The Decision-Maker Has Moved
Introducing the Agentic Shelf, and why twenty-five years of “get found” strategy no longer determines who gets chosen
Tim de Rosen and Paul Sheals · July 2026
For twenty-five years, digital marketing rested on one assumption nobody said out loud: the customer was the one who decided.
Every technology that has mediated commerce has been dismissed, at first, as marginal. Advertisers said no one would sit for hours watching a box in the corner of the room. Retailers said no one would hand a stranger their card details over a wire. Both were wrong for the same reason: they mistook a new format for a smaller version of the old one. Something similar is happening now, and it is being dismissed the same way, as an extension of search, a new surface to optimise, a new set of keywords to chase. It is not that. It is a change in who makes the decision.
The Google-Era Decision Chain
For a quarter of a century, a purchase decision ran through a chain the customer controlled at every link. They typed the query. They read the metadata on the results page and judged which listing looked credible. They opened a page and read content a brand's own team had written and controlled. They went back, opened another, compared. At the end, they decided what to buy and who to buy from.
Every dollar spent on SEO, content, and conversion optimisation was spent influencing a human who remained, at every step, the one doing the choosing. Reaching the first page of Google mattered because more visibility in front of that chooser increased the odds they would choose you. That was the entire logic of the industry, and it was sound, because the assumption underneath it, that the customer decided, was true.
The Agentic Shelf
That assumption no longer holds for an increasing share of purchase journeys.
When someone asks an AI assistant for a recommendation, the model performs every step of that chain itself. It decides what information to retrieve. It decides what to compare. It decides how much context to share about competitors the user never sees named at all. It narrows, condenses, and reasons through a conversation, and at the end it offers a recommendation. The customer's remaining act is to accept it or decline it.
This is not a new channel. It is a new shelf, and it behaves differently from the two that came before it.
The agentic shelf is the first shelf in the history of commerce where a brand is judged, chosen, or discarded, and the company has no way to watch it happen.
That claim needs a precise reading, so it is stated precisely once, here, and held to for the rest of this piece. Passive observation, the equivalent of glancing at a shelf position or checking a search ranking, does not exist for the agentic shelf. There is no dashboard to look at. Deliberate testing does exist, and is described later in this article. The difference between the two is the difference between a shelf you can walk past and one you have to interrogate on purpose. Invisible, in this article, means the first. It has never meant the second.
The physical shelf, you could walk into a shop and see what was on it. The digital shelf, you could check your ranking and know exactly where you stood. Both were competitive, but both were passively visible; you could watch the contest happen without deciding to look. The agentic shelf offers no aisle to inspect and no ranking to check by habit. There is a conversation deciding commercial outcomes, at scale, every day, that no brand sees unless it goes looking.
What We Found
AIVO tested this directly. Across 1,427 brand probes spanning ten industries, 95.7% of brands were recognised by the models tested, brand knowledge was never the bottleneck. But recognition and recommendation diverged sharply. Of the brands present when a buying conversation began, 87.3% were displaced before the model delivered its final recommendation (95% CI 85.5–88.9%; “Beyond Visibility,” Sheals and de Rosen, AIVO Meridian Research, DOI 10.5281/zenodo.20761232).
That figure is not uniform across models, and is not presented as one. Gemini displaced 86.1% of previously present brands before the final recommendation. ChatGPT displaced 81.7%. Perplexity displaced 56.7%, a materially better result, consistent with its reliance on live retrieval at answer time rather than on knowledge fixed at training. The variance across platforms is treated in the underlying research as evidence for the mechanism, not as noise to be averaged into a single number.
The more consequential finding sits underneath the headline figure. In most displacement cases, the model was not unaware of the brand it dropped. It had demonstrated possession of the relevant facts earlier in the same conversation. What failed was not knowledge, it was the model's ability to carry that knowledge forward and deploy it at the moment it reasoned toward a recommendation. We call this the Linkage Gap.
Ruling Out the Obvious Objection
A fair challenge to this finding is that displacement might simply be correct filtering, a model reasonably narrowing its answer to criteria a brand no longer meets, rather than a failure to retrieve information it still held. If that were the explanation, the finding would describe the model doing its job well, not a gap worth closing.
The underlying research tested this directly, rather than assuming an answer. In a counterfactual follow-up conversation, researchers reintroduced one previously possessed fact about the brand at the exact moment of recommendation, without additional framing, advocacy, or argument for the brand's inclusion. Two distinct patterns emerged. In the majority of cases, the brand returned to the recommendation once the fact was made available, and did so consistently, evidence that the original failure was retrieval, not judgment. In a minority of cases in the test set, the brand did not return even once the fact was reintroduced, because the model held a structural judgment about the brand's category fit, treating a travel comparison tool as distinct from a holiday provider, for instance, that no single reintroduced fact could override.
These are two different problems, and this paper is careful to name both rather than collapse them into one statistic. The first, where a reintroduced fact changes the outcome, is the Linkage Gap: a retrieval and activation failure, and an infrastructure problem that structured evidence can close. The second, where it does not, is a Reasoning Gap: a genuine, often defensible, model judgment that no amount of evidence-seeding will change, and that this paper does not claim to solve. The 87.3% headline figure describes brand-level displacement generally. The Linkage Gap specifically, the addressable share of that figure, is the narrower and more commercially relevant claim, and it is the one this paper is making.
Why This Is Not a GEO Problem
The natural response inside most marketing organisations has been to treat this as an extension of search optimisation. Optimise the content, structure the pages, earn the citations, and the AI will find you the way Google once did. That instinct is not wrong. It is incomplete.
Generative engine optimisation, and the visibility tools built on it, answer a real and necessary question: does the model know a brand and mention it. That work still matters. A brand invisible to the model at the first turn has no chance of surviving to the last one. But GEO was built to influence a human chooser reading a results page, and on the agentic shelf, the results page has been replaced by a conversation the model conducts with itself. Mention is no longer the decision point. It is one input into a reasoning process the brand cannot see and the visibility tool cannot measure.
This is the precise sense in which twenty-five years of “get found” strategy has stopped being sufficient. It has not stopped mattering. It has stopped being enough, because the thing it was built to win, a human's attention on a page, is no longer where the decision gets made.
None of this happens in a commercial vacuum, and it would be incomplete to describe the agentic shelf as a purely disinterested reasoning process. AI platforms are still early in monetising these interactions, and sponsored placement, affiliate arrangements, and paid promotion inside agentic recommendations are already emerging alongside organic reasoning. A brand with comprehensively structured, machine-readable evidence can still lose a recommendation turn to a competitor who has paid for preferential placement. Structured evidence closes the Linkage Gap. It does not, on its own, counter a paid arrangement. Brands operating on this shelf should expect to need both levers, not one.
The Room That Was Closed
Everything above describes a genuine loss of passive visibility. Brands that could once glance at their own competitive position, on a shelf, in a search ranking, cannot glance at the conversation that now decides their commercial fate.
But unglanceable is not the same as unmeasurable. The same conversation that decides a brand's fate can be deliberately tested, by asking the questions real customers ask, in the sequence they ask them, and recording what the model does at every turn: which brands it reaches for first, where in the conversation a brand is dropped, what the model says about a brand when no one is checking, which competitor it chooses instead and on what basis, and, critically, whether a dropped brand returns when its evidence is made available at the decision turn or whether it does not.
None of that needed to be actively tested under the Google-era model of measurement, because the customer was doing the deciding, and the customer could be surveyed, interviewed, and observed directly. On the agentic shelf, the decision-maker is the model, and the model has to be interrogated on its own terms, across the full arc of a conversation, not sampled at a single prompt, and not assumed to be reasoning fairly without being tested for it.
A Note on Limitations
This paper's central figures describe brand-level outcomes aggregated across platforms and industries, and readers evaluating the finding for their own category should weight the per-platform breakdown above more heavily than the aggregate. The Reasoning Gap distinction was established on a subset of the full probe set; broader replication of that specific test, at the full 1,427-probe scale, is ongoing and will be reported separately. The findings describe displacement mechanics as tested; they do not, on their own, quantify the separate and additive effect of paid placement discussed above, which is a distinct research question this paper does not attempt to answer.
What Companies Need to Ask Now
The practical implication is a change in the question companies bring to their AI strategy. The question has not been “are we visible” for some time, even if most organisations have not caught up to that fact. The question is whether a brand's presence early in a conversation survives to the point where the model commits to a recommendation, and if it does not, why not.
Brands that spend the next year still optimising for the shelf the customer has stopped looking at will be measuring a contest that no longer determines the outcome. Brands that start watching the agentic shelf itself, treating the reasoning conversation as the thing to be measured and won, will be operating with information almost no one else in their category has.
The shelf did not disappear. The ability to see it did.
That is the gap this Journal, and the research behind it, exists to close.
References
Sheals, P. and de Rosen, T. (2026). Beyond Visibility: The Linkage Gap and the Case for a Third Layer of AI-Native Brand Infrastructure. AIVO Meridian Research. DOI: 10.5281/zenodo.20761232.
de Rosen, T. (2026). The Linkage Gap: Brand Displacement Across Multi-Turn AI Buying Conversations. AIVO Standard Working Paper WP-2026-14. Zenodo.
de Rosen, T. (2026). The Human Agentic Gap. AIVO Journal. DOI: 10.5281/zenodo.20920098.
de Rosen, T. (2026). Revenue at Risk from AI Displacement (RaR-AID). AIVO Standard Working Paper WP-2026-19. Zenodo.
tim@aivostandard.org