If you track AI visibility for a brand, you have probably watched a number fall this year without being able to say why. Nothing changed on your site. Nothing changed with your competitors, as far as you could tell. The share of AI answers that named you just drifted down and stayed there.
Here is the most likely explanation, and it is not about you. The answers got shorter. The engines started naming fewer brands per response, and when a list goes from seven names to three, somebody has to fall off. Twice in the last year, that somebody was most of the market.
On 18 October 2025, ChatGPT started treating brands as entities. In plain terms, the model got a structured sense of which company is which, rather than matching names as strings of text. Profound tracked millions of prompts across the change and the effect on answers was immediate.
The average number of brand mentions per response dropped from six or seven to three or four. Average visibility across the brands they tracked fell 31 percent. More than 85 percent of brands saw a decrease. About 13 percent took a large hit, more than ten points of visibility. Another 58 percent lost somewhere between two and ten.
Two things about that are easy to miss. First, the gap between the top spot and tenth place narrowed by roughly 23 percent. The engine did not just cut the list. It compressed the difference between the brands that survived the cut, so being number one bought less than it used to. Second, the citation side moved at the same time. Wikipedia and Reddit pulled further ahead as sources, and the long tail of everything else stretched out.
The Summer 2026 Profound Index, built from more than 1.9 billion real conversations across 50-plus industries, recorded a second contraction. After a June ChatGPT update, citations per answer fell 10.4 percent, and the count of unique domains cited dropped in 56 of 58 industries tracked.
That second figure is the one worth sitting with. It was not a shift in which industries got attention. It was nearly every industry, at once, getting a shorter source list. Fewer citations per answer and fewer distinct sites cited means the engine is consolidating trust on a smaller set of pages, and the pages that were marginal before are now gone.
The number of seats in a given answer depends heavily on which engine you are looking at. Profound's citation overlap study, run on 100,000 prompts across the major platforms in mid-2025, counted how many distinct domains each engine cites in a typical response. A later study in May 2026 tracked 15,155 brand configurations daily across Google's three AI surfaces and counted brands rather than domains. The two are not the same measurement, but together they give a fair picture of how much room there is.
Copilot is the tight one. Two or three sources per answer leaves almost no room for a second-tier brand. AI Overviews and Perplexity are roomier, though the Google number has been drifting lower and the ChatGPT figure predates both of the cuts described above. If you are tracking a single blended "AI visibility" score across engines, it is averaging over rooms of very different sizes, which is one reason those blended scores are hard to act on.
Not everyone lost. The same Summer 2026 index found that in the median industry, nine of the ten leading brands held their position quarter over quarter. Contraction hit the middle and the bottom of the list far harder than the top. A few brands moved a long way: one cybersecurity company nearly doubled its share of category mentions in a single quarter, and an e-commerce platform went from nineteenth to sixth. Those are the exceptions, and Profound's data does not say why they moved, so neither will this piece.
What the data does support is that the leaders are sticky. The brands the engines already trust survived two rounds of list-shortening almost untouched. Everyone else is now competing for fewer seats than they were a year ago, against incumbents who did not lose theirs.
None of this is a reason to give up on AI brand mentions. It is a reason to be more deliberate about where the effort goes. Each of the following comes from a specific Profound finding, not from general SEO instinct.
Across 11.84 billion citations analyzed between April and July 2026, 57 percent went to brand-operated websites. On Gemini it was 69 percent. Even ChatGPT, the lowest, still drew 47 percent of its citations from company sites. When the seat count shrinks, the engine keeps the sources it trusts most, and the biggest single bucket of those is your own domain. A clear, current, well-structured page on the exact question is still the most reliable way into the shortlist.
In roughly 730,000 ChatGPT conversations from late 2025, the first turn was two and a half times more likely to produce citations than the tenth turn, and about four times more likely than the twentieth. Conversations get less sourced the longer they run. If your content only answers the deep follow-up, it is competing for the turns where the engine has mostly stopped citing anybody.
Profound's overlap study found that only 11 percent of citations were shared between ChatGPT and Perplexity; nearly 89 percent were unique to one or the other. Google AI Overviews and Copilot overlapped just 6 percent. A page that lands in that shared sliver is doing something structurally right that all the engines reward. Pages that show up on one engine only are often riding a quirk of that engine, and quirks get patched.
Between June and July 2025, 40 to 60 percent of the domains cited for identical prompts changed, depending on the engine. Over six months the drift ran 70 to 90 percent. That is the baseline noise you are measuring against. A separate Profound experiment found that checking a prompt ten times a day instead of once improved visibility precision by only about 10 percent. The composition of the prompts you track mattered more than how often you ran them. Build a prompt set that represents real buying questions, then read the monthly line.
Two and a half seats on Copilot is a different contest from seven or eight on AI Overviews. A brand that is fourth in its category has a real chance of being named on Perplexity and almost none on Copilot. Rather than a single blended score, know which engine your customers actually ask and treat each one as its own race with its own seat count.
A visibility drop that lands on the same week as a platform update is not a verdict on your content. It is a verdict on how many seats exist. The honest way to read your trend line is against the seat count: are you holding a share of a smaller list, or losing share of a list that stayed the same size? Those need different responses, and most dashboards cannot tell them apart.
Scoring each page against the things the engines reward, then running your actual buyer prompts across ChatGPT, Gemini, Perplexity and Claude on a schedule, is how you separate the two. That is roughly what Prism Optimizer does, but the method matters more than the tool. Track seats, not just mentions, and the drops start making sense.
My own read: the shortlist keeps shrinking. Every engine has a commercial reason to give a cleaner answer with fewer names, and none of them has a reason to go back. Plan for three seats, not seven.
See how many seats you hold.Put your site through the Prism and get a page-by-page read on where you are cited, where you are absent, and what to fix first.
Request a demo