You are not competing for the question your customer asked
The model rewrites it into several questions you never see, then answers those. Here is what it adds.
April 23, 2026 · 3 min read
This is the most useful thing we know about how AI answers get built, and almost nobody outside the industry has been told it.
When someone types a question into Google’s AI Mode, Google does not go and search for that question. In Google’s own words:
“Under the hood, AI Mode uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf.”
Deep Search, they add, “can issue hundreds of searches.”
So there is a step between your customer and the web that nobody works on, because until recently nobody could see it.
In April 2026, Peec AI collected five million of these fan-outs across ChatGPT, Perplexity and Grok. The average number of sub-queries generated from a single user question: Perplexity 1.4, ChatGPT 2.1, Grok 6.8.
But the interesting data is not the count. It is what the models add. These are words that appeared in the machine’s sub-queries and not in the human’s question:
"best" 15 times in 100, and up to 24 when advice is being asked for
"what" 9 times in 100
"review(s)" 7 times in 100
"2026" 5 times in 100
"top" 5 times in 100
"comparison" 4 times in 100
"vs" 4 times in 100
"company/companies" 4 times in 100
The model asks for the “best X” on your customer’s behalf about one time in six, even when your customer never said the word. On questions phrased as asking for advice, closer to one in four.
Now the consequence, and it is measurable. In March 2026 Ahrefs analyzed 863,000 keyword result pages and four million AI Overview citations. Only 38 in every 100 AI Overview citations now come from Google’s top ten results. Eight months earlier it was 76. About a third come from positions 11 to 100. About a third from past position 100 altogether, pages nobody would ever scroll to.
Their attribution is that Gemini 3’s arrival in January 2026 expanded the reliance on fan-out: AI Overviews now lean “less on the direct search results and more on the sources showing up in fan out query SERPs.”
And a separate study of 1.4 million ChatGPT questions compared two things: how closely a page’s title matched the question the model had rewritten, and how closely it matched what the person actually typed. The rewritten version won. For guessing which pages get cited, how the model rephrased the question matters more than how the customer asked it.
What this changes, practically.
Ranking first for the phrase your customer types is no longer the game. The pages pulled into an answer are the ones that match the model’s rephrasing: comparing, ranking, dated, “X vs Y”, “best of”. That is why one kind of page dominates. In December 2025 Ahrefs found nearly half of all cited pages were “best X” lists, by a wide margin the largest group.
Which leads to an uncomfortable strategic conclusion. You cannot write every “best plumber in Columbus” article, and what you write about yourself is worth less than what somebody else writes about you. What you can do is be in them. This is the clearest argument we know of for why the work away from your own site, getting into guides, local media, trade directories and other people’s round-ups, is not an extra bolted onto a website project. It is where a large share of the citations actually live.
One caveat on that finding, because it sits in the same study: more than a third of the cited lists came from sites with no standing at all, “many of which were highly questionable.” Ahrefs themselves declined to adopt the tactic, choosing original research and awards instead. Being in a bad list is not the same as being in a good one, and a model reading a spam round-up is a weaker signal than a model reading a local newspaper’s guide.
And a caveat on the whole picture. This is changing quickly. Ahrefs notes that AI Overviews change their citations, their text and the questions they ask themselves roughly every two days. The numbers above will move. The mechanism will not.
