains unresolved. Planned growth initiatives require evaluation against established compliance frameworks before deployment, not after traffic has become dependent on questionable practices.

Mature organizations increasingly integrate compliance reviews into their operational governance. Sponsorship structures undergo search compliance review before launch. Scaled publishing systems are assessed for quality before expansion. Historical content is evaluated on a recurring basis. Acquisition due diligence includes policy exposure analysis alongside financial review.

Own the conversation before your competitors.

See where your brand appears, where it doesn’t, and exactly how to win more visibility across search, AI, local, social, and every channel that matters.

Start your free trial

Compliance is a business imperative

This level of discipline and vigilance matters because manual spam actions rarely arrive at convenient moments. More often than not, undesirable Google scrutiny coincides with critical periods: just before a long-planned commercial expansion, in the run-up to a migration project, ahead of an acquisition, as the peak retail season begins, or shortly before investor reporting deadlines.

This is hardly intentional. It’s simply a matter of unfortunate timing. Google doesn’t align search quality enforcement with business planning calendars. Google cares primarily about user experience. For every website that loses its top position, there is usually another capable of providing users with a similarly compelling experience.

Businesses that ignore unresolved policy exposure often discover the problem the hard way, only after search visibility has collapsed and online sales have followed suit. At that point, recovery becomes a far more prolonged, expensive, and operationally disruptive undertaking than ongoing compliance reviews would have been prior to penalization.

Nevertheless, the work must be done. The one silver lining is that, in many cases, the process proves cathartic. Once the penalty has been resolved and the website’s SEO signals have become more consistent, the removal of legacy issues often allows rankings not merely to recover, but to exceed their previous highs.

Read more at Read More

How AI is merging paid and organic visibility

/0 Comments/in , /by
How AI is merging paid and organic visibility

The idea that AI is killing advertising misses the bigger shift. As AI expands across search, assistants, productivity tools, and transactions, advertising is moving with it.

Ad density may be changing within AI experiences, but advertising opportunities are expanding across a growing number of surfaces.

At the same time, paid and organic are becoming harder to separate. The same AI systems increasingly power ad campaigns, search experiences, and brand visibility across Google’s ecosystem.

That changes how brands should think about visibility.

Paid and organic are no longer separate channels competing for the same click. They are increasingly different ways of influencing the same AI systems, which means the signals shaping organic visibility may also affect paid performance.

The old model: Paid and organic on one finite SERP

Google’s SERP was a finite surface: 10 organic blue links, a few ad slots, and a knowledge panel on the right. The user landed, scanned, and clicked.

Paid and organic teams operated on separate budgets, separate tools, and separate quarterly reports, and rarely talked to each other because manual Google Ads kept the paid specialist busy full time. Titles, descriptions, bids, and campaign structure were all chosen by hand and required constant attention, which is why the organic team had no part in any of it.

DSA changed that for me. It read my organic pages to decide which ads to run, who to show them to, when, at what bid, and what title to use. I controlled the descriptions. The engine decided everything else, and it did it better than I would’ve done manually because it was reading the same signals the organic side was already optimizing for.

When someone at Google in Singapore explained how PMax worked, I thought, “That’s exactly what I was doing.”

PMax took the DSA logic and extended it across every Google surface simultaneously: Search, YouTube, Gmail, Display, Maps, and Shopping, all in one campaign, with the engine making every placement decision from your assets and audience signals.

AI Max brought the same intelligence into Search campaigns, specifically, with Gemini underneath instead of rules. PMax and AI Max run on the same Gemini brain: one focused on Search, the other spread across every surface, applying the same funnel logic to different contexts with different signal layers on top.

And if Gemini’s understanding of your brand is thin, it fills those decisions with whatever it thinks will work, which isn’t necessarily your brand narrative, and you have no direct way to override it. You train it, or you lose control of your own ads.

Be the brand AI recommends.

See where your brand appears in AI search, where competitors are winning, and what it takes to become the answer AI recommends.

See your AI visibility

The new model: Gemini sits inside every surface, and it carries ads with it

Gemini now sits inside every layer of the Google ecosystem: 

  • Discovery (Search, Maps, YouTube, Lens, News, Discover, and Shopping), productivity (Gmail, Docs, Drive, Photos, and Calendar).
  • Distribution (Android, Chrome, Google Play, Pixel, Wear OS, Google TV, and Nest).
  • Transaction (Google Pay, Wallet, Flights, Hotels, and Travel).
  • Assistive surfaces themselves (AI Mode, AI Overviews, Assistant, NotebookLM, and the Gemini app). 

That’s how many connected consumers spend most of their workday, and most of those surfaces either carry ads now or have the infrastructure to start carrying them.

Microsoft Advertising sits inside Copilot across Bing, Edge, Windows Consumer, Office Consumer, Teams Free, and GitHub. 

OpenAI Ads launched in February for logged-in users on Free and Go tiers in the U.S., placing ads below ChatGPT responses and clearly labeling them as sponsored. By May, OpenAI had opened a self-serve Ads Manager and was expanding internationally.

The ads layer travels with the engine, the engine is everywhere, and ads therefore have the potential to be everywhere. Most brands still treat paid as a separate channel run by a separate team on a separate dashboard, which is a search-era inheritance that was never ideal but now needs to be dropped. 

Performance Max already runs the auction across YouTube, Display, Search, Discover, Gmail, and Maps as one campaign type. Search is one surface among many, and the “ads are dying in AI search” narrative is measuring the wrong thing. It sees ad slots compress inside the assistive interface while ignoring that the surface base has multiplied by an order of magnitude.

Ad density follows the delegation the user has made to the machine

The dominant narrative in 2026 is that ads are dying because AI is replacing search, and ads inside AI are a problem nobody has fully solved yet. That’s partially correct: Ad density per session drops as AI takes more control, and nobody – including Google – has yet figured out how to insert ads into the AI response itself without killing the experience that makes the AI valuable in the first place.

But this is the part the analysis gets wrong: This doesn’t add up to fewer ads overall.

Search ads are Google’s goose with the golden egg, and the goose may be slowing down — though nobody outside Google actually knows, because Google doesn’t break out search ad revenue from YouTube, Display, and the rest. That ambiguity is doing a lot of work.

What we do know is that total ad revenue has kept growing even as AI has taken over more of the search experience, which proves the flock is already working.

Kodak invented the digital camera and then buried it to protect film-processing revenue, and we know how that ended. Google appears to be doing what Kodak didn’t: building the replacement while the original is still profitable.

Every surface Gemini sits inside is a new bird in the flock, each laying a smaller egg that grows over time, and when Google finally cracks ads inside the AI response itself, that’s one more goose. The surface base has expanded faster than density has dropped, and the ad-density problem in Search and AI is temporary.

The more the user delegates decisions to the machine, the less room the machine has to surface a paid option. Search keeps the user in charge, so the engine surfaces ads the user might pick. Assistive narrows the options, so a sponsored slot still has a chance. Agentic executes the decision, so the ad has nobody to persuade. Ad density follows that delegation, mode by mode, with AI deciding which brands win at each mode.

Ad density follows the delegation the user makes to the machine
Ad density follows the delegation the user makes to the machine.

Google is running two moves at once, and it seems most people have noticed only the first one. Gemini is taking over the recommendation, targeting, and auction logic on surfaces that have carried ads for years. And Google is adding ads to surfaces where they were previously absent, with AI Overviews now eligible for ads above, below, and within the answer, and AI Mode testing conversational ad formats.

The first move is AI taking over the existing ad business. The second is the ad business expanding into surfaces it never occupied. The net effect is more AI-driven ads across more of the stack than ever before.

The freemium system still works, but the ad is becoming part of the surface

The monetization model that works at consumer internet scale is simple: pay with money, or pay with attention.

  • YouTube is Google’s clearest example — and proof that it works: free with ads, paid without, and the vast majority of users have always chosen ads. 
  • Gmail draws the same line: Where the user pays directly, Google doesn’t insert ads. Where the user pays with attention, Google monetizes it.

I learned about freemium the hard way. When our children’s media company, Boowa & Kwala, survived the dot-com crash, we added a paid tier that removed the ads. Out of a million unique visitors a month, a few hundred paid. Almost nobody chose to pay. 

The freemium contract — free access in exchange for ads — is the deal they actively prefer, and the numbers prove it. And for ad-driven businesses, pure volume makes the money. In Big Tech, Google has the clear advantage.

  • ChatGPT is already running ads on free tiers. 
  • Gemini is ad-free without login, but that’s a launch state, not a permanent model. 
  • Perplexity is blocking users instead of monetizing them, which is a different bet on the same problem — and a bet with a limited runway. 

Every AI surface is in the process of landing on the same answer because there is no other answer.

What changes is how the ads appear. The classic SERP ad was clearly labeled and set off in a colored panel. The Gemini recommendation that surfaces a product inside a Gmail context, the Copilot suggestion that names a vendor inside a Word document, and the agent that picks a supplier on the user’s behalf are something else entirely. 

The ad becomes ambient. It dissolves into the surface, and what advertising looks like becomes harder to identify as advertising. Gemini reads context and intent with enough precision that an ad placed in a meeting summary can feel useful rather than disruptive, which is a risk profile Google’s rules-based systems could never have accepted.

At Boowa & Kwala, when we scaled free ad-supported views from 100 million to 1 billion, revenue multiplied by roughly two, and costs rose by around 20%. Surface (a.k.a. pageviews) multiplied tenfold, revenue doubled, costs grew by a fifth, and we went from profitable to significantly more profitable. 

The aim was never to push revenue up at the same rate as surface expansion. It was to keep expanding the surface, knowing the incremental delivery cost was negligible. 

Google’s ratios at planetary scale differ from ours, but the structural shape almost certainly doesn’t: surface expansion plus near-zero incremental cost equals profit growth, regardless of whether revenue per surface keeps pace.

Get the newsletter search marketers rely on.

See terms.

Cohort, intent, and profit drive both paid and organic

PMax, AI Max, AI Overviews, AI Mode — Gemini is driving all of them. The AI optimizing your paid campaigns is the same AI evaluating your organic content, reading the same user, in the same moment, with the same intent.

The engine reads three signals: 

  • Cohort.
  • Intent.
  • Profit. 

In paid, you declare all three explicitly when you structure your campaigns. In organic, the engine infers all three from behavior: clicks, dwell time, and return-to-search serve as proxies for the profit signal that is missing there. Google denied using behavioral signals for years. Its own court case documentation told a different story.

Which means the organic discipline the whole series has been building — the funnel query pathway, the entity home, and the corroboration stack — has always been pointing at one thing: engineer the page so precisely for the right cohort that the behavioral signal does the same job as a correctly structured PMax campaign. The user lands, stays, converts, and doesn’t go back and research the same thing again. Google reads that behavior and infers your profit tier.

My bet, and I want to be clear it’s a bet rather than a documented fact, is that Gemini can’t serve a paid ad in real time without grounding against current search results because the ad has to match the organic context it’s appearing in. 

If it doesn’t ground, the ad is inconsistent with what the user sees organically, which breaks the experience and loses the click. So the grounding process for paid is the same process as for organic: same knowledge graph, same search index, same LLM. 

That means training Gemini on your brand through organic improves your paid performance through the same mechanism. One training investment, two outputs. I’ll be proven right on this eventually, and this article is the timestamp.

The same AI runs your organic and your paid. Train it once, win twice.
The same AI runs your organic and your paid. Train it once, win twice.

You can’t directly target Gemini in AI surfaces. You can only train it.

Across AI-driven placements, Gemini decides everything: where to show your ad, what to show, how to show it, who to show it to, when, and at what bid. The advertiser feeds it information and sets the parameters, but Gemini makes every decision that matters.

What you’re buying when you spend on Google Ads in 2026 is the right to feed a recommendation system that analyzes your brand on its own terms. The explicit signals you declare in paid — cohort, intent, and profit — are a real advantage over organic, where the engine has to infer all three from behavior. 

But your ability to dominate through pure campaign structure is vastly reduced when Gemini doesn’t understand or trust your brand. The control has shifted: you guide it through signal clarity, not through the settings dashboard, and that guidance works best when your organic foundation is solid.

Use paid to find the combinations that work, build organic pages around them

In a correctly structured PMax or AI Max campaign, you declare cohort, intent, and profit margin explicitly: this audience, this goal, this margin, in the same campaign. You don’t mix a luxury hotel and a budget guesthouse in the same ad group because the cohort is different, the profit margin is different, and handing the engine a mixed signal makes it spend your budget resolving a contradiction you created.

Organic doesn’t let you declare profit directly. The engine infers it from who landed, who stayed, who converted, and who never came back to search for the same thing. That behavioral signal is the only proxy it has for the profit tier, and it’s a thin signal compared to the explicit declaration you make in paid.

The smartest move for any brand running both is to treat them as a single loop. Run paid to find which cohort-intent-profit combinations actually convert. Build the organic pages around those combinations, designed so precisely for the right cohort that the behavior on the page sends the engine the same signal the paid campaign explicitly declared.

The paid shortcut in the funnel

The paid side becomes cheaper because organic pages provide the behavioral confirmation the engine needs. The organic side gets stronger because the paid data tells you exactly which pages to build and for whom, and then feeds the engine the same signal the paid campaign declared explicitly, for free.

Most travel sites serve the same page template to a budget traveler looking for a €30 guesthouse in Bangkok and a wealthy traveler looking for a €3,000 suite at the Peninsula. Same layout, same fields, same photo grid, same review format. 

The engine has to infer which cohort the page serves mostly from behavior because the differentiation of the pages is limited. Build the page for the person rather than the query, and you hand the engine the cohort signal it’s currently having to guess. That’s not a UX decision. That’s your profit margin declaration to an engine that can’t see your margins any other way.

And you win on all three fronts simultaneously. A page built precisely for the right person converts better because it works better for the human.

Better conversion behavior sends cleaner implicit signals to the engine, which improves your organic ranking for that cohort. And cleaner organic signals reduce your paid CPC because the engine has less to guess about. Better pages, more organic, cheaper paid – the same work produces all three.

When Gemini isn’t convinced about you, you pay on both sides simultaneously

The three revenue taxes — the doubt tax, the ghost tax, and the invisibility tax — operate on the organic side. Because the engine powering your organic results is the same one powering your paid placements, you pay all three on both sides simultaneously.

  • The doubt tax: When the engine hedges on basic facts about you organically, it rewrites your paid creative to soften the same claims.
  • The ghost tax: When the engine prefers competitors in organic comparisons, your paid creative gets passed over even when your bid is competitive.
  • The invisibility tax: When the engine doesn’t surface you organically, it doesn’t show your ad either. You’re not in the running.

Paid surfaces carry two additional taxes that don’t exist on the organic side, and one discount you earn when you get it right.

The taxes and discounts in AI-driven paid search

The taxes and discounts in AI-driven paid search include:

  • The mistrust tax: What you pay when the engine’s confidence in your brand is low. A CPC premium because Quality Score penalizes low entity trust, and message distortion because the Gemini Filter rewrites your creative away from your intended positioning. You can’t turn the filter off. The practical answer isn’t constraining it. It’s improving the entity confidence that the engine reads when deciding how to filter.
  • The intent tax: This is self-inflicted. Build an ad group with mixed intent, and you hand the engine a contradiction. Gemini will spend your money figuring out a mess you made. Each ad group should align on cohort, intent, and profit margin — any mix across those three, and Gemini is billing you to resolve the confusion.
  • The confidence discount: This is the blade cutting the other way. Every properly defined ad group is secretly doing two jobs: it buys you an efficient placement today, and it teaches the engine which cohort you serve tomorrow. When the engine trusts you, it stops second-guessing your ads, your CPC drops, and your creative lands cleaner. That’s worth more than any bid adjustment you make.

If AI can’t find you, customers won’t either.

Track your visibility across AI search, uncover missed opportunities, and grow your presence where customers are asking questions.

See your AI visibility

Google has a structural advantage that Microsoft and OpenAI can’t match

Google has all the cards: the model, the surfaces, and the ads platform, all owned and tuned together in absolute harmony. Microsoft has the surfaces but lacks the LLM to drive them at the same level. 

OpenAI has the model and launched a real ads business in February 2026, but lacks the surfaces – no Gmail, no YouTube, no Maps, no Play – and without surfaces, an ads business can’t compound at scale. Only Google has all three working as one system.

Paid and organic are now inseparable. The goose is fading, but Google can afford to let it. They know it rises like a phoenix, and in the meantime, they’ve got the biggest gaggle.

This is the 18th piece in my AI authority series.

Read more at Read More

How AI helped build hreflang XML sitemaps at scale

/0 Comments/in , /by
How AI helped build hreflang XML sitemaps at scale

As AI tool usage has become more common, I’ve seen impressive examples of people building tools to automate complex processes that once required significant manual effort. I’ve also seen teams adopt AI simply because it’s available, often with little practical benefit.

My approach is to focus on AI applications that save time and solve real problems.

Recently, I needed to align the SEO architecture for more than a dozen websites across three separate businesses, eight regional domains, and multiple languages, including three English dialects, Italian, Japanese, Spanish, Thai, French, and Korean.

Historically, mapping thousands of URLs to create cohesive hreflang XML sitemaps would have required specialized software or days of spreadsheet work. Instead, I used Google Gemini to build a custom Python script that handled the heavy lifting.

Here’s how the project evolved from an initial prompt into a highly customized automation tool, and what it taught me about using AI for technical SEO.

Where AI delivers the most value

I use AI primarily for practical, time-saving tasks, including:

  • Generating regex patterns when I need a quick solution without researching syntax from scratch.
  • Creating complex spreadsheet formulas for reporting workflows that rely on manual data exports.
  • Accelerating research and planning for projects that require competitive analysis across multiple business lines.
  • Building custom automation tools for recurring SEO and data-processing tasks.

The hreflang project discussed here falls into that final category.

Be the brand customers find first.

Track, grow, and measure your visibility across Google, AI search, social, local, and every channel that influences buying decisions.

Start your free trial

Mapping hreflang at scale

The challenge was clear: map thousands of URLs across more than a dozen multilingual websites into accurate hreflang XML sitemaps.

Rather than tackling the project manually, I used Google Gemini to help build a custom Python solution.

Here’s how the process unfolded.

Phase 1: Asking for an approach, not just a script

A common pitfall when using generative AI for coding is asking it to sprint before it knows the route. If you simply type, “Write a Python script to create an hreflang sitemap,” you’ll get a generic, fragile piece of code that breaks the moment it encounters real-world data.

Instead, I started by asking for an approach. I explained the scenario: multiple regional domains, organic growth over several years resulting in mismatched URL slugs, translated subfolders, and appended revision years.

Gemini suggested a multi-step, data-driven approach:

  • Crawl the websites to collect live URLs and their metadata.
  • Use Python in Google Colab to process the raw data.
  • Run an exact match cluster first to group identical slugs.
  • Use an advanced semantic AI model (such as SentenceTransformers) to fuzzy match translated pages based on their titles and normalized URLs.

Phase 2: Crawling and data collection

Following the strategy, I used a crawler to spider all the regional websites. The goal was to generate a unified comma-separated values (CSV) file containing the live URLs, status codes, title tags, and H1s. Screaming Frog worked perfectly for this application.

A critical point: Your AI output is only as good as your crawl data (remember the old saying, “garbage in, garbage out”).

An AI script will fail to map an obvious “exact match” if the target URL is a 404 or a 301 redirect in your source data. You must filter your CSV to include only indexable content before feeding it to the script.

Dig deeper: International SEO in 2026: What still works, what no longer does, and why

Get the newsletter search marketers rely on.

See terms.

Phase 3: The Google Colab sandbox

Google Colab provides a free, cloud-based Jupyter notebook environment where you can write, paste, and execute Python code without worrying about local installations or environment variables. You can access it through Google Drive. I found the free version had enough capacity to handle this project.

I uploaded the CSV to Colab, and Gemini provided the initial Python script. The script used a domain-mapping routine to assign language codes, clean the URLs, and generate an XML tree. The initial output was far from perfect.

Phase 4: The iteration (where the real work happens)

If you expect AI to deliver a flawless, edge-case-proof script on the first try, you’ll be disappointed. You’ve probably heard the comparison of AI tools to interns, meaning you need to check their work. That’s very true.

The real value of AI lies in the iteration. As we ran the script, we encountered several unmatched URLs, leaving pages orphaned rather than grouping them with their international counterparts.

Here’s how I iteratively trained the AI to handle the nuances of human-managed websites.

The directory flattening problem

The U.S. site had recently reorganized its blog into topical folders, while the Mexican and Italian sites hadn’t yet been reorganized.

I prompted Gemini with these specific mismatched examples. It responded by adding a URL flattener function to the script, which stripped the topical folders behind the scenes so the translated slugs could align cleanly.

The aggressive semantic trap

To prevent the AI from mixing up different topics, we implemented concept traps. Initially, they were too strict. A UK article about the manufacturing sector wouldn’t match an Italian article because the U.S. title was slightly more generic.

I instructed Gemini to loosen the traps for generic industries while keeping them strictly enforced for critical acronyms (such as “SEO” versus “SEM”). This gave the AI the breathing room it needed to match creative translations.

The translated slug epiphany

The biggest breakthrough came while auditing the Mexican blog orphans. For example, the Spanish URL /detras-de-escenas-historias... is a direct translation of the English /behind-the-scenes-stories... I pointed this out to Gemini.

Instead of forcing me to hard-code hundreds of manual matches, Gemini updated the script to create a “Combined Semantic Signature.” It dynamically translated core operational phrases in the slugs, effectively bridging the language gap for the semantic matching model and connecting dozens of orphaned pages almost instantly.

Dig deeper: Cultural SEO: A practical framework for Spanish markets in AI search

Own the conversation before your competitors.

See where your brand appears, where it doesn’t, and exactly how to win more visibility across search, AI, local, social, and every channel that matters.

Start your free trial

Lessons from building an AI-assisted SEO tool

The project reinforced a simple lesson: AI works best when it’s treated as a collaborator rather than a shortcut.

  • Be the strategist, let AI be the coder: Don’t just demand a final product. Discuss the architecture, edge cases, and logic first. Treat AI like a junior developer that needs clear architectural direction.
  • Provide concrete examples: When a script fails, don’t just say, “It’s broken.” For this project, I provided either exact URLs that failed and the URLs they should have matched with, or groups of URLs with mismatches. AI needs concrete patterns to fix its logic.
  • Embrace the iterative loop: Expect to run the code, identify anomalies, and feed them back into the prompt. Each iteration makes the tool significantly smarter.
  • Leverage Google Colab: You don’t need to be a Python expert to use Python for SEO. Colab bridges the technical gap, allowing you to run complex data science libraries directly in your browser.

By the end of the project, we had a robust, highly customized Python script that could process a massive CSV and generate a cross-referenced hreflang XML sitemap in minutes.

AI isn’t going to replace technical SEOs anytime soon. However, SEOs who know how to collaborate with AI to build custom, scalable, and useful tools will have a significant advantage.

Dig deeper: How AI search defines market relevance beyond hreflang

Read more at Read More

Why next-question intent matters for AI search visibility

/0 Comments/in , /by
Why next-question intent matters for AI search visibility

Much of the GEO conversation focuses on how AI systems discover, extract, cite, and recommend content. That work matters. But visibility also depends on what the content contains once it’s found.

Next-question intent is a way to test whether a page provides enough information to support the user’s next decision, not just the initial query.

The first search is often only the starting point. Real decisions happen in the follow-up questions, comparisons, constraints, and objections that come next.

Content that helps answer those questions gives AI systems more useful material to summarize, compare, cite, and recommend.

From results to narratives: Traditional search vs. AI search

Traditional search was built around a results page: a ranked set of links users could scan, compare, and interpret for themselves. AI search is increasingly built around a synthesized answer drawn from multiple sources.

That changes what content must do. A page can rank, index, and appear technically sound, yet still fail to provide the information needed to support an AI-generated answer. That’s where next-question intent matters.

Search intent asks, “What is this user trying to do?”

Next-question intent asks, “What will the user need to know next before they can trust, compare, choose, buy, book, or move on?”

That question is becoming increasingly important because AI systems don’t simply match queries to pages. They assemble answers, comparisons, qualifications, and recommendations.

In that environment, content must support the full answer path, not just the first query.

Be the brand AI recommends.

See where your brand appears in AI search, where competitors are winning, and what it takes to become the answer AI recommends.

See your AI visibility

The first query is often only the doorway

A user’s first search is often broad, incomplete, or simply exploratory. It signals a direction. Real value appears in what comes next: the follow-up, the objection, the comparison, the constraint, the “practical anxiety,” the “Yes, but what about my very specific situation?” moment.

As the simplest example, someone searches “best CRM software for small business.” The first query becomes a doorway. But the actual buying process begins with the follow-up questions.

  • Which platform is easiest for a two-person team?
  • Which integrates best with QuickBooks?
  • Which one works for a business without a formal sales department?
  • Which one is best for a local service company rather than a software startup?
  • Which one won’t make an owner, office manager, or intern quietly resent tech?

These queries aren’t add-on or side questions. They’re the actual decision path.

Otherwise competent content fails at this stage. It answers the query, but doesn’t help complete the conversation. A page can define the category, mention benefits, include a few keywords, and still omit information buyers need to make decisions.

In traditional search, the user might click a few results and assemble context manually. In AI search, the system will assemble it for them. If your content lacks that useful context, it gives the system less to work with and may appear less visible.

Next-question intent is not just a writing exercise

The risk with any new content framework is that it becomes a fresh label for familiar advice. Next-question intent should do more than remind you to “write better content.” It should help you test whether a page contains enough context to support the next step in a user’s decision.

In practical terms, next-question intent means asking whether the content is answer-ready.

Answer-ready content addresses the user’s initial need, anticipates the next layer of decision-making, and provides specific, verifiable, and contextual information to support a synthesized answer.

This distinction matters because AI search visibility isn’t exclusively about rankings. It’s also about citations, mentions, recommendations, and whether a brand is recognized as a trusted answer in a given context.

Those outcomes require something more than volume. They depend on whether the brand’s content provides the system with enough substance to understand what the brand does, who it serves, when it’s useful, why it’s trustworthy, and how it compares to alternatives.

Where good content goes thin

Most brands have decent content that’s accurate, readable, and optimized around a keyword. There may even be an FAQ section, like a useful but decorative basket of afterthoughts.

In AI search, decent may not be enough.

AI systems need extractable clarity, but they also need context. They must understand what something is, who it’s for, when it’s useful (and when it’s not), what evidence supports the claim, and what the user should consider next.

This level of context is where many pages go thin.

As an example, a service page says, “We offer customized marketing strategies.” But what does customized mean?

  • A real strategy?
  • A lightly personalized template?
  • A monthly call where everyone nods at a dashboard no one has time to interpret?

The product page says “safe for families.” Safe for whom?

  • Babies?
  • Pets?
  • People with health issues?

A software page says, “built for small businesses.” What business?

  • A solo bookkeeper?
  • A nonprofit?
  • A 40-person heating and cooling company?
  • A founder doing payroll late at night after working all day?

Broad claims offer humans little to trust and AI systems little to use. Specific, structured, evidence-backed content offers something better.

Get the newsletter search marketers rely on.

See terms.

How to audit for next-question intent

A next-question audit looks beyond keyword coverage and asks whether a page contains the information needed to support the next step in the user’s journey.

For every important page, you should ask:

  • What’s the primary question this page answers?
  • What would a serious buyer, reader, or researcher ask next?
  • What objection would stop them from acting?
  • What comparisons would help them understand the category?
  • What proof would make this answer trustworthy?
  • What detail would make this financially, technically, locally, or personally relevant?
  • Where are we applying broad language because we haven’t done the harder thinking?

The best inputs for the audit often come from inside the business, not from keyword tools alone. Customer reviews, comparison queries, demo questions, sales calls, support tickets, chat logs, internal site search, and objection patterns can all reveal the questions real people ask when making decisions.

That information is often closer to the buyer’s actual path than a neat spreadsheet of keywords.

Examples of next-question content across industries

For a local service business, next-question content might involve service areas, prices, appointment windows, insurance, reviews, emergency availability, or what happens after someone books.

B2B software may invest in next-question content that involves integrations, user roles, implementation times, costs for switching, security, support, or whether a lower-tier plan is useful.

For higher-trust categories like medical, financial, and legal, next-question content involves scope, credentials, risk, regulation, evidence, or when to speak with a qualified professional.

The point isn’t to stuff pages with every possible question. It’s to build content around how people actually decide.

AI search rewards content that completes the answer

Next-question intent helps you avoid one of the least useful responses to AI search: publishing more content because visibility feels uncertain. The better move is more specific, decision-ready content.

If your page says, “I/we help small businesses grow,” explain which small businesses, what kind of growth, what constraints, what proof, what trade-offs, and what alternatives.

For example:

  • “We help local service businesses without in-house marketing teams improve search visibility and generate more qualified appointment requests by clarifying their website content, answering the questions clients actually ask, and building pages that support both traditional and AI-generated search. This is best for businesses looking for durable visibility rather than a quick paid-ad spike.”

In that same line of thought, if a page says “We’re eco-friendly,” explain the materials, sources, use cases, certifications, limitations, disposal issues, and even circumstances where that claim doesn’t apply.

If a page says “This is AI-powered,” explain what that AI tool actually does, what it automates, what remains human-led, what data it uses, and where users will still need judgment.

This isn’t writing for bots. It’s writing for real people whose decisions are increasingly being mediated by AI-generated answers. The goal is to make your expertise, relevance, and trustworthiness easier to understand and use.

If AI can’t find you, customers won’t either.

Track your visibility across AI search, uncover missed opportunities, and grow your presence where customers are asking questions.

See your AI visibility

The new visibility test

Traditional SEO asked whether a page could rank. AI search asks whether a page can contribute to the answer.

Any page can be indexed, optimized, and technically sound, yet still fail if it lacks substance. It might answer the initial query, but ignore the information users need to make a decision.

The opportunity isn’t to chase every new acronym or rebrand every content plan as a new discipline. It’s to build answer-ready content.

That means clearer definitions, stronger examples, honest comparisons, better proof, more precise positioning, and direct answers to the questions customers ask every day.

In traditional search, visibility belonged to the page that best matched the query. In AI search, it increasingly belongs to the content that helps people move forward.

Read more at Read More

Google says llms.txt files won’t harm or help your search rankings

/0 Comments/in , /by

Google updated its AI Search optimization guide to clarify that llms.txt files neither help nor hurt Google search rankings. It also confirmed that Google Search does not use llms.txt files.

What Google wrote. I bolded the portion that is new, where Google wrote that Google Search does not use AI text files, markup or Markdown files.

  • “You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn’t use them. Note that Google may discover, crawl, and index many kinds of files in addition to HTML on a website: this doesn’t mean that the file is treated in a special way.”

Google also added a new note that reads:

  • “It’s completely fine if you decide to create and maintain LLMS.txt files (or other similar files) for other services or systems that use these files. Doing so won’t harm (nor help) your visibility or rankings in Google Search, as Google Search ignores them.”

Here’s a screenshot of this section:

Dig deeper.

Why we care. There’s been a lot of confusion about how Google Search handles llms.txt, markdown, and other AI-related files. In short, Google Search may discover, crawl, and index these files, but it does not use them in any special way. Having them on your site won’t help your rankings, and it won’t hurt them either.

Read more at Read More

ChatGPT Opens Ads for All: How to React to This Shift

/0 Comments/in /by

For the past several months, advertising on ChatGPT meant getting an invitation. A small group of brands had access. Everyone else waited.

Self-serve access is now open to all advertisers, and the dynamics that made early access valuable are already starting to shift.

Key Takeaways

  1. ChatGPT surpassed $100 million in annualized ad revenue in its first six weeks, generated from less than 20 percent of eligible users seeing ads daily.
  2. Around 85 percent of free and Go tier users are eligible to see ads, meaning current revenue represents a small fraction of eventual ad capacity.
  3. Self-serve access launched in May 2026, opening the platform beyond the initial group of managed pilot brands via a new OpenAI Ads Manager.
  4. OpenAI removed the $50,000 minimum spend requirement entirely, opening the door for businesses of any size.
  5. ChatGPT now reaches 800 million weekly active users, processing 2.5 billion prompts daily.
  6. First-mover advantage is real, and it will not last long once self-serve competition normalizes pricing.

The Numbers Behind the Launch

ChatGPT crossed $100 million in annualized ad revenue in six weeks, which is a strong opening number on its own. The context makes it more striking. That figure came from less than 20 percent of eligible users seeing ads daily. With roughly 85 percent of free and Go tier users eligible to see ads, the platform is operating at a fraction of its eventual capacity.

OpenAI launched its self-serve Ads Manager in early May 2026, removing the significant minimum spend thresholds that had previously locked out most advertisers. During the pilot phase, entry required a $50,000 commitment minimum, which limited access to large brands and agency partners including Dentsu, Omnicom, Publicis, and WPP. 

OpenAI's ad manager.

Source

That barrier is now gone. Any U.S. business can sign up, set their own budget, and launch campaigns without going through a partner agency.

The platform has also added CPC and CPM bidding options alongside conversion tracking, pixel-based measurement, and attribution capabilities. That infrastructure shift matters. It transforms ChatGPT advertising from an experimental awareness product into a channel capable of performance measurement, which is what allows ad ecosystems to scale properly.

Geographic expansion is already underway, with OpenAI confirming rollout to Canada, Australia, New Zealand, the United Kingdom, Japan, South Korea, Brazil, and Mexico. For international advertisers, the time to start building familiarity with the platform is now, before it reaches your market.

Why This Channel Works Differently

Dropping your existing search or social creative into ChatGPT and expecting it to perform is a mistake. The environment is fundamentally different.

ChatGPT is a conversational platform. Users are having a dialogue, asking follow-up questions, getting synthesized answers, and making decisions based on what the platform surfaces. When someone clicks a Google ad, they are often at the beginning or middle of their research journey. When someone encounters an ad in ChatGPT, they have already spent time in a specific, multi-turn conversation that has narrowed their problem. The AI has done the educational and comparison work. The user is ready for a direct answer or a specific solution.

A branded answer in ChatGPT.

That intent depth is what makes ChatGPT advertising different from display or social. It also means that landing pages and creative designed for top-of-funnel traffic will underperform. The user who arrives from a ChatGPT ad is further along the decision process than most of your other paid traffic. Your messaging and destination need to match where they are.

The targeting model is also distinct. ChatGPT uses contextual matching based on current conversation topics, past chat history, and previous ad interactions rather than traditional keyword targeting or demographic signals. That combination of conversational depth and behavioral context creates a quality of intent signal that search and social cannot fully replicate.

A graphic asking whether people are using ChatGPT for search over Google.

OpenAI has been tracking ad quality closely. Fewer than seven percent of ads are currently rated as low relevance by users, and the company says improving that metric alongside user trust is an active priority. Early pilot results showed no negative impact on consumer trust metrics and low ad dismissal rates, which OpenAI interpreted as signals to move forward with expansion.

The Two Ad Formats Currently Running

Two formats are currently live inside ChatGPT. Both appear below the AI’s response, clearly labeled as sponsored and visually separated from the organic answer.

The first is a shopping product carousel with integration for checkout. This format is well-suited for ecommerce brands selling products with clear visual appeal and straightforward purchase paths.

The second is a conversational banner that includes a call-to-action and an “Ask ChatGPT about this ad” button. When a user clicks that button, they enter a conversation powered by information the advertiser has pre-loaded: product details, FAQs, and service specifics. ChatGPT answers user questions on behalf of the brand using that uploaded data. A user who asks about pricing, sizing, or features gets a direct, brand-informed answer without leaving the platform. This format is particularly powerful for high-consideration purchases and B2B categories where questions are complex and the buying cycle is long.

Where the Early Opportunity Is Clearest

The categories with the clearest early opportunity are the ones where users already turn to ChatGPT for research and decision-making. B2B software, professional services, financial products, health and wellness, travel and hospitality, and high-consideration consumer purchases all fit that profile. These are categories where the buying decision is complex, the conversation context is rich, and users are asking detailed questions across multiple sessions.

A study pie chart about ChatGPT ad presence.

High-consideration e-commerce also performs well, particularly where users compare specifications or ask the AI to evaluate options. Brands selling commodity goods or low-price impulse purchases will find the signal-to-noise lower, at least in the early stages before format options expand.

Start by identifying the specific questions users ask ChatGPT that relate to what you sell. Use ChatGPT itself to research those queries: the language the AI naturally uses to discuss your category is a preview of the context your ads will appear in. Align your messaging with that language. Those query moments are the equivalent of high-intent keywords in early search, and right now the auction pressure around them is low.

A graphic talking about where queries contain ChatGPT ads for commercial terms.

Set a test budget and treat it as education. A modest budget in the early months of self-serve access should be viewed as learning what works in conversational ad contexts, not as a channel expected to deliver strong ROAS immediately. The data you build now will be more valuable as the platform scales.

The Bigger Picture

ChatGPT’s ad launch is part of a broader shift in how discovery works. The platform now processes 2.5 billion prompts daily from 800 million weekly active users. That is not a niche experiment. It is a mainstream consumer behavior that brands need to account for.

The parallel to early search advertising is not a stretch. Google Ads in 2002, Facebook Ads in 2007, and ChatGPT Ads in 2026 follow the same pattern: access was initially limited, costs were low, and the brands that moved early built structural advantages that compounded over time. OpenAI is targeting $2.5 billion in ad revenue for 2026, with longer-horizon projections reaching $100 billion by 2030. For context, AI-driven search ads are projected to reach $26 billion by 2029, equivalent to 13.6 percent of total U.S. search ad spend.

The window for low-competition early adoption is open now. It will not stay that way.

FAQs

Do ChatGPT ads affect what the AI says in its responses?

No. OpenAI has been explicit on this point: ads do not influence ChatGPT’s answers. Sponsored content is always visually separated from the organic response and clearly labeled. Advertisers receive only aggregated performance data. Individual conversations stay private.

Who can see ChatGPT ads?

Currently, ads are shown to logged-in adult users on the Free and Go plans only. Users on Plus, Pro, Business, Enterprise, and Education plans see no ads. That means the addressable audience is the tens of millions of people on the free version of ChatGPT.

How is ChatGPT ad targeting different from Google or Meta?

ChatGPT targets based on current conversation context, past chat history, and previous ad interactions rather than demographics or keywords. This gives you access to a deeper intent signal than behavioral or interest-based targeting can provide.

What should my landing page look like for ChatGPT traffic?

Not like a generic homepage. Users arriving from ChatGPT ads have already had a specific, contextual conversation. Your landing page should acknowledge that context directly: match the problem they were discussing, provide the specific answer or solution they are looking for, and make the next step clear.

Conclusion

$100 million in annualized revenue from less than 20 percent of eligible users in six weeks is not a modest start. When self-serve scales, the minimum spend barrier is removed, and the eligible audience expands, those numbers move fast.

Move early. Set benchmarks. Learn how conversational advertising works in your category. The cost of waiting is higher than the cost of testing.

Read more at Read More

How a €30,000 underspend taught Simran Harichand the importance of the basics

/0 Comments/in , /by

Watch this video on YouTube

While managing a major B2B SaaS account, Hallam PPC Lead, Simran Harichand tightened a target CPA to improve efficiency but failed to monitor the impact. The change dramatically reduced spend, leaving the account €30,000 short of its monthly budget target.

When underspending becomes a business problem

Underspending isn’t just a media issue — it can affect a client’s future budgets. In this case, unused funds had to be returned to finance, making it harder for the marketing team to justify similar investment levels in future planning cycles.

The hardest part wasn’t the mistake

The most difficult moment came when Simran had to explain the situation to the client. Rather than making excuses, she took full responsibility for the error and acknowledged the impact it had on their goals.

Trust is built after the mistake

Although the client was understanding, trust had been damaged. Simran rebuilt confidence by introducing weekly budget pacing updates, showing transparency and proving the issue wouldn’t happen again.

Why the “brilliant basics” matter

The experience reinforced the importance of fundamentals such as budget pacing, account monitoring and conversion tracking. No matter how advanced advertising platforms become, strong basics remain the foundation of good performance.

What she’d do differently today

Looking back, Simran says she underestimated how much influence a target CPA change could have on delivery. Today, she treats any spend-related adjustment as a significant account change that requires close monitoring.

The danger of relying on AI without oversight

Simran supports testing AI-powered tools but warns against blindly adopting every new feature. She believes advertisers should balance experimentation with human oversight and strategic thinking.

Why conversion tracking remains the industry’s biggest blind spot

One of the most common issues she sees in account audits is poor tracking implementation. Inaccurate conversion data can lead to flawed optimisation decisions, making reliable measurement more important than ever.

The human side of client relationships

Strong client relationships can help teams navigate difficult moments when mistakes happen. Building trust through communication and honesty often matters just as much as delivering strong performance.

The bottom line

Mistakes are inevitable in PPC, but accountability and learning from them are what matter most. For Simran, the experience was a reminder that long-term success is built on mastering the fundamentals and maintaining trust.

Read more at Read More

Headline formats and Google Discover: What 3.4 million articles reveal

/0 Comments/in , /by
Google Discover headline formats

You’ve probably seen some version of these three claims:

  • Quote-led headlines outperform plain declarative ones by nearly 29%.
  • Question headlines underperform both, sometimes by 24%.
  • Format drives the result: Rewrite a statement as a quote, or add that magic word, and you should expect a real lift.

We tested all three against 1,674,518 English editorial articles and 1,690,295 French articles from the 1492.vision Discover corpus (November 2025 to May 2026): about 3.4 million editorial articles with at least one capture across our fleet.

They share a deeper flaw than any of their numbers.

All three treat headline format as a cause — a lever you pull to gain visibility. But the data shows, layer after layer, that a format’s measured effect is almost entirely a proxy for something else: which publisher used it, for which audience, and on which Discover surface.

The headline is a symptom of those choices, not an independent driver.

The clearest demonstration is Simpson’s paradox. Once you see it, you find it throughout the dataset.

A note on what we measure

Our metric isn’t clicks from Discover; no third party has that data. It’s hits per article: how often an article appears across the 1492.vision fleet we observe, a proxy for visibility.

The corpus is limited to editorial articles. YouTube and X are excluded because their headlines follow different conventions. We’ll return to both at the end—they sharpen the point more than anything else.

A word on why the volume matters: the entire argument depends on being able to slice 3.4 million articles by publisher, Discover surface, topic, and language while still retaining enough data in each segment for meaningful comparisons. That’s the difference between a number and an insight — and between a real format effect and a statistical mirage.

The number is real, at the wrong altitude

Mean hits by headline format, EN and FR

Pool all publishers together, and a clean gradient emerges: quote-led headlines at the top, statements at the bottom.

Lang Format Articles Mean hits Median vs statement
EN Quote-led 38,044 13.0 4 +37%
EN Quote inside 75,463 11.5 4 +21%
EN Question 53,081 10.2 4 +7%
EN Statement 1,674,518 9.5 3 baseline
FR Quote-led 179,472 52.8 13 +48%
FR Quote inside 223,052 49.9 12 +40%
FR Question 103,117 41.3 11 +16%
FR Statement 1,690,295 35.7 9 baseline

The commonly cited +29% is conservative for pure editorial articles: quote-led headlines show a +37% lift in English and +48% in French. Questions, far from underperforming, also outperform statements (+7% EN, +16% FR).

At this level of aggregation, claim 1 looks understated and claim 2 looks plainly wrong.

This is the level of aggregation where most headline advice is born. Hold onto that +37% figure — the rest of this piece is about what it’s actually measuring.

Hidden variable 1: which publisher

The aggregate can’t answer a crucial objection on its own: the publishers that use quotes aren’t the same publishers that don’t.

Celebrity media, regional dailies, and buzz-driven sites lean heavily on quotes and earn more Discover hits per article regardless of headline format. Pure-play publishers, wire services, and utility-focused sites favor declarative headlines and tend to sit lower.

The raw comparison, then, isn’t quote versus statement. It’s one publisher population versus another.

Simpson’s paradox in one chart

This is a textbook Simpson’s paradox: a strong trend in the aggregate that weakens, disappears, or reverses once you segment by group.

To get anywhere near the effect of headline format itself, the grouping variable has to be the publisher.

So make each publisher its own baseline: compare quote versus statement within the same site, holding audience and topic mix constant.

Across 324 English and 439 French publishers with enough of both formats — at least 50 quote and 200 statement articles each:

Within-publisher: does quote beat statement at the same site?
Lang Publishers Quote wins (median site) Quote wins (mean site) Median within-publisher Δ
EN 324 31.5% 55.9% +3.1%
FR 439 47.6% 57.4% +5.5%

In English, statements outperform quotes at 68% of publishers by the median; quote-led headlines hurt more often than they help. In French, the result is close to a coin flip.

That leaves the underlying format effect at roughly +3% to +5%—about five to nine times smaller than the aggregate figure.

(The mean is higher than the median because a minority of publishers see large gains from quotes. The median is the more reliable measure of the typical publisher.)

Stop here and the lesson sounds like “segment your data.” But the collapse points to something larger.

If three-quarters of a +37% effect was really a publisher effect, the obvious next question is: what else is the headline metric standing in for?

The rest of this article is a tour of those hidden variables. And by this point, the answer to claim 3 is already coming into view: the format itself isn’t the driver.

The same substitution, in reverse: questions

Questions: same Simpson, opposite direction

The conventional advice says questions underperform by roughly 24%. The aggregate view of our data says the opposite: questions outperform statements (+7% EN, +16% FR).

Both conclusions are wrong for the same reason. Question headlines are disproportionately used by high-engagement publishers, which inflates their aggregate performance.

Within publishers, the picture settles.

In English, question headlines show a modest real underperformance (-3.7%), winning at only 29.3% of sites. In French, the effect is essentially neutral (-0.5%), with questions outperforming at 46.2% of sites.

The conventional advice gets the direction roughly right in English and neutral in French, but its usual magnitude is about sixfold too large.

The question mark isn’t the cause. The kind of publisher using it is. Same hidden variable, opposite sign.

The effect won’t even hold still

Monthly within-publisher quote advantage

Even that modest within-publisher effect drifts from month to month.

In English, it peaks at +2.5% and turns negative in March 2026, while statements outperform questions at 55% to 60% of sites each month. In French, it ranges from +3% to +12% — strongest in December and February, weakest in March — with no clear trend.

A genuine causal lever shouldn’t wobble like this. A correlation tied to a shifting content mix should.

Hidden variable 2: Which audience

Top EN publishers, quote vs statement

The +3-5% average hides a sharp, consistent split. In English:

  • Gainers: International general news (BBC +85%, Forbes +46%, CBS News +43%, Boston Globe), Yahoo aggregators, mass-market magazines (Parade, Good Housekeeping), Gizmodo.
  • Losers: Specialist sport (RugbyPass, Planet F1, ThisIsAnfield), entertainment (IMDb, TVInsider, People), and factual-leaning dailies (Standard, Washington Post).
Top FR publishers, quote vs statement

French data follows the same pattern in a different market.

  • Gainers: Regional newspapers (La Dépêche, La Montagne, L’Écho Républicain) and general-interest magazines (Grazia).
  • Losers: Specialist sports outlets (Foot National, le10sport, MadeInFoot), technology publishers (Les Numériques), and service-oriented titles (Journal des Femmes, Femme Actuelle).

The pattern is editorial, not algorithmic. Quotes tend to work where the audience comes for commentary, reaction, and framing, and fail where the audience comes for facts.

A publisher built around “what someone said” benefits from a quoted headline. One built around “what just happened” usually doesn’t.

The convergence between English and French is the giveaway. This isn’t a language effect; it’s a reader-intent effect.

What looks like a headline-format effect is, in this case, an audience effect wearing the clothes of a headline.

Hidden variable 3: Which Discover surface

Discover isn’t a single feed. It’s a collection of pipelines, each selecting articles in different ways:

  • Editorial curation (moonstone, mustntmiss).
  • The main topic-personalization engine (aura).
  • Related-reading context (paginationpanoptic, content).
  • Similarity-based recommendation (relatedcontentruby, userpersonascontent).

First, rule out the obvious alternative explanation. Are quote-led articles simply being routed to higher-value Discover surfaces, making the apparent bonus a placement effect rather than a headline effect?

The data says no.

Comparing where quote and statement articles actually appear, the distributions are nearly identical. In English, the largest differences are small: content.f (+2.2 percentage points), aura.f (-1.9), and moonstone.f (+0.6).

Pipeline mix by format

The bonus isn’t about placement: quotes and statements appear on the same surfaces in the same proportions. It’s about intensity — how each format performs once it’s on a surface. There, the overall +3% to 5% breaks into a wide range: from +22% to -14% in EN and from +25% to -12% in FR.

Quote bonus by pipeline, EN, full picture

Grouped into functional families, the pattern is readable:

Pipeline family EN FR
Editorial curation (moonstone, mustntmiss, astria, news…) +3.4% +9.7%
Related reading / context (paginationpanoptic, content…) +2.0% +6.7%
Trends / freshness (deeptrends, freshvideos…) +4.4% +2.3%
Main personalization (aura) +0.6% +1.8%
Similarity-based recommendation (relatedcontentruby, userpersonas…) -1.6% -1.9%

Quote-led headlines win where multiple headlines compete for attention at once — curation carousels, news clusters, and other surfaces where the title carries a social signal: someone said this. They lose on similarity-based recommendations, where the surface sells continuity (“because you read X, you’ll read Y”) and a quote disrupts the topic-clear promise with an out-of-context citation.

The largest pipeline by volume, Aura, ranks on topic affinity and barely reacts to format at all, with gains of just +0.6% to +1.8%.

Why is the net effect so small?

A single quote-led FR article doesn’t get one number; it gets a blend:

  • +10 to +25% on its curation share (moonstone, mustntmiss, astria)
  • ~0% on its aura share, the largest slice of volume
  • -3% on its relatedcontentruby share (≈ 10% of captures)
  • -2 to -6% on shopping/viewer-related surfaces

Integrate those and you land at +4% to +7% net. The curatorial gains are real but partly offset by recommendation losses, which is why the aggregate is nowhere near +29%. The same format is both an asset and a liability, depending entirely on the surface serving it.

And +4–7% overstates how much the format itself matters because each pipeline’s ranking is a compound of signals unrelated to the title: engagement, scroll depth, topic affinity, E-E-A-T, entities, reading history, location, timing, and prior interactions.

A quote in the headline is, at best, one weak signal competing with all of those. Long before an article reaches a feed, it’s largely swamped by everything else.

Questions by pipeline, same story sharper

Question vs statement bonus, by pipeline

These are within-publisher medians (each publisher against itself), so they aren’t a crude artifact of FR using more questions. The format follows the same pipeline logic as quotes, but in a more polarized form:

  • FR curation leans positive on questions; EN curation leans negative. astria.f, the same pipeline in both languages, runs +9% in FR and -1% in EN; FR mustntmiss.f is +14%, EN moonstone.f is -13%.
  • Similarity-based recommendation penalizes questions everywhere, harder than quotes: relatedcontentruby.f FR -11.5% (306 publishers), EN -6.1% (119); itemitemcollaborativefiltering.f FR -14.5%.
  • aura stays neutral in both (+3.5% FR, -0.6% EN).

Two caveats point in the same direction:

  • A fleet-capture metric can’t distinguish an algorithmic penalty from an audience-eviction effect: readers see a question mark, decide “not now,” and scroll past. The fact that relatedcontentruby — which serves already-engaged readers — penalizes questions this heavily points to a behavioral signal, not just ranking.
  • Within-publisher pairing controls for each publisher against itself, but the median is still computed across a different set of publishers in FR and EN, on partly different surfaces. So “FR rewards questions, EN doesn’t” describes the publishers and topics occupying each cell, not an inherent property of the language or the question mark. It’s another hidden variable mistaken for a format effect.

Hidden variable 4: Which editor, and which judgment

Even the honest +3% to 5% comes with a caveat that outweighs its size. When a publisher writes a headline as a quote, they choose the best available quote for that story. So the within-publisher figure compares the best quote an editor selected with the average of all that publisher’s statements, not the same article written two ways.

It’s the subject-line A/B testing problem: a good alternative beats a bad one, but the average alternative doesn’t. Convert every headline to quote-led and you’d be writing average quotes, so most of the gain would disappear. The +3–5% is an upper bound on a selective practice, not the return from a blanket rule.

That’s the final reason “do it everywhere” fails:

  • Not every article has a quote. A sports result, a press release, a market analysis, a product test: forcing one means fabricating it.
  • The editor-selection bias above: The measured bonus is the best quote chosen, not a property of the format.
  • Recommendation pipelines are long-tail levers. relatedcontentruby and friends are how an article redeploys after its initial peak, the main mechanism for extending Discover lifetime. Optimizing the headline for the curation peak while breaking the promise on these surfaces can net negative.
  • The largest pipeline barely reacts. aura is 11% to 15% of FR captures and 7% to 9% of EN, with a +0.6% to 1.8% quote effect. A universal quote rule optimizes secondary surfaces while ignoring that the biggest one runs on topic affinity.

The clincher: the same format, opposite meaning

YouTube and x.com, quote bonus

We excluded YouTube and X from the main corpus, but their results are the clearest proof of the thesis. The same quote-led format produces opposite effects depending entirely on what the title is trying to do.

Domain Lang Quote articles Statement Mean hits quote Mean hits stmt Δ
YouTube EN 43,476 734,986 11.6 10.2 +14%
YouTube FR 16,509 93,912 59.0 29.1 +103%
x.com EN 34,156 268,175 5.2 4.9 +6%
x.com FR 32,201 114,914 21.4 24.6 -13%

On YouTube, the title is effectively a text thumbnail that has seconds to create curiosity. A quote serves as a content promise — “here’s the line worth hearing” — which helps explain the +103% result in French. On X, the title is the post itself, and a detected quote usually indicates that someone is repeating or responding to another person’s words, diluting the original message. That correlates with a -13% result.

Same characters. Same regex. Opposite outcome. The format didn’t change; the job it was doing did.

(Methodological footnote: a naive audit that folded YouTube into the editorial corpus would inflate the overall quote bonus by 20–30 points, while one that folded in X would dilute it. Any serious headline study has to isolate editorial articles before measuring headline effects.)

The headline was never the variable

Put the layers together. Three-quarters of the +37% raw bonus was explained by publisher differences. What remained split again by audience, then by Discover surface, then by which quote the editor selected, and finally reversed entirely when the title served a different function on another platform. At every step, removing context shrank or flipped the apparent format effect.

There’s no clean residue at the bottom where the headline acts independently. The effect is inseparable from the context that creates it.

That’s not a measurement failure; it’s the finding. We just saw the mechanism. Headline format is one weak signal among many stronger ones, all moving through pipelines that often pull in opposite directions.

The consequence is the point. An article’s visibility is the running score of that entire contest, not the verdict of any headline rule. A number measured across publishers is downstream of everything that travels with the format: who published it, what topic it covers, what the audience expects, the newsroom’s style and habits, and the conventions of the language itself.

So when an aggregate reports “+29% for quotes,” it isn’t isolating the quotation marks. It’s measuring a correlation with that whole bundle of factors and quietly relabeling it as causation.

None of this means aggregate data is the enemy. Everything above comes from aggregate data, just analyzed at the right level.

The trap is narrower: treating a single cosmetic variable, averaged across publishers that don’t belong in the same category, as a causal lever.

The same index that exposes that mistake also reveals the signals that genuinely drive Discover: which topics a publisher wins on, which entities are accelerating, who dominates a given surface, and what’s trending before it peaks. Those signals aren’t cosmetic, and they aren’t drowned out by stronger forces. They’re the underlying demand that headline format only weakly approximates.

The lesson isn’t “ignore the data.” It’s “stop averaging the wrong variable across the wrong population.”

This is why no cross-publisher average, corrected or not, converts into a rule for your site:

  • Visibility isn’t traffic. Two sites can earn identical Discover visibility on the same article and see very different CTRs because their audiences click for different reasons.
  • No two audiences are the same. A quote that reads as insider commentary to a magazine reader may read as vague or irrelevant to someone scanning sports scores.
  • A cross-publisher average of one cosmetic feature is the average of audiences you don’t have. Segment by your audience, your topics, and your surfaces, and it becomes information again.

The only test that answers your question is the one you run on your own site, with your own audience. Know who you’re writing for, then measure them. Slice the data by your audience, your topics, and your surfaces — not by a single number averaged across everyone.

So what about the three claims?

Each is real as a correlation and useless as a cause:

  • “Quotes beat statements by ~29%”: True in aggregate — larger than +29%, in fact — but mostly explained by publisher differences. At the publisher level, the residue is +3% to 5%, and even that compares the best quote an editor selected against the average of all statements, not the format itself.
  • “Questions underperform”: Directionally true in EN, neutral in FR, but the magnitude is about 6x too large. The actual effect is roughly -4% in EN and ~0% in FR.
  • “The format itself is the driver”: The claim the dataset refutes. The same article from the same publisher, mechanically rewritten as a quote, would not gain the aggregate effect.

The honest version, if you want one sentence to keep:

A quote-led headline can earn roughly +3% to 7% additional Discover visibility for audiences that value commentary and framing (general news, magazines, regional press), especially on curation surfaces, and lose for factual audiences (sports, tech, utility) and on similarity-based recommendation surfaces. There is no universal gain from quotation marks; the popular ~+29% figure overstates the format effect by roughly an order of magnitude. The useful question isn’t “Should I use a quote?” but “Who am I writing for, and which Discover surface drives my traffic?” The only place to answer that is with your own site, not anyone else’s average.

Methodology

  • Data and period: 1,674,518 EN and 1,690,295 FR editorial articles with Discover visibility from 1492.vision proprietary data, collected between 2025-11-01 and 2026-05-19. Editorial articles only; excludes ads, videos, AI Overviews, and showcases. Domain exclusions: x.com, twitter.com, m.twitter.com, youtube.com, www.youtube.com, and m.youtube.com (reported separately above).
  • Headline format detection (regex): Quote-led: title starts with a multi-word quoted phrase (“…”, «…», ‘…’, or ‘X…’:). Quote inside: a quoted phrase appears but not at the start. Question: ends with ?. Statement: everything else. Titles under 20 or over 300 characters are excluded. Detection deliberately errs toward false negatives in the quote bucket, biasing against finding a quote effect, so the +3–5% is conservative.
  • Three layers of analysis: (1) Raw aggregate: all publishers pooled, producing +37% / +48%. (2) Within-publisher: quote vs. statement inside each publisher with ≥50 quote and ≥200 statement articles; we report the share of publishers favoring quotes and the median per-publisher Δ. This neutralizes publisher-mix bias. (3) Monthly evolution: the same pairing, recomputed monthly with relaxed thresholds (≥10 quote, ≥40 statement).
  • Pipeline layer: Captures come from 1492.vision proprietary data, with each row representing one capture on a specific pipeline. For each (pipeline, format, publisher), captures per article = pipeline captures ÷ distinct articles. Within-publisher pairing includes publishers with ≥20 quote (or question) and ≥60 statement articles on that pipeline. A pipeline is shown only if ≥5 publishers qualify. Pipeline families are an empirical grouping (editorial curation, related reading, trends, similarity-based recommendation, and main personalization) that reflects how each surface behaves.
  • Metric: A “hit” is one capture of an article on Discover by the 1492.vision device fleet. It is a visibility proxy, not a visit.
  • Known limitations: (1) No traffic data: the metric is Discover visibility, not clicks, so a format could affect CTR independently without appearing here. (2) Regex detection misses edge cases and is biased toward under-counting quotes. (3) Within-publisher effects compare the best quote an editor selected against the average statement, not the counterfactual of making every headline quote-led. (4) Some negative pipelines have small publisher samples (<10); the consistent direction matters more than any individual magnitude.

Read more at Read More