Tag Archive for: AI

If you’re a content strategist, you might feel this isn’t your territory. Keep reading, because it is. Everything you build feeds these five gates, and the decisions the algorithms make here determine whether the system recruits your content, trusts it enough to display it, and recommends it to the person who just asked for exactly what you sell.

The DSCRI infrastructure phase covers the first five gates: discovery through indexing. DSCRI is a sequence of absolute tests where the system either has your content or it doesn’t, and every failure degrades the content the competitive phase inherits.

The competitive phase, ARGDW (annotation through won), is a sequence of relative tests. Your content doesn’t just need to pass. It needs to beat the alternatives. A page that is perfectly indexed but poorly annotated can lose to a competitor whose content the system understands more confidently. 

A brand that is annotated but never recruited into the system’s knowledge structures can lose to one that appears in all three graphs. The infrastructure phase is absolute: pass, stall, or degrade. The competitive phase is Darwinian “survival of the fittest.”

The DSCRI infrastructure phase determines whether your content even gets this far. The ARGDW competitive phase determines whether assistive engines use it.

Up until today, the industry has generally compressed these five distinct processes into two words: “rank and display.” That compression muddied visibility into several separate competitive mechanisms. Understanding and optimizing for all five will make all the difference in the world.

The competitive turn: Where absolute tests become relative ones

The transition from DSCRI to ARGDW is the most significant moment in the pipeline. I call it the competitive turn.

In the infrastructure phase, every gate is zero-sum: does the system have this content or not? Your competitors face the same test, and you both pass or fail. But the quality of what survives rendering and conversion fidelity creates differences that carry forward. 

The differentiation through the DSCRI infrastructure gates is raw material quality, pure and simple, and you have an advantage in the ARGDW phase when better raw material enters that competition.

At the competitive turn, the questions change. The system stops asking “Do I have this?” and starts asking “Is this better than the alternatives?” 

Every gate from annotation forward is a comparison. Your confidence score matters only relative to the confidence scores of every other piece of content the system has collected on the same topic, for the same query, serving the same intent.

You’ve done everything within your power to get your content fully intact. From here, the engine puts you toe to toe with your competitors.

Multi-graph presence as structural advantage in ARGD(W)

The algorithmic trinity — search engines, knowledge graphs, and LLMs — operates across four of the five competitive gates: annotation, recruitment, grounding, and display. Won is the outcome produced by those four gates. Presence in all three graphs creates a compounding advantage across ARGD, and that vastly increases your chances of being the brand that wins.

The systems cross-reference across graphs constantly. An entity that exists in the entity graph with confirmed attributes, has supporting content in the document graph, and appears in the concept graph’s association patterns receives higher confidence at every downstream gate than an entity present in only one.

This is competitive math. If your competitor has document graph presence (they rank in search), but no entity graph presence (no knowledge panel, no structured entity data), and you have both, the system treats your content with higher confidence at grounding because it can verify your claims against structured facts. The competitor’s content can only be verified against other documents, which is a higher-fuzz verification path — more interpretation, more ambiguity, lower confidence.

For me, this is where the three-dimensional approach comes into its own, and single-graph thinking becomes a structural liability. “SEO” optimizes for the document graph. Entity optimization (structured data, knowledge panel, and entity home) optimizes for the entity graph. 

Consistent, well-structured copywriting across authoritative platforms optimizes for concept graph. Most brands invest heavily in one (perhaps two) and ignore the others. The brands that win at the competitive gates are stronger than their competitors in all three at every gate in ARGD(W).

Your customers search everywhere. Make sure your brand shows up.

The SEO toolkit you know, plus the AI visibility data you need.

Start Free Trial

Get started with

Annotation: The gate that decides what your content means across 24+ dimensions

Annotation is something I haven’t heard anyone else (other than Microsoft’s Fabrice Canel) talking about. And yet it’s very clearly the hinge of the entire pipeline. It sits at the boundary between the two phases: the last gate that applies absolute classification, and the first gate that feeds competitive selection. Everything upstream (in DSCRI) prepared the raw material. Everything downstream in ARGDW depends on how accurately the system can classify it.

At the indexing gate, the system stores your content in its proprietary format. Annotation is where the system reads what it stored and decides what it means. The classification operates across at least five categories comprising at least 24 dimensions.

Canel confirmed the principle and confirmed there are (a lot) more dimensions than the ones I’ve mapped. What follows is my reconstruction of the categories I can identify from observed behavior and educated guesses.

Canel confirmed the Annotation gate back in 2020 on my podcast as part of the Bing Series, in the episode “Bingbot: Discovering, Crawling, Extracting and Indexing.”

• “We understand the internet, we provide the richness on top of HTML to a lot, lot, lot of features that are extracted, and we provide annotation in order that other teams are able to retrieve and display and make use of this data.”
• “My job stops at writing to this database: writing useful, richly annotated information, and handing it off for the ranking team to do their job.”

So we know that annotation is a “thing,” and that all the other algorithms retrieve the chunks using those annotations.

Annotation classification runs across five types of specialist models operating simultaneously per niche: 

• One for entity and identity resolution (core identity).
• One for relationship extraction and intent routing (selection filters).
• One for claim verification (confidence multipliers).
• One for structural and dependency scoring (extraction quality).
• One for temporal, geographic, and language filtering (gatekeepers). 

This five-model architecture is my reconstruction based on observed annotation patterns and confirmed principles. The annotation system is a panel of specialists, and the combined output becomes the scorecard every downstream gate uses to compare your content against your competitors.

Gatekeepers 

They determine whether the content enters specific competitive pools at all:

• Temporal scope (is this current?).
• Geographic scope (where does this apply?).
• Language.
• Entity resolution (which entity does this content belong to?). 

Fail a gatekeeper, and the content is excluded from entire query classes regardless of quality.

Core identity

This classifies the content’s substance: entities present, attributes, relationships between entities, and sentiment. 

For example, a page about “Jason Barnard” that the system classifies as being about a different Jason Barnard has perfect infrastructure and broken annotation. The content was there, and the system read it, but filed it in the wrong drawer.

Selection filters 

They add query routing: intent category, expertise level, claim structure, and actionability. 

For example, content classified as informational never surfaces for transactional queries, regardless of how well it performs on every other dimension.

Extraction quality

Think:

• Sufficiency (does this chunk contain enough to be useful?)
• Dependency (does it rely on other chunks to make sense?)
• Standalone score (can it be extracted and still work?)
• Entity salience (how central is the focus entity?)
• Entity role (is the entity the subject, the object, or a peripheral mention?)

Weak chunks get discarded before competition begins.

Confidence multipliers 

These determine how much the system trusts its own classification: verifiability, provenance, corroboration count, specificity, evidence type, controversy level, consensus alignment, and more.

Two pieces of content can be classified identically on every other dimension and still receive wildly different confidence scores based on how verifiable and corroborated their claims are.

An important aside on confidence

Confidence is a multiplier that determines whether systems have the “courage” to use a piece of content for anything.

Once upon a time, content was king. Then, a few years ago, context took over in many people’s minds.

Confidence is the single most important factor in SEO and AAO, and always has been — we just didn’t see it.

To retain their users, search and assistive engines must provide the most helpful results possible. Give them a piece of content that, from a content and context perspective, appears to be super relevant and helpful, but they have absolutely no confidence in it for one reason or another, and they likely will not use it for fear of providing a terrible user experience.

What happens when annotation fails you (silently)

Annotation failures are the most dangerous failures in the pipeline because they are invisible. The content is indexed. But if the system misclassifies it, every competitive decision downstream inherits that misclassification.

I’ve watched this pattern repeatedly in our database: a page is indexed, it appears in search results, and yet the entity still gets misrepresented in AI responses.

Imagine this: A passage/chunk from your website is in the index, but confidence has degraded through the DSCRI part of the pipeline, and the annotation stage has received a degraded version. 

The structural issues at the rendering and indexing gates didn’t prevent indexing, but they were degraded versions of the original content. That degradation makes the annotation less accurate, less complete, and less confident. That annotative weakness will propagate through every competitive gate that follows in ARGDW.

When your content is included in grounding or display, and it’s suboptimally annotated, your content is underperforming. You can always improve annotation.

Measuring annotation quality in ARGDW

Annotation quality is the most important gate in the AI engine pipeline, but unfortunately, you can’t measure annotation quality directly. Every metric available to you is an indirect downstream effect.

The KPIs I suggest below are signals that clearly show where your content cleared indexing and failed annotation: the engine found the page, rendered it, indexed it, and then drew the wrong conclusions from it.

That distinction matters: beware of “we need more content” when the real problem is “the engine misread the content we have.”

Your brand SERP tells you exactly what the algorithm understood

These signals reveal how accurately the AI has understood who you are, what you do, and who you serve. The brand SERP (and AI résumé) is a readout of the algorithm’s model of your brand and, because it is updated continuously, makes it a great KPI.

• Brand SERP shows incorrect entity associations: wrong competitors, wrong category, wrong geography.
• AI résumé is noncommittal, hedged, or incomplete.
• AI outputs underestimate your NEEATT credentials.
• Knowledge panel displays incorrect information.
• AI describes your brand using a competitor’s framing or category language.
• Entity type is misclassified (person treated as organization, product treated as service).
• AI can’t answer basic factual questions about your brand and offers without hedging.

If the algorithm can’t place you in a competitive set, it won’t recommend you

These signals reveal which entities the system considers comparable — a direct readout of how annotation classified them. Annotation places entities into competitive pools, and if your brand doesn’t appear in comparison sets where it belongs, the engine classified it outside that pool. Better content won’t fix that. Improving the algorithm’s ability to accurately, verbosely, and confidently annotate your content will.

• Absent from “best [product] for [use case]” results where you qualify.
• Absent from “alternatives to [competitor]” results.
• Absent from “[brand A] vs. [brand B]” comparisons for your category.
• Named in comparisons but with incorrect differentiators or misattributed features.
• Consistently ranked below competitors with weaker real-world authority signals.

For me, that last one is the most telling. Weaker brand, higher placement.

Once again, what you’re saying isn’t the problem, how you’re saying it and how you “package” it for the bots and algorithms is the problem.

If the algorithm can’t surface you unprompted, you’re invisible at the moment of intent

These signals reveal whether the AI can place your brand at the point of discovery, before the user knows you exist. Clearing indexing means the engine has the content. Failing here means annotation didn’t connect that content to the broad topic signals that drive assistive recommendations. 

The difference between a brand that appears in “how do I solve [problem]” answers and one that doesn’t is whether annotation connected the content to the intent.

• Absent from “how do I solve [problem your product solves]” answers, even as a passing mention.
• Not surfaced when the AI explains a concept you coined or own.
• Absent from AI-generated roundups, guides, and “where to start” responses for your core topic.
• Named as a generic example rather than a recommended solution.
• The AI discusses your subject area at length and doesn’t name you as a practitioner or source.
• Entity present in the knowledge graph but invisible in discovery queries on AI platforms.

The three taxes you’re paying with sub-optimal annotation

Three revenue consequences follow from annotation failure, one at each layer of the funnel. 

• The doubt tax is what you pay at BoFu when a buyer reaches your brand in the engine and the AI presents a confused, incomplete, or misframed version of what you offer. 
• The ghost tax is what you pay at MoFu when you belong in the consideration set and the algorithm doesn’t prominently include you. 
• The invisibility tax is what you pay at ToFu when the audience doesn’t know to look for you and the algorithm doesn’t introduce you. 

Each tax is a direct read of how well annotation worked — or didn’t.

For you as an SEO/AAO expert, you can diagnose your approach to reduce these three taxes for your client or company as: 

• BoFu failures point to entity-level misunderstanding. 
• MoFu failures point to competitive cohort misclassification.
• ToFu failures point to topic-authority disconnection.

Annotation should be your focus. My bet is that for the vast majority of brands, the gate in the pipeline with the biggest payback will be annotation. 99% of the time, my advice to you is going to be “get started on fixing that before you touch anything else.”

For the full classification model in academic depth, see: 

• Annotation as the Confidence Fulcrum: How AI Systems Classify Digital Content and Why It Determines Recommendation Outcomes
• Annotation Cascading: Hierarchical Model Routing, Topical Authority, and Inter-Page Context Propagation in Large-Scale Web Content Classification

Recruitment: The universal checkpoint where competition becomes explicit

Recruitment is where the system uses your content for the first time. Every piece of content the system has annotated now competes for inclusion in the system’s active knowledge structures, and this is where head-to-head competition begins.

Every entry mode in the pipeline — whether content arrived by crawl, by push, by structured feed, by MCP, or by ambient accumulation — must pass through recruitment. No content reaches a person without being recruited first. We could call recruitment “the universal checkpoint.”

The critical structural fact: it recruits into three distinct graphs, each with different selection criteria, different confidence thresholds, and different refresh cycles. The three-graph model is my reconstruction. 

The underlying principle (multiple knowledge structures with different characteristics) is confirmed by observing behavior across the algorithmic trinity through the data we collect (25 billion datapoints covering Google’s Knowledge Graph, brand search results, and LLM outputs).

The entity graph stores structured facts with low fuzz — who is this entity, what are its attributes, how does it relate to other entities, binary edges — and knowledge graph presence is entity graph recruitment, with entity salience, structural clarity, source authority, and factual consistency as the selection criteria.

The document graph handles content with medium fuzz — passages and pages and chunks the system has annotated and assessed as worth retaining — where search engine ranking is the visible output, and relevance to anticipated queries, content quality signals, freshness, and diversity requirements drive selection.

The concept graph operates at a different level entirely, storing inferred relationships with high fuzz — topical associations, expertise patterns, semantic connections that emerge from cross-referencing multiple sources — with LLM training data selection as the mechanism and corroboration patterns as the primary selection criterion.

The same content may be recruited by one, two, or all three graphs. Each graph has its own speed of ingestion and its own speed of output. I call these the three speeds, a pattern I formulated explicitly this year but have been observing empirically across 10 years of brand SERP experiments: 

• Search results are daily to weekly.
• Knowledge graph updates are monthly. 
• LLM updates are currently several months (when they choose to manually refresh the training data).

Grounding: Where the system checks its own work in real time

Recruitment stored your content in the system’s three knowledge structures. Grounding is where the system checks whether it should trust your content, right now, for this specific query.

Search engines retrieve from their own index. Knowledge graphs serve stored structured facts. Neither needs grounding. Only LLMs have the (huge) gap between stale training data and fresh reality that makes grounding necessary. 

The need for grounding will gradually disappear as the three technologies of the algorithmic trinity converge and work together natively in real time.

In an assistive Engine, the LLM is the lead actor. When the user asks a question or seeks a solution to a problem, the LLM assesses its confidence in its own answer. 

If confidence is sufficient, it responds from embedded knowledge. If confidence is low, it sends cascading queries to the search index, retrieves results, dispatches bots to scrape selected pages, and synthesizes an answer from the fresh evidence (Perplexity is the easiest example to see this in action — an LLM that summarizes search results).

But that’s too simplistic. The three grounding sources model that follows is my reconstruction of how this lifecycle operates across the algorithmic trinity.

The search engine grounding the industry currently focuses on is this: the LLM queries the web index, retrieves documents, and extracts the answer. That’s high fuzz.

Now add this: Knowledge graph allows a simple, quick, and cheap lookup: low fuzz, binary edges, no interpretation required, and our data shows that Google does this already for entity-level queries.

My bet is that specialist SLM grounding is emerging as a third source. We know that once enough consistent data about a topic crosses a cost threshold, the system builds a small language model specialized for that niche, and that model becomes a domain-expert verifier. It would be foolish not to use that as a third grounding base.

The competitive implication is huge. A brand with entity graph presence gives the system a low-fuzz grounding path. A brand without it forces the system onto the high-fuzz path (document retrieval), which means more interpretation, more ambiguity, and lower confidence in the result. The competitor with structured entity data gets verified faster and more accurately.

In short, focus on entity optimization because knowledge graphs are the cheapest, fastest, and most reliable grounding for all the engines.

---

Display: Where machine confidence meets the person

Your content has been annotated, recruited into its knowledge structures, and verified through grounding. Display is where the AI assistive engine decides what to show the person (and, looking to the future that is already happening, where the AI assistive Agent decides what to act upon).

Display is three simultaneous decisions: format (how to present), placement (where in the response), and prominence (how much emphasis). A brand can be annotated, recruited, and grounded with high confidence and still lose at display because the system chose a different format, placed the competitor more prominently, or decided the query deserved a different type of answer entirely.

This is essentially the same thing as Bing’s Whole Page Algorithm. Gary Illyes jokingly called Google’s whole page algorithm “the magic mixer.” Nathan Chalmers, PM for the whole page algorithm at Bing, explained how that works on my podcast in 2020. Don’t make the mistake of thinking this is out of date — it isn’t. The principles are even more relevant than ever.

UCD activates at display

You may have heard or read me talking obsessively about understandability, credibility, and deliverability. UCD is absolutely fundamental because it is the internal structure of display: the vertical dimension that makes this gate three-dimensional.

The same content, grounded with the same confidence, presents differently depending on who is asking and why.

A person arriving with high trust — they searched your brand name, they already know you — experiences display at the understandability layer, where the engine acts as a trusted partner confirming what they already believe, which is BOFU.

A person evaluating options — they asked “best [category] for [use case]” — experiences display at the credibility layer, where the engine presents evidence for and against as a recommender, which is MOFU.

A person encountering your brand for the first time — a broad topical question in which your name appears — experiences it at the deliverability layer, where the system introduces you, which is TOFU.

The user interaction reveals the funnel position. The funnel position determines which UCD layer fires.

This is why optimizing only for “ranking” misses reality: Display is a context-sensitive presentation, not a list, and the same piece of content can introduce, validate, or confirm depending on who asked.

The framing gap at display

The system presents what it understood, verified, and deemed relevant. The gap between that and your intended positioning is the framing gap, and it operates differently at each funnel stage.

• At TOFU, the gap is cognitive: the system may know you exist, but doesn’t associate you with the right topics. 
• At MOFU, the gap is imaginative: the system needs a frame to differentiate your proof from the competitor’s, and most brands supply claims without frames. 
• At BOFU, the gap is about relevance: the system cross-references your claims against structured evidence, and either confirms or hedges.

After annotation, framing is the single most important part of the SEO/AAO puzzle, so I’ll talk a lot about both in the coming articles.

Won: The zero-sum moment where one brand wins and every competitor loses

Everything I’ve explained so far in this series collapses into a zero-sum point at the “won” gate. Here, the outcome is binary. The person (or agent) acts, or they don’t. One brand converts, and every competitor loses. 

The system may have mentioned others at display, but at the moment of commitment, there can only be one winner for the transaction.

Three won resolutions in the competitive context

Won always resolves through three distinct mechanisms, each with different competitive dynamics.

Resolution 1: Imperfect click

• The AI influences the person’s thinking at grounding and display, but the person decides independently: they choose one of several options offered by the engine, they walk into the store, or they book by phone. 
• This is what Google called the “zero moment of truth,” where the competitive battle happens at display, where the engine has influenced the human, but the active choice the person makes is still very much “them.”

Resolution 2: Perfect click

• The AI recommends one brand and the person takes it. This is the natural next step, what I call the zero-sum moment. 
• This fires inside the AI interface, where the engine filtered for intent, context, and readiness, presented one answer, and the person converted.

Resolution 3: Agential click

• The AI agent acts autonomously on the person’s behalf. No person at the decision point, an API settlement between the buyer’s agent, and the brand’s action endpoint. 
• The competitive battle happened entirely within the engine: whichever brand had the highest accumulated confidence, the strongest grounding evidence, and a functional transaction endpoint is the winner. The person doesn’t choose. The system chooses for them.

The trajectory runs from oldest to newest: Resolution 1 was dominant up to late 2025, Resolution 2 is taking over, and Resolution 3 gained a lot of traction early 2026. Stripe and Cloudflare are laying the transaction and identity rails. Visa and Mastercard are building the financial authorization infrastructure. 

Anthropic’s MCP is providing the coordination layer. Google’s UCP and A2A are defining how agents communicate across the full consumer commerce journey. Apple has the closed-loop infrastructure to make it seamless on a billion devices the moment they choose to. 

Microsoft is locking in the enterprise and government layer through Copilot in a way that will be extremely difficult to displace. No single company turns Resolution 3 on — but all of them together make it inevitable.

Competitive escalation across the five ARGDW gates

The competitive intensity increases at every gate — a progressive narrowing, a Darwinian funnel where the field shrinks at each stage. The narrowing pattern is my model based on observed outcomes across our database. The underlying principle (competitive selection intensifies downstream) is structural to any sequential gating system.

• The field is large at annotation, where the algorithms create scorecards and your classification versus competitors’ determines downstream positioning.
• Recruitment sets the qualifying round: multiple brands enter the system’s knowledge structures, but not all, and the selection criteria already favor multi-graph presence.
• Grounding narrows the shortlist as confidence requirements tighten — the system verifies the candidates worth checking, not everyone.
• Display reduces to finalists, often one primary recommendation with supporting alternatives.
• Won is the binary outcome. The zero-sum moment you’re either welcoming with open arms or fearful of.

ARGDW: Relative tests. The scoreboard is on.

Five gates. Five relative tests. Competitive failures in ARGDW are significantly harder to diagnose than infrastructure failures in DSCRI because the fix is competitive positioning rather than technical.

• Annotation failures mean the system misclassified what your content is or who it belongs to — write for entity clarity, structure claims with explicit evidence, and use schema markup to declare rather than expect the system to guess.
• Recruitment failures increasingly mean you’re present in one graph while competitors have two or three — build entity graph presence (structured data, knowledge panel, entity home), document graph presence (content quality, topical coverage), and concept graph presence (consistent publishing across authoritative platforms) as a coordinated program.
• Grounding failures mean the system is verifying you on the high-fuzz path — provide structured entity data for low-fuzz verification, and MCP endpoints if you need real-time grounding without the search step.
• Display failures mean the framing gap is costing you at the three layers of the visible gate — assuming you fixed all the upstream issues, then closing that framing gap at every UCD layer is your pathway to gain visibility in AI engines.
• Won failures mean the resolution mechanism doesn’t exist — Resolution 1 requires that you rank (good enough up to 2024), Resolution 2 requires that you dominate your market (good enough in 2026), and Resolution 3 requires a mandate framework and action endpoint (needed for 2027 onward).

See the complete picture of your search visibility.

Track, optimize, and win in Google and AI search from one platform.

Start Free Trial

Get started with

After establishing the 10-gate AI engine pipeline, what’s next?

The aim of this series of articles is to give you the playbook for the DSCRI infrastructure phase and the strategy for the ARGDW competitive phase. This 10-gate AI engine pipeline breaks optimizing for assistive engines and agents into manageable chunks.

Each gate is manageable on its own. And the relative importance of each gate is now clear for you (I hope). In the remainder of this series of articles, I’ll provide solutions to the major issues at each gate that will help you manage each individually (and as part of the collective whole).

Aside: The feedback I have had from Microsoft on this series so far (thank you, Navah Hopkins) reminded me of something Chalmers said to me about Darwinism in Search back in 2020.

• “… it’s not as mechanical as the way you outlined it.” – Nathan Chalmers

My explanations are often more absolute and mechanical than the reality. That’s a very fair point. But then reality is unmanageably nuanced, and nuance leads to a lack of clarity and often paralyzes people to the extent that they struggle to identify actionable next steps. I want to be useful.

I suggest we take this evolution from SEO to AAO step by step. Over the last 10+ years, I’ve always done my very best to avoid saying “it depends.”

People often say it takes 10,000 hours to become an expert. The framework presented here comes from tens of thousands of hours analyzing data, experimenting, working with the engineers who build these systems, and developing algorithms, infrastructure, and KPIs.

The aim is simple: reduce the number of frustrating “it depends” answers and provide a clear outline for identifying actionable next steps.

This is the fifth piece in my AI authority series. 

• The first, “Rand Fishkin proved AI recommendations are inconsistent – here’s why and how to fix it,” introduced cascading confidence. 
• The second, “AAO: Why assistive agent optimization is the next evolution of SEO,” named the discipline. 
• The third, “The AI engine pipeline: 10 gates that decide whether you win the recommendation,” mapped the full pipeline. 
• The fourth, “The five infrastructure gates behind crawl, render, and index,” walked through the first five gates.
• Up next: “The brand’s digital footprint: Entity home, entity home website, and the content map.”

Read more at Read More