Part five of the series AI Content: 10 Strategies That Work. A new part every Friday.
This is the area that has changed most since 2023. Anyone planning content today from handbooks written back then is optimising for a results page that no longer looks like that.
What happened to pulled-out answers
Featured snippets, the answers pulled out above the results, were the main prize in SEO for years. According to measurements by Ahrefs they collapsed from 15.41 % of searches in January 2025 to 5.53 % in June 2025. In five months they disappeared from two thirds of the queries where they used to show.
Over the same period AI Overviews grew from 3.93 % to 27.43 %. The correlation between that collapse and that growth is 0.9, and the turning point came in March 2025. These are not two independent trends running side by side, this is one trend measured from two directions.
Google confirmed it in January 2026 through Rajan Patel, VP of engineering for Search: „On occasion we fall back to featured snippets when we are unable to generate an AI Overview.“ The featured snippet is now the fallback for when the AI summary fails to generate.
The practical consequence is that optimising for a pulled-out answer is not dead, it just moved up a floor. The target was renamed, the mechanism stayed similar.
A top 10 position stopped being the ticket
The more substantial change concerns the relationship between classic positions and citations in AI summaries. In 2025 roughly 76 % of the URLs cited in AI Overviews came from the organic top 10.
Measurements across 863,000 results pages from March 2026 put it at 38 %. Most cited sources are therefore nowhere to be found in the classic top ten. YouTube alone accounts for 18.2 % of citations that appear nowhere in the top 10.
Ranking first on Google is ceasing to be the same thing as being seen. That is an uncomfortable finding for the whole field, because positions are easy to measure and citations in an AI answer are hard to measure, so attention naturally stays where the light is.
It does not mean positions no longer matter. It means a position stopped being the only way into the game, and that a site aiming exclusively at it is playing for a smaller slice than it used to.
What Google itself recommends
It pays to be precise here, because a great deal of paid consulting has grown up around this topic. Google’s official documentation on AI features states that there are no additional requirements and no special optimisation for appearing in AI Overviews or AI Mode.
No new machine-readable files, no „AI text file“, no special structured data. Anyone selling you a technical solution that will get you into AI answers is selling you something Google describes as a category that does not exist.
More useful than hunting for tricks is understanding the mechanism Google calls query fan-out. AI Overviews and AI Mode do not run one search but several related ones across sub-topics, and they assemble their links from all of them.
Content that covers a topic including its side questions therefore has more places where it can catch than content aimed at a single phrase. Which incidentally brings honest depth back into play after years of being squeezed out by writing for a keyword.
What the data means for writing
A short direct answer right under the heading, a definition paragraph of forty to sixty words, a clear table instead of a paragraph full of numbers: all of that used to win featured snippets.
Today it raises the chance of a citation in an AI summary. So it is not featured snippets versus AI Overviews, it is one way of writing with two possible outputs. Anyone who picked up the habit early lost nothing.
The second thing the data suggests is a shift in weight from links to mentions. In a correlation analysis by Ahrefs across 75,000 brands, the strongest signals are YouTube mentions at around 0.74 and web mentions of the brand at 0.66 to 0.71. The number of pages on a site correlates at around 0.19, which is to say barely at all, and backlinks and Domain Rating both came out surprisingly weak.
The authors themselves stress that correlation is not causation, and that deserves to be taken seriously. But the direction is consistent with how language models work: they operate on what is written about you and where, not on how many links point at you.
Cannibalisation: what it actually is
The term keyword cannibalisation is not Google’s term and has no basis in its documentation. It emerged in the SEO community as a description of a real phenomenon, but it gradually turned into a bogeyman used to justify not publishing.
It has nothing to do with publishing volume. Fifty articles on fifty different topics will not cause cannibalisation. Three articles on practically the same topic will, because Google usually picks one page from a site for a given query.
Mueller played it down considerably in September 2025: „If you have 3 different pages appearing in the same search result, that doesn’t seem problematic to me just because it’s ‚more than 1.'“ He added that pages are not duplicates merely because they turn up in the same result.
The practical conclusion: do not treat cannibalisation as a bogeyman. Ask instead whether your three texts answer three different questions or answer one question three times. The first is topic coverage, the second is a mess.
Will arrivals from AI replace what search has lost?
Not so far, and it needs saying plainly, because a fair number of strategies rest on the opposite assumption. Referral traffic from ChatGPT grew by 206 % between January 2025 and January 2026, which sounds impressive.
The absolute numbers put it in context. Thirty per cent of all ChatGPT referrals go to ten domains, and 21.6 % to Google alone. Measurements across a thousand domains showed search referral traffic falling from 12 billion to 11.2 billion visits year on year, while AI referral traffic to news and media sites came to just under 36 million over the same period.
That is a difference of three orders of magnitude. Anyone building a strategy on arrivals from AI tools replacing arrivals from search is building it on numbers that are far too small.
It is more sensible to assume that the overall volume of search visits will keep falling and to set expectations and internal reporting accordingly. How to measure all this without the numbers pushing you into panicky decisions is the subject of part eight.
Previous part: Personalisation that is more than a different salutation. Next Friday: Where AI ends and a human begins.
The whole series: AI Content
1. Why most advice about AI content no longer holds · 2. A publishing plan that assumes volume does not work · 3. The prompts that make a difference, and the fact-checking · 4. Personalisation that is more than a different salutation · 5. Visibility in the year people stopped clicking · 6. Where AI ends and a human begins · 7. Images, and how to spot a marketing myth · 8. Measurement that tells you something · 9. One piece of content, several channels, without copy-paste · 10. Law and ethics, specifically, as they stand in August 2026 · 11. Build a system, not a pile of articles · 12. What to take from the whole series