Why we asked the people who built search
The practitioners who built search engine optimisation (SEO) are the first to feel the ground move. To understand how AI answers are changing demand, the PRISM team interviewed senior search leaders and practitioners across agencies and in-house teams. Their accounts converge on a small set of findings that matter to any marketing leader.
The headline is uncomfortable but consistent. Good search performance is no longer a reliable proxy for visibility in AI answers, and the practitioners closest to the data were the first to see it.
Finding one: search is necessary but no longer sufficient
Clean search foundations still matter, and arguably matter more, but they no longer guarantee the outcome. Practitioners report that a crawlable site with sound structured data is the entry ticket: if an AI crawler cannot read your pages, the model never considers you. AI crawlers are less forgiving of messy markup than search crawlers.
Yet sound foundations are not enough. In our pilot work comparing the same queries, only around a third of the pages ranking in a search engine’s top ten overlapped with the pages an answer engine actually cited. The model was drawing the majority of its citations from elsewhere.
Finding two: the dashboards stay green
The most dangerous pattern practitioners describe is the dashboard that keeps rising while the result that matters falls. Search metrics were once leading indicators: what they showed today predicted pipeline in a quarter. They are now closer to lagging indicators of a buyer behaviour that has already shifted.
The effect within a company is two teams, both right. Marketing reports a healthy quarter from the search dashboard. Sales asks why the qualified pipeline is soft. They are measuring different worlds.
Finding three: third parties shape the answer, not your site
Practitioners have observed that answer engines frequently rely on third-party sources instead of the company’s own webpages. When a model makes a statement about a company, the cited authority is often a review site, a community discussion, an aggregator, or an analyst document. In some cases, the information may be several years old.
This reframes the approach. A company that identifies a single unifying thread within a community can significantly influence how AI describes that community. Instead of implementing numerous minor changes, the focus should be on one key improvement. The change lies in what the model recognises as the authority in your category. It’s not about publishing more pages on your own website.
Finding four: one model is not the market
Optimising for a single model is now a strategic risk rather than a convenience. Practitioners point to fragmentation. By April 2026, ChatGPT’s share of AI chatbot web traffic had fallen to about 55%. In early 2025, it was about 77%. Gemini, over the same period, rose from under 6% to about 27% (Similarweb data, April 2026, via Momentic). That share counts visits to the chatbot websites themselves; it excludes mobile apps and API use. Regional models matter in regional markets — Qwen in its home markets, for instance.
Because each model treats the same query differently, a company can be dominant on one and nearly invisible on another. Single-platform measurement reports only part of the picture.
The compounding-loss thesis
Question: If AI visibility is still emerging, why treat it as urgent rather than wait and see?
Answer: because the loss compounds. Each time a model is retrained, the patterns it has learned about companies are reinforced or shifted. Companies with strong signals at training time become the model’s default understanding of the category. Once that association settles, displacing an incumbent becomes harder.
This is the thesis that shaped PRISM. Visibility today becomes the model’s assumption tomorrow, so the cost of waiting does not stay flat. It grows as competitors accumulate the signals that future models will treat as settled fact. The team is candid that the compounding rate is not yet precisely known. The direction across all observed updates has been towards greater concentration on incumbents.
What practitioners are doing about it
The practitioners adapting fastest share a pattern. They keep search foundations clean, then add work that search never required. That work has three parts:
-
Persona-aware content — written for the questions buyers actually ask AI.
-
A deliberate citation strategy — aimed at the third-party sources models trust.
-
Multi-model measurement — spanning several answer engines, not one.
The people who built search keep reaching the same conclusion. Search best practice has become the price of entry. It no longer counts as a strategy on its own. For the framework that turns these findings into a single score, see “Search is no longer being optimised. It is being answered.” For the scoring detail, see the companion methodology piece “AEO with PRISM – How we score each of the five pillars.”
