Tag Archive for: AI

The AI engine pipeline runs 10 gates from discovered through won.

• Discovered is the bot finding your page. Alongside selected, crawled, rendered, and indexed, those five infrastructure gates get you legible to the machine. 
• Annotated, recruited, grounded, and displayed are the four competitive gates where the algorithm decides whether your brand is the brand it’s prepared to put in front of a buyer. 
• Won is the gate that pays for everything: the click happens, the recommendation lands, and the agent transacts.

Won has changed beyond recognition in the last 24 months. It used to mean a click on a search result, the human doing the picking, the brand competing for attention against nine blue links. It still means that, sometimes.

It can also mean an assistive engine naming your brand to a user who accepts the recommendation. Or it can mean an Agent transacting on the user’s behalf.

All of this is about delegation: how much we delegate to machines, and when. Delegation in the context of search and AI is far from new. We’ve delegated “finding the books to the librarian” since AltaVista.

What’s new is that the boundary of what and where we delegate is now flexible: the user can hand more of the journey to the engine than ever before, and the brand that wants to win has to be ready for every option on the delegation spectrum.

What hasn’t changed: The point of search

Underneath the three mechanisms sits the same commercial truth that’s been the point of search since Sergey Brin first articulated it: get the user to the best solution to their problem as efficiently as possible.

AI hasn’t changed the point of search. AI assistive engines and agents simply get the user to the best solution to their problem significantly more efficiently than search: a 15-minute purchase journey through ChatGPT that would have taken a week through traditional Search. In essence, AI removes an enormous amount of friction inherent in search.

The delegation boundary is the line between what the user does for themselves and what they hand to the engine. The further the user pushes it toward the engine, the less work the user does, and the faster won arrives. The further the user holds it back, the longer won takes.

From problem to purchase in 15 minutes with ChatGPT

I’m a double bass player, and I have a huge bass amp in my living room. I also have a guitar my father gave me when I was 18, which I’ve never played much, because I became a professional double bass player in the 1990s, so the guitar has lived as a keepsake for most of its life.

A few months back, I got an offer to play a solo gig at the jazz club opposite my flat. I had a guitar, but no guitar amp, and I didn’t want to buy a guitar amp for one gig. I figured I could just use the bass amp.

Here’s my conversation with ChatGPT:

• Me: Can I play guitar through my bass amp? Will I break it?
• ChatGPT: No, you won’t break it. But it’ll sound terrible.
• Me: How do I make it sound good?
• ChatGPT: Three pedals: reverb, compression, equalization.
• Me: Which ones?
• ChatGPT: Boss, JHS, TC Electronic.
• Me: Price?
• ChatGPT: Around $250 for the set.
• Me: Cheaper. I’m more of a singer than guitarist.
• ChatGPT: I can get you there for $125.
• Me: I need delivery by Friday, guaranteed. Saturday gig.
• ChatGPT: Thomann in Europe, Sweetwater in the US. Both will hit Friday.
• Me: Europe.
• ChatGPT: [Thomann link]
• Me: Clicks. Buys.

The Google team specifically asked me to make this point in the keynote, and I want to land it here too because they care about it for a reason most marketers haven’t worked out yet.

The engine made decisions for me all the way down the funnel. It decided whether the question even had an answer, which pedals were worth shortlisting at which price tier, and who could meet a Friday deadline.

My delegation boundary sat at the perfect click. ChatGPT owned the entire research-and-recommendation funnel. I owned the buy button, and only because Thomann doesn’t yet have an agential checkout. If they had, the agent would have transacted while I was making coffee.

The point isn’t that the funnel compressed from a week to fifteen minutes. The point is what happened inside the compression: the engine made dozens of small decisions on my behalf, each one quietly closing off options that might otherwise have stayed open. 

A different engine might have surfaced different brands, recommended a different price tier, or picked a different supplier. The brand that wins isn’t the one the user chose. It’s the one that survived every one of those upstream decisions the engine made before the user ever saw a recommendation.

Two things put my boundary that far to the right. 

• Emotional weight: The pedals just needed to be good enough, not perfect. 
• Domain expertise: I’m not a guitarist. I have no opinions about boutique pedal brands. The comparison work a serious guitarist would relish is friction I’d happily have someone else do.

A working professional would have approached the same purchase entirely differently. A studio musician whose tone is their living would have gone into the shop, plugged the pedals in, and decided in the room. A geeky enthusiast would have spent the week on Reverb forums comparing JHS to Strymon to Walrus, because for them, the research is the fun, the comparing is the point.

The point here is that you have the same purchase made by three buyer personas, each with a completely different position on the delegation boundary, and Thomann needs to be ready to win all three.

I’ve bought over €2,000 of equipment from Thomann since.

The single-mode assumption is dead, three modes coexist now

For two decades, “optimize for search” was the whole of the job: get in the top 10 and win the click at each stage of the funnel. Exhausting work, and it’s no longer enough.

That single approach has been replaced with three modes running in parallel. Search hasn’t gone anywhere. It just has assistive and agential sitting alongside as alternatives for all or part of the journey for each use.

• Search tolerates a fuzzy or unclear brand because the human will do the sorting. 
• Assistive tolerates less of it, because the AI is recommending you to the user, and the AI’s credibility is on the line every time it uses your name. 
• Agent tolerates none of it, because the agent transacts without asking, and a fuzzy brand is exactly the kind of risk the agent will quietly route around, and the user will never know.

Won looks different in each mode. 

• In search, won is the click, and then the user finishes the journey on your site, and you need to deal with friction (objections, questions, and clarifications). 
• In assistive, won is the AI engine naming you and the user accepting the name without further verification, and so most of the friction has already been dealt with. 
• In agent, won is the transaction completed without anyone consulting the user, friction reduced to zero from the user’s point of view.

The pedal-buying journey sat at the assistive end, and could have been at the agential end. The professional studio musician sits at the search end. The geeky enthusiast sits at the search end too, because for him, the friction is the entertainment.

The delegation boundary moves with every purchase, person, and culture

A user might delegate a coffee order to an agent without thinking, ask an assistive engine for advice on a kitchen renovation but make the final supplier call themselves, then spend an hour happily window-shopping for a bracelet that’s “just what I always wanted.” Same person, same week, three positions on the boundary, and all three engines doing useful work.

The diagram shows that one person sets the delegation boundary at completely different positions depending on the purchase. A wedding venue lives at the far left of search mode, because the decision is emotional and irreversible, and you wouldn’t want AI to make that journey for you. 

A few notches to the right, still in search mode, you’ve got someone who’s a sock collector and is window-shopping the listings because the choosing is the point. Further right, still in search mode but inching toward assistive, you’ve got the strategic business contract where only the human holds the strategic context that the AI doesn’t have access to.

Cross over into assistive mode, and you’ve got the holiday I’m planning, where I’m asking the AI for advice but cross-checking through search to challenge the results before I commit. A notch further right, still inside assistive mode, you’ve got the kitchen renovation where I’m taking the AI’s advice on what’s possible, but I’m picking the supplier myself, because the supplier relationship is one I’m going to live with for years.

Cross again into agent mode, and you’ve got the holiday I already know: the agent books the cheapest familiar hotel without consulting me, because I’ve stayed there before and it’ll be fine. 

Further right, the pragmatic sock purchase where I just need the right pair in the right size by Tuesday, and the brand doesn’t matter a great deal. And right at the far end, the taxi from A to B, where I genuinely don’t think about it, the agent picks the ride, the agent pays, and I get to my destination with zero fuss or thought.

Same person, eight different positions on the boundary. The variability is the whole AAO game: you have to have a strategy that will win the click at every stage across the Delegation Boundary. The boundary tracks risk, human preference, emotional weight, reversibility, and a half-dozen other things.

You have to wrap that into your strategy on top of the cascading confidence, the 10 gates, the entity home, the push-layer entry modes, the framing gap, and the funnel flip this series has been piling up since February 2026.

7 factors tell you where the delegation boundary sits in your category

Score yourself against seven broad factors (and add your own, if you like):

• Emotional weight: The more the purchase touches identity, family, or values, the harder to delegate.
• Domain expertise required: The more specialized the decision, the more users either delegate fully (because they know they don’t know) or refuse to delegate (because they think they do).
• Price relative to income: A $2 coffee delegates easily, a $20,000 car doesn’t.
• Purchase frequency: Habitual purchases delegate readily, one-offs need scrutiny.
• Reversibility: Returnable goods delegate easily, wedding venues don’t.
• Regulatory context: Financial, medical, and legal categories carry compliance constraints.
• Cultural context: Trust in agents varies by market and demographic.

Evaluate your category and you get a credible read of where your audience sits, how much friction they’re willing to hand to the engine, and where you need to focus your work. You’ve also done something less obvious, but even more important than scoring: you’ve grouped your audience by behavior at the decision moment rather than by demographic or category labels.

For two decades, we built paid and organic strategies based on how humans see the world: by category, by demographic, by geography. In Google Ads, five-star hotels in Bali went in one ad group, hostels in Bali in another, and five-star hotels in Thailand in a third, because that’s how a marketer thinks.

With AI, the engines stopped thinking that way. For Performance Max and AI Max, you absolutely must group by intent to get performance. Five-star hotels in Bali sit in the same cohort as five-star hotels in Thailand, because the intent (luxury accommodation) holds the cohort, and the geography doesn’t.

Gemini thinks in intent cohorts, not human categories, and the same lesson applies across Organic, ChatGPT, Claude, and other LLMs.

What matters here: AI groups by intent, AI builds the cohorts, and the brand optimizing against the old human structure is competing in categories the engines have left behind.

For years, Google representatives have told us that SEO (or assistive agent optimization, the newer discipline that extends SEO to cover assistive engines and agents) is all about intent, and the point they perhaps didn’t make clear enough is: that’s because it’s how AI thinks.

The user delegates, the engine commits

Here’s the point most brands haven’t seen yet: the user and the engine work in tandem, and your strategy needs to account for that.

Nothing happens without the user’s mandate. The agent doesn’t take over because the assistive engine has decided it’s time. The user pushes the boundary by handing the engine a mandate, and the engine acts within that mandate. The delegation boundary isn’t drifting toward the engine on its own: it is being intentionally moved by the user on a case-by-case basis.

Every engine has been delegated to, and delegation forces commitment. We hand search the job of finding the 10 best links for our query, and Search has to commit to which 10. We hand the assistive engine the job of recommending one brand from a comparison set, and it has to commit to which one. We hand the agent the job of executing the transaction, and it has to commit to which transaction.

Different layers of delegation, different consequences when the commitment is wrong, but the act of commitment is universal across the delegation spectrum, because every engine has been handed a job that ends in an answer (even if that answer is giving us a choice of 10).

The algorithms learn at three levels, and they learn differently across the three engines (search, assistive, and agential):

• Individual level: What this specific user has accepted, overridden, ignored, repeated. The narrowest scale, fully personalized. Heaviest in Agent mode (where the agent acts on this specific user’s behalf), significant in Assistive mode, negligible in Search.
• Cohort level: What users with this combination of intent signals have accepted on average across thousands of comparable journeys. Wider than the individual, narrower than everyone. Decisive in Assistive mode (the engine needs to know “what works for users like this one”), important in Agent mode as a fallback when individual signal is sparse, lighter in Search.
• Global level: What the algorithmic trinity has encoded about the brand from the world’s record. Not user behavior but encoded knowledge, stored in three places simultaneously: the LLM as parameters baked into model weights during training, the search engine as the indexed corpus and ranking signals, and the knowledge graph as entity nodes, relationships, and attributes. Three storage layers, three update cadences, three signal types, all encoding the same fundamental thing: what the AI has come to believe about the brand from everything brands publish and everything others say about them.

Three concentric circles. The user at the center. The cohort around the user. The global world around the cohort. The first two layers are about behavior: what users do with you.

The third layer is about knowledge: what the algorithmic trinity has encoded about you across LLM parameters, search index, and knowledge graph.

The brand has direct leverage at the third layer (everything you publish, everything written about you, everything that becomes part of the corpus the trinity reads from) and indirect leverage at the first two (through the experiences that shape what users do).

Important: The third level is the one that compounds the longest. Every won event Thomann has earned across every customer in every cohort over the last 28 years isn’t just training the cohort signal: it’s feeding the global priors that every future model trained on widely available data will inherit. 

Brand confidence work done today (and every “today” past and present) compounds into future models on substrates the brand never directly trains. That’s the deeper structural reason systematic effort compounds in the AI era. Train the substrate, not just the signal.

Why the cohort signal isn’t destiny, and where your opportunity lies

If the engines weighted what they had already learned as the only voices that mattered, no challenger brand could ever break in. 

The dominant brands would dominate forever, the individual histories, the cohort patterns, and the global priors would lock in, and AAO would be a closed game where only the incumbents win.

Luckily, that isn’t the reality. What the engine has already learned is heavy, but it isn’t absolute. The brand that establishes a stronger claim than the brands the engine recommends and advocates for can break in, and the layers the brand can move are cohort and global. (The individual layer can be influenced through other communications channels, but is effectively closed in the AI engines because they are increasingly walled gardens.)

The route in is to claim, frame, and prove. Make your claim, frame the claim, and prove it with evidence across enough sources that the engine has the corroboration to move you up. 

Underneath that work sits the understandability, credibility, and deliverability process:

• Understandability built through clear knowledge of who you are and what you do.
• Credibility built through N-E-E-A-T-T.
• Deliverability built through topical ownership. 

Get all three. Once your confidence clears the threshold:

• The cohort signal that currently favors your competitor stops being the only voice the engine hears.
• The global layer starts absorbing your contribution alongside the incumbents’.
• Your claim becomes framed and proved in a way the engine can resolve and prioritize. 

That’s the only way an outsider gets in. It’s also the route every dominant brand took to become dominant in the first place.

What wins at won is confidence

When push comes to shove, what decides the outcome at won is the confidence the algorithm has in your brand at the moment of decision. 

Confidence at the moment of:

• Ranking in search. 
• Recommending in assistive. 
• Action in agent. 

The mode, channel, and cohort change, but the question doesn’t.

Content and context are table stakes today. Content is what you’ve published, where, in what structure: every brand doing digital marketing seriously has been doing it for years, since Bill Gates wrote “Content is king” in 1996. 

Context is the match between your content and the user’s intent at the moment they’re asking. Both are table stakes, so neither differentiates at the recruitment, grounding, display, and won gates.

Confidence is what’s left, and confidence is what we can measure at the bottom of the funnel. Three things, you measure at the bottom of the funnel brand results (the due diligence rabbit hole):

• You measure the accuracy of the results about your brand.
• You measure the positive sentiment in the results about your brand.
• You measure the consistency of the results about your brand across all of the engines.

How right, how positive, how consistent: that is your measurement of confidence across the AI spectrum.

This is exactly where Thomann won my pedal purchase. ChatGPT was confident enough in Thomann’s information to commit to a Friday delivery on Thomann’s behalf, on a Tuesday, with my Saturday gig riding on it. 

Hundreds of European suppliers could, in theory, have gotten those pedals to me by Friday. ChatGPT picked Thomann because they had published the stock data, shipping times, warehouse logistics, country-by-country delivery commitments, returns policy, price, and credibility signals, in the kind of structured, accurate, consistent detail that makes an algorithm willing to put its name behind a recommendation. 

Confidence work laid down years before the moment ChatGPT needed it was compounded into a single won event in 15 minutes flat.

Who’s making the decision at won — and when

Map your intent cohorts. Score each one against the seven factors. Work out which mode each cohort will use to make which decision. Then train the seven AI employees (Google, ChatGPT, Perplexity, Claude, Copilot, Siri, and Alexa) to represent you with positive sentiment, accurate facts, and consistent narrative across every one of those moments, because they’re already working 24/7, they’re already talking to your customers, and the only question left is whether they’re recommending you or your competitor.

Untrained employees cost you money every day they’re untrained. Trained employees generate revenue every day they’re trained. Won is the moment you’ve trained them for, or the moment they’ve handed to your competition.

The next question is how you measure all of this, and it turns out my answer rewrites how brands should think about measuring AI-era search-assistive-agential success entirely. That’s the next article in the series.

This is the 11th piece in my AI authority series. 

• Part 1, “Rand Fishkin proved AI recommendations are inconsistent – here’s why and how to fix it,” introduced cascading confidence. 
• Part 2, “AAO: Why assistive agent optimization is the next evolution of SEO,” named the discipline. 
• Part 3, “The AI engine pipeline: 10 gates that decide whether you win the recommendation,” mapped the full pipeline. 
• Part 4, “The five infrastructure gates behind crawl, render, and index,” walked through the infrastructure phase.
• Part 5, “5 competitive gates hidden inside ‘rank and display’,” covered the competitive phase.
• Part 6, “The entity home: The page that shapes how search, AI, and users see your brand,” mapped the raw material.
• Part 7, “The push layer returns: Why ‘publish and wait’ is half a strategy,” extended the entry model. 
• Part 8, “How AI decides what your content means and why it gets you wrong,” covered annotation — the last gate where you’re alone with the machine. 
• Part 9, “Why topical authority isn’t enough for AI search,” opened the competitive phase proper with topical ownership.
• Part 10, “The funnel flip: Why AI forces a bottom-up acquisition strategy,” named the process.
• Part 11, “The framing gap: Why AI can’t position your brand” exposed the gap between evidence and recommendation.
• Part 12, “The 10-gate AI search pipeline: Find where your content fails,” showed you how to find (and repair) your F grades in the AI engine pipeline.
• Up next: The funnel query pathway: How to measure what matters in the AI era.

Read more at Read More