Brand is the only compounding asset in agentic search

Sundar Pichai told Patrick Collison in April that "a lot of what are information-seeking queries will be agentic in Search. You'll be completing tasks. You'll have many threads running." Two months later, on Decoder, he committed to the opposite promise: "Everything we do across all, you will see us five years from now sending a lot of traffic out to the web."

Both things are being said by the same person, in the same product cycle, about the same surface. The first describes a world where the user never visits your site because an agent visited it for them. The second promises that your site will still get visits. The gap between those two statements is not a contradiction Google has failed to notice — it is the gap inside which every business currently doing SEO is being asked to make five-year investment decisions.

I have spent eighteen years watching the industry react to shifts like this. The pattern is consistent. A platform announces a direction. The optimisation industry repackages the platform's own talking points as a methodology. Vendors emerge to sell tools that measure the new thing. Two years later, everyone discovers that the talking points were partially true, the methodology was partially nonsense, and the only people who actually benefited were the ones who had already built something the platform couldn't easily replace.

That last part is the argument of this piece. In the agentic web — the one where search becomes an agent manager, where Chrome browses on your behalf, where Deep Research compiles reports without you ever clicking through — there is exactly one asset that does not commoditise. Everything else does. And almost nobody in the SEO industry is talking about it in those terms, because the industry's economic incentives push toward selling services that do commoditise. Brand is the only compounding asset in agentic search, and the industry's silence on this is the loudest tell about who the optimisation conversation is actually structured to benefit.

The state of play, briefly

Three Google executives have now said versions of the same thing in the last eight weeks. Pichai described Search as an agent manager. Nick Fox, the SVP who oversees Search, Ads, and Commerce, told Ben Smith at Google Marketing Live that "the way to optimise for AI search is the same way to optimise for search. Create great content." Liz Reid, the VP of Search, told publishers the rules are the same: make content people actually want to read, don't be the 1,000th copy, don't block the crawlers.

I've already written about why "create great content" is a tautology when said at that altitude, and why the "one playbook" framing papers over real shifts in how the product behaves. I won't repeat those arguments here. What I want to do instead is take the convergence claim seriously — really seriously — and ask what it implies about which assets compound in this environment and which ones don't.

Because if Pichai is right, and agentic search is the direction, then the question every business should be asking is not "how do I rank in AI Mode." It is: what do I own that an agent cannot bypass, cannot summarise, cannot route around, and cannot replace with a cheaper substitute?

The answer is uncomfortable for most of the industry, because the answer is not a tactic.

---

What agentic search actually does to discovery

Start with what the agent is doing on behalf of the user. A Deep Research query fires off a batch of sub-queries, pulls the pages that recur across them, and assembles a report with citations. AI Mode handles the longer, more exploratory questions that don't fit a single click. Chrome auto-browse fills forms and completes bookings. In each case, the user delegates the decision-making process to a system that reads more, faster, and decides what to surface.

The consideration phase is being automated. The question is which assets survive that automation.

What the agent is doing economically is harder to look at directly. It is performing exactly the function that, until now, a buyer would have performed by visiting four or five sites, reading reviews, comparing options, and forming a preference. The agent is doing the consideration phase of the funnel on the buyer's behalf.

This matters because the consideration phase is where most marketing actually works. The awareness phase creates demand. The consideration phase decides who captures it. SEO, content marketing, comparison content, review hunting, category education — almost everything the optimisation industry sells exists to influence what the buyer does between "I have a problem" and "I'm choosing this vendor."

When an agent runs that phase, it doesn't read content the way a person does. It retrieves, weighs, summarises, and presents — and the presentation is shaped by what the model is confident about, what it has seen repeatedly across sources, and what carries enough specificity that it cannot be confidently generated from training data alone.

The consideration phase is being automated. The question is which assets survive that automation.

Three things, broadly, do not survive it. Generic informational content gets summarised and stripped. Comparison content gets ingested and replaced by the agent's own comparison. Category education gets absorbed into the model's prior knowledge and reproduced without attribution. If your traffic and pipeline depend on any of those three, the agentic web is bad news, and no amount of schema markup is going to fix it.

What survives is everything that carries something the model cannot generate on its own. Jono Alderson has been saying this for over a year, and Nick Fox echoed it almost verbatim at GML: original data, first-person experience, named-entity specificity, takes the model is not confident enough to produce. That is the floor.

But there is a ceiling above that floor, and the ceiling is what nobody wants to talk about, because it is harder to sell as a service.

The ceiling: what agents trust by default

Watch a Deep Research run. Watch what gets cited and what doesn't. The pattern, repeatedly, is that the model defaults to sources it has seen before, in volume, across contexts, with consistent characterisation. Wikipedia. Major publications. Brands the model has encountered enough times to have internalised a stable representation of.

This is not a Google policy. It is not a tweakable parameter. It is how large language models behave under uncertainty. When the model is unsure what to retrieve or which source to weight, it falls back on density and familiarity. The signal is essentially: how often has this entity appeared in training data, in what contexts, with what associations, alongside what other entities. That is a working definition of brand strength from the inside of a model.

The Cornell research on retrieval poisoning makes this from the opposite direction. Thirteen words on a recurring community page can plant a brand into 38–51% of AI-generated reports on a topic. Why? Because the agent is looking for entities it can resolve, and the planted entity becomes the path of least resistance. The defence against poisoning is not technical — researchers tested three approaches and none worked without degrading the product. The defence is recognising which entities are legitimately referenced across enough independent sources that the model develops a stable representation that a single planted mention cannot dislodge.

That is brand. That is what brand does inside a retrieval-augmented system. It is the prior the model leans on when retrieval is ambiguous, contested, or poisoned. It is the reason ChatGPT cites Reddit at one rate and your blog at another. It is the reason, when Google tweaked an API and Reddit lost two-thirds of its ChatGPT citations, the citations rebounded — because the brand-level association between Reddit and community opinion was strong enough that the model kept reaching for it as soon as retrieval re-stabilised.

Brand is the prior. Everything else is a query against the prior.

Why the industry won't say this

If brand is the moat, the optimisation industry has a problem, because the optimisation industry does not know how to sell brand-building.

It knows how to sell audits. It knows how to sell content packages, link-building retainers, technical fixes, schema implementations, and increasingly, "GEO" services that are largely SEO with a different deck. These are all things that can be delivered on a 30-day cycle, billed monthly, and reported on with a dashboard. They produce visible activity. They make the agency look busy.

Brand-building does not work like that. It is slow. It is hard to attribute. It compounds across years, not months. It is more about what you say no to than what you say yes to. It often looks, from the outside, like the company is not doing very much — until five years later, when the company has become the default mental shortcut for its category and competitors cannot understand why their identical product cannot win the same deals.

This kind of work does not fit a retainer structure cleanly. So the industry, in aggregate, talks about it less than it should. Every agency that publishes a "2026 GEO playbook" will mention E-E-A-T, structured data, original research, and probably some variation of "build topical authority." Almost none will say the quiet thing out loud: that the businesses winning in AI search are mostly the businesses that were already winning in their categories, and the reason they're winning is not their schema implementation. It is that ten years of brand-building put them into the model's prior with enough density that the retrieval layer reliably reaches for them.

This is not a criticism of agencies trying to do good work. It is an observation about what the economic structure of the industry rewards. If you sell a service, you have to be able to describe its deliverable. "We built your brand" is not a deliverable. "We published twelve pieces of thought leadership and earned three high-affinity placements" is. The deliverables look like tactics. The tactics are not the thing.

The Fishkin point, generalised

Rand Fishkin's recent SparkToro piece on audience affinity vs. traffic gets at this from a different angle. He found, looking across eight industries, that "hidden gem" publishers with 5K–10K monthly visitors delivered 1.7× higher audience affinity than major outlets with 130× the traffic. The takeaway in his framing is that earned-media strategy should weight relevance, not just reach.

The deeper takeaway, in the framing of this piece, is that the brand signals that matter to a model are not the same as the brand signals that matter to a marketer's vanity. A placement in a 500K-visitor mainstream publication might generate a backlink and a brief traffic spike. A placement in a 10K-visitor niche outlet that your buyers genuinely read does something different — it strengthens the association between your brand and the specific topic, audience, and context in which you want to appear. The model is paying attention to the second kind of signal far more than the first.

This generalises. Every brand-building activity has two layers: the audience-facing layer (what your buyers see) and the model-facing layer (what gets ingested, embedded, and associated). The two layers used to be tightly correlated, because what your buyers saw was, broadly, what the web indexed. They are decoupling. A podcast appearance might reach 800 listeners, but the transcript might be ingested into training corpora and shape how your brand appears in retrieval for years. A single piece of original research might be cited by twenty other sites, each citation strengthening the brand–topic association inside the model.

The implication is that earned media, original research, expert positioning, and consistent topical association across high-affinity outlets are doing double duty. They reach your audience and they configure how the retrieval layer represents you. The tactics that look like classical brand-PR work are also, structurally, the strongest possible GEO play. Nobody is calling them that, because the GEO conversation has been hijacked by vendors who'd rather sell you a citation-tracking dashboard.

---

What this means for measurement

The hard part of all this — and the part where I want to be honest — is that brand-as-prior is genuinely difficult to measure. You can track brand searches. You can track unaided recall in surveys. You can track citation share in AI answers, with all the caveats about fake crawler traffic and Search Console's limited AI impression data. You cannot, however, directly measure "how stable is our representation inside GPT-4's weights." That is not a thing you can dashboard.

This is part of why the industry avoids the topic. SEO grew up on measurability. Keyword rankings, organic sessions, conversion rates, attribution models — everything got pulled into a measurement stack that, however imperfect, produced numbers you could put in a slide. Brand-building does not produce those numbers on the same cadence. It produces an outcome you observe years later when you notice that competitors who were equivalent five years ago are now bidding 40% higher on the same keywords because their click-through rates are lower than yours.

I am not going to pretend there is a clean measurement framework here. There isn't. The honest answer is that brand-building in the agentic search era requires a longer attention span than most quarterly-reporting cultures can sustain, and the businesses that will benefit most are the ones whose owners can hold a five-year view and resist the demand for monthly proof that the brand work is "working."

The closest thing to a leading indicator is this: monitor how AI systems describe your brand and your category, unprompted. Not "do you appear in citations" but "when ChatGPT or Gemini is asked an open question in your space, what entities does it reach for, in what order, with what characterisation?" If your brand appears with stable, accurate, positive associations across multiple models and multiple query formulations, the prior is solid. If it appears inconsistently, gets confused with competitors, or shows up with characterisations you didn't write, the prior is weak and you have work to do.

That work, again, is not schema markup. It is sustained, consistent, multi-channel presence in the contexts where your category gets discussed.

The counterargument I want to take seriously

The strongest pushback on this argument is that brand-building has always been the moat, in every channel, in every era, and saying so in the context of AI search is just dressing up an old truth in new language.

Brand is not newly important — it is newly unsubstitutable.

That is partly fair. The general principle that strong brands compound is not new. What is new is the magnitude of the effect and the loss of compensating mechanisms. In classical SEO, a strong technical foundation, smart content strategy, and aggressive link-building could beat a stronger brand in the SERP for a long time. There were paths up the mountain that did not require brand strength as a precondition. You could win on tactics.

In agentic search, the tactical paths are narrower. The model is not ranking ten blue links and letting the user decide. It is making the decision and presenting an answer. The mechanism by which a weaker brand could outflank a stronger one — by ranking higher on a high-intent keyword the stronger brand had ignored — still exists, but it is compressed. The agent reads more sources, weights them by familiarity and consistency, and presents fewer options. Tactical wins are still possible, but they have shorter shelf lives and produce smaller deltas.

So yes, brand has always mattered. What is new is that the substitutes for brand have weakened. The tactical SEO playbook used to be a meaningful counterweight for businesses without strong brands. In the agentic environment, the same playbook produces less leverage. Brand is not newly important — it is newly unsubstitutable.

The second counterargument is that this position is convenient for incumbents and harsh for challengers. Also partly fair. The agentic web does favour established brands. New entrants face a steeper climb than they did five years ago, because the model's prior is, by definition, built from the past. The honest implication is that challenger strategies in this environment need to be more concentrated, more distinctive, and more willing to forgo broad reach in favour of dominant association in a narrow niche. Generic challenger strategies — "we'll do what the incumbent does but cheaper" — do not work in an environment where the model has never heard of you.

What this actually looks like as work

The practical implications, if you take this argument seriously, look less like an SEO programme and more like a sustained editorial and PR operation with technical support.

Publish original research, on a cadence, in your category. Not because original research ranks well — though it does — but because original research creates citations, citations create entity associations, and entity associations are what the model's prior is built from. The Fishkin piece I linked above is itself an example: SparkToro is not the highest-traffic brand in audience research, but every report they publish strengthens the association between SparkToro and audience data done with methodological transparency. That association is the moat.

Show up consistently in the high-affinity outlets in your space, even when the traffic looks unimpressive. Podcasts, niche publications, expert roundups, conference talks, community forums where your buyers actually congregate. Each appearance is a vote for your brand–topic association in the model's training data, in addition to whatever immediate audience value it produces.

Use your own platforms — your site, your newsletter, your social presence — to establish positions that are yours. Not category-summary content the model can produce without you. Positions the model has to retrieve specifically from you because the take, the data, the experience, or the framing does not exist anywhere else.

Audit your brand presence inside AI systems quarterly. Run twenty unprompted queries in your category across ChatGPT, Gemini, Claude, and Perplexity. Note which entities appear, in what order, with what characterisations. Look for instability, confusion with competitors, or misattribution. Those are the signals you need to address with sustained brand work, not with a one-off content sprint.

Resist the temptation to spend your AI search budget on tools that promise to measure citation share. Most of them are log analysis with a markup, the data underneath them is unreliable for reasons I've written about elsewhere, and the budget is better spent on the work that strengthens the prior than on the dashboard that measures the prior's current state.

Close

The convergence of search and agents that Pichai described, and that Fox confirmed, is real. The "same playbook" line is also, in a narrow sense, true — the fundamentals that have always driven discovery still drive discovery. What the convergence statement obscures is which fundamentals matter most, and how the weighting has shifted.

The fundamentals that matter most in an agentic environment are the ones that configure how the retrieval layer represents your brand. Schema helps. Technical foundations help. Original content with a real point of view helps. But the compounding asset — the one that survives every shift in retrieval architecture, every Google update, every new agent product, every model retraining — is the brand. Not brand in the trade-dress sense. Brand in the sense of consistent, repeated, multi-source association between your name and the thing you want to be known for.

This is uncomfortable for an industry that grew up selling tactics. It is uncomfortable for businesses that want a 90-day deliverable. It is uncomfortable for me to write, because the consultancy work I sell also looks like tactics when you describe it on an invoice. But it's the argument the evidence pushes me toward, and not making it because it inconveniences my own pitch deck would be exactly the kind of self-interested industry posture I started this piece complaining about.

Brand is the only thing the agentic web can't commodify. Everything else — every tactic, every tool, every dashboard, every framework — is a query against the prior. Spend your next five years building the prior. The rest of the work gets easier when you do.