It Already Knew
The market is done with diagnostics. It wants the gap closed, and closing it requires knowing exactly what kind of gap it is
The head of GEO at one of the world's largest beauty companies said it to us directly. His management is done with diagnostics. Another dashboard, another score, another 40-page report telling them what they already suspect, that isn't what they're asking for anymore. They liked our decision-stage analysis more than anything else they'd seen. What they actually want is someone who fixes it.
That is the correct instinct, and the industry serving them has not caught up to it. There is a lot of research right now on how ChatGPT's search actually works underneath the interface, which backend source serves a given query, how often that routing changes. It deserves attention. It is also not the layer that determines whether a brand gets recommended, and treating it as though it were is part of why this market is still stuck producing diagnostics when its actual clients are asking for interventions.
A Real Question, Asked About the Wrong Layer
Independent research published in the past week traced ChatGPT's search telemetry across a thousand prompts, run ten times each, and found that the backend retrieval source is not fixed. Most prompts route consistently to the same source. A meaningful minority do not, and when the backend changes, the set of URLs and domains retrieved changes substantially alongside it. There is no single, stable “ranking in ChatGPT.” There are several partially overlapping retrieval ecosystems.
That finding describes what happens when a model goes looking for information on the open web at the moment a query is asked. It has almost nothing to do with why a brand a model already knows well still loses the recommendation. Conflating the two is how an entire category of vendors ends up optimising the wrong layer while their clients quietly lose patience.
Two Different Failure Modes
In a large share of the buying conversations we test, the model already has the relevant facts about a brand before any search happens, carried in its training rather than fetched at query time. The question that determines the outcome is not whether the model can find something out there. It is whether the model deploys what it already knows at the specific turn in a conversation where the decision gets made.
These are not the same problem, and the difference is not academic. If the failure were retrieval noise, the fix would be about improving what gets indexed. If the failure is deployment, indexing more content changes nothing, because the model was never short of the fact in the first place. Our research says the second failure mode is the dominant one, and we can show it, not just assert it.
What We Actually Found
Across 1,427 brand probes spanning ten industries, 95.7% of brands tested were recognised by the models we ran against them. Possession was never the bottleneck. 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%). Gemini displaced 86.1% of previously present brands, ChatGPT 81.7%, Perplexity 56.7%. Perplexity's materially better result tracks its heavier reliance on live retrieval at answer time, itself evidence that retrieval matters only where the model is actually retrieving, which most of the time, it is not.
The model was not unaware of the brand it dropped. It had already demonstrated it knew. It simply didn't carry that knowledge to the turn where the decision was made.
We tested whether this was retrieval failure or something else directly. In a counterfactual follow-up, we reintroduced one previously possessed fact about a displaced brand at the exact moment of recommendation, with no additional framing or argument for the brand's inclusion. In the substantial majority of cases, the fact changed the outcome. That is the Linkage Gap: a deployment failure, not a knowledge failure, not a retrieval failure. No new search occurred. The fact was already there. It simply had not been surfaced at the turn that mattered.
A minority of displaced brands did not return even once the fact was reintroduced. That is a different problem, a genuine model judgment about category fit, and no amount of evidence changes it. We call it the Reasoning Gap, and we are not in the business of pretending it doesn't exist or claiming to fix it. The distinction matters because it tells a brand exactly what kind of gap it has, and only one of the two is what remediation closes. Most of the 87.3% is the first kind. That is precisely why closing it is worth doing at scale, not despite the distinction, because of it.
The Market Has Already Moved Past Diagnosis
Almost everything sold under the banner of AI visibility or generative engine optimisation stops at measurement. It will tell a brand whether it is mentioned, how often, and how that compares to competitors. That was useful once. It is no longer enough, and the clients buying this work are the ones saying so, not us. A head of GEO at a major beauty company does not need another study telling him his brand has a gap. He needs the gap closed, brand by brand, SKU by SKU, before his management stops asking for the diagnostic altogether and starts asking why nobody delivered a fix.
We don't stop at the diagnostic. Once the Linkage Gap is identified for a specific brand and category, our system maps exactly what evidence is missing at the point of decision, sources that evidence from the brand's own repositories where it already exists, flags what has to be created where it does not, and seeds the resulting evidence directly into the high-authority sources these models draw from. This runs automatically once the pipeline is in place. It is not a report. It is an intervention, and it is the thing the market is now explicitly asking for.
Where This Is Live Today
This is not a framework waiting for a first customer. We are directly engaged in beauty, working with brand teams on the specific evidence gaps their own audits surface, closing them, not just reporting them. In healthcare, automotive, and financial services, our measurement runs underneath live client work through our partnership with a global market research company, including a top-five US bank audited end to end across ChatGPT, Gemini, and Perplexity.
Beauty is a direct AIVO engagement. Healthcare, automotive, and financial services are engagements where our measurement is the instrument underneath. Both are real. A brand in any of these categories should know which relationship applies to them, and should ask their current AI measurement vendor a direct question: after the diagnostic, what actually gets fixed, and who does it.
The Throughline
Retrieval infrastructure is a real layer of how these systems work. It is not the layer that explains why a brand a model already knows well still loses the recommendation, and it is not what the industry's actual buyers are asking for solved. They have been diagnosed enough. The work now is closing the gap, automatically, brand by brand, before the patience for another dashboard runs out entirely.
Sheals, Paul and de Rosen, Tim