Before an AI answer is built, the prompt gets broken into searches you never see. On ChatGPT, 91 percent of those searches are one-offs. Your headings have to work for the question the person asked and the query the machine actually ran.
Most advice on how to rank in AI Overviews treats the person's question as the thing you are optimizing for. Write the answer, structure it well, get cited. That is half the job. The other half is the part Google described in May 2025 when it explained that AI Mode works by "breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf." The engine does not search for the question. It searches for a handful of rewrites of the question, and those rewrites decide which pages are even in the running.
Google calls the technique query fan-out. It has not published what those rewritten queries look like. Profound has, for three other engines, by tracking 10,000 prompts over 14 days in the spring of 2026 and logging every search ChatGPT, Perplexity and Copilot fired behind the scenes. The patterns are different enough from each other that "optimize for AI" turns out to mean three different things.
The single most useful number in the study is word overlap: how much of the person's original wording survives into the search the engine actually runs. On Perplexity, 88 percent of the words carried over. On Copilot, 50 percent. On ChatGPT, 13 percent. That is not a small spread. Perplexity is essentially searching for what you typed. ChatGPT is searching for something it composed, and it composes a different one each time.
Question shape changes too. Give Copilot a "how do I" prompt and 93 percent of the time it searches for a keyword string with no question in it. Perplexity does the same 78 percent of the time. ChatGPT is the odd one out: it keeps the question as a question in about 31 percent of searches, and rewrites the rest into something else entirely.
Even the word "best" is not safe. A prompt framed as "best [product]" stayed in that form 65 percent of the time on Perplexity, 52 percent on Copilot and 39 percent on ChatGPT. The rest of the time the engine dropped the superlative and searched for the product category, a comparison, or a feature.
A fair objection: this study did not measure Google. Google has confirmed the fan-out technique for AI Mode and for Deep Search, which "can issue hundreds of searches," but has not said how AI Overviews rewrites a query, and Profound's tracking did not include it. What we do know from Profound's other work is that the engines' source choices overlap far less than most people assume. ChatGPT's cited sources matched Google's results only 39 percent of the time in a 250-million-answer analysis, and its alignment with Google's index rose from 12 percent to 33 percent across 2025, so the two are converging but still mostly independent.
So treat the three-engine study as a floor, not a ceiling. If Perplexity, the most literal engine, still rewrites a third of "best" prompts, Google's fan-out is doing at least that much. And the practical response is the same regardless of which engine is doing the rewriting: write pages that survive the rewrite.
The H2 is the question a person would ask. The first sentence under it answers in the keyword form Copilot and Perplexity would search for. "How do I choose a running shoe for flat feet?" as the heading; "Stability running shoes for flat feet need a firm midsole and a straight last" as the opening line. One block of text now matches both the 31 percent of searches that stay a question and the 93 percent that become keywords.
Because the superlative survives between 39 and 65 percent of the time, a "best X" page still earns its title. But the other 35 to 61 percent of searches went looking for the category, the comparison or a feature, so the plain category page needs a comparison table and a features section of its own. Do not make the listicle the only place the comparison lives.
At 1.4 to 2.4 searches per prompt, every question is really a small cluster: the main ask, a constraint (for flat feet, under a budget, this year), and often a comparison. A page that only answers the main ask is competing for one search of the two or three. Sections on the edge cases and the implied follow-ups are what put you in the running for the others.
Ninety-one percent of ChatGPT's searches were one-offs. There is no phrasing to target, because the engine will not run the same one twice. What survives is topical coverage: the page that plainly covers stability, overpronation, arch support and midsole firmness will match a search built from any of those words. Synonyms and the vocabulary from People Also Ask boxes are not padding here. They are the match surface.
Perplexity preserved the person's intent 74 percent of the time on product and shopping prompts and 71 percent on food and recipe prompts. Copilot held lifestyle intent only 43 percent of the time. Commercial intent is the most stable thing in the study, which means a product page that states what it is, who it is for, and what it costs is matching the search that actually gets run. Vaguer lifestyle content gets rewritten into something you may not have written for.
For the rewritten query, the engine still needs an index to search, and being in the top ten matters more than being first. In Profound's 240-million-citation analysis, Google's first position drew 27.5 percent of human clicks but only 10 percent of ChatGPT citations; tenth position drew 2.5 percent of clicks but 4 percent of citations. The click curve is steep and the citation curve is nearly flat. A page ranking seventh for the rewritten query is in contention. A page ranking eleventh is not.
The fan-out study is 10,000 prompts over two weeks in spring 2026, on three engines, with intent classification done by a model rather than by hand. That is a solid sample for the behavior it measured and a weak basis for predicting what Google does. The position figures are for ChatGPT citations against Google rankings, not for AI Overviews directly. And all of it describes engines that ship changes monthly; the ChatGPT rewrite rate could look different by the time you act on it.
What holds up across every one of Profound's studies is the direction. The engines are moving away from the person's literal words and toward their own reconstruction of what was meant. Pages built to match a phrase are betting against that. Pages built to cover a question, its constraints and its follow-ups, in plain language a machine can lift, are betting with it. Scoring a page for how well it covers the fan-out, the edge cases and the implied questions is a reasonable proxy for that, and it is the lens Prism Optimizer uses, but the habit matters more than the tool.
If I had to reduce it to one instruction: write for the query you would run if you were the engine, not the one the person typed.
See the queries the engines actually run for your pages.Fan-out queries, People Also Ask and AI Overview data fetched live on every scan, then scored against what you have written.
Request a demo