What do AI systems search for before they answer, and what does that mean for SEO?

What happens between question and answer

A customer asks: “Which online shop delivers lightweight trail running shoes in size 39 to Bern within two days?”

An AI system with web search does not send this sentence to a search engine unchanged. It breaks it down into its own, shorter search queries, such as “lightweight trail running shoes test” and “running shoes online shop Switzerland delivery time”. Google calls this approach query fan-out. The results of these sub-searches are retrieved, read and condensed into an answer.

How widespread this is can be seen in an analysis by AirOps (April 2026): around 16,900 queries in ten categories were each put to ChatGPT three times. For 88.6% of the queries, the system generated exactly two sub-searches; for 8.8%, none.

Rank in the sub-searches is closely linked to citations

From the same analysis:

• Pages in position 1 of a sub-search were cited 58% of the time, pages in position 10, 14% of the time.
• Across all runs, 58% of pages were never cited and 25% were cited every time they appeared in the search results. Only around one sixth lie in between.
• Pages whose heading clearly matches the question were cited 41% of the time, weakly matching ones 29%.
• Very long pages (over 5,000 words) were cited somewhat less often (29%) than pages with 500 to 2,000 words (33 to 34%).

An analysis of an earlier AirOps dataset by ALM Corp (March 2026) adds: 56% of the pages cited by ChatGPT were in Google's top 20 for the original question or at least one sub-search. A third of the cited pages appeared only in the results of the sub-searches, not in those of the original question.

For Google, there is an analysis by Ahrefs. According to it, 76% of the sources in Google AI Overviews still came from the top 10 for the question asked in July 2025; in February 2026 the figure was 38%. Ahrefs cites the Gemini 3 update and sources from sub-searches as reasons, and points out that part of the change may be down to improved data capture.

What follows from this: SEO does not become superfluous for AI answers. Pages that rank well in the searches the system runs itself are cited considerably more often. Rank is not a guarantee, though: around 44% of the cited pages were not in Google's top 20 for any of these searches.

Being found requires that there is something to read

A page can rank well on Google and still be empty for AI systems. Google renders JavaScript. According to an analysis by Vercel (December 2024), the crawlers of OpenAI, Anthropic and Perplexity did not; Gemini uses Googlebot's infrastructure. If a shop only loads its product data via JavaScript, these crawlers see nothing but a shell.

In one of our pilot measurements, the shop of a Swiss manufacturer that sells to businesses and to private customers delivered the same empty page without JavaScript on its home, category and product pages. Perplexity did not name the provider once in 146 answers. Other systems whose crawlers also do not execute JavaScript, by contrast, named it occasionally, mainly via the API. We did not measure whether the empty page plays a role for Perplexity. What is certain: what is not delivered cannot be read by a crawler without JavaScript.

More on the technical background: Client-side rendering

Three points classic SEO does not cover

1. The sub-searches are invisible to keyword tools

According to the ALM analysis, 95% of the sub-searches had no measurable monthly search volume. That is only logical: no person types these searches; a system formulates them. That is why they hardly show up in keyword research. Anyone who plans their content solely by search volume misses part of the searches that AI systems actually run.

The only way to know these sub-searches is to observe them: during real queries, log which searches a system triggers and which pages it retrieves in the process.

2. Not every system searches on Google

In practice, “SEO” usually means Google SEO. AI systems, however, rely on different search indexes. As of October 2026, as far as publicly known:

System	Search index	How confirmed
Google AI Overviews, AI Mode, Gemini	Google	official
Microsoft Copilot	Bing	official
ChatGPT	mixed: own crawler and index, plus several external search providers	weighting not disclosed; own index according to an analysis by Peec AI (September 2026)
Claude	Brave Search	Anthropic lists Brave Search as a subprocessor for web search (as of October 2026)
Perplexity	own crawler and index	according to Perplexity; whether third-party sources are also used is not disclosed

For Swiss shops, this has a simple practical consequence: a website that does well on Google but is not fully indexed on Bing may be missing from some AI answers from the outset. Whether your own website is indexed in Bing and Brave can be checked in a few minutes: for Bing via Bing Webmaster Tools, for Brave with a site: search on search.brave.com. New or changed pages can be actively submitted via IndexNow.

3. Found is not chosen

The ALM Corp analysis also shows: of all the pages ChatGPT retrieves in the course of an answer, only around 15% end up being cited. And cited does not yet mean recommended. A system can read your page, name your brand and still recommend another shop.

On the question of why someone is found but not chosen, the ranking position alone says little. Obvious candidates are what is on the page and what third parties say: whether the purchase conditions are unambiguous, whether product data is complete and machine-readable, whether brand and shop are recognisable as one company, and whether comparison portals or trade media place the provider in context. The contribution of the individual factors has not been established so far.

What this looks like in practice is shown by a case from one of our pilot measurements: When does a well-known shop drop out of AI recommendations?

Two different problems, two different responsibilities

This gives a useful distinction for any diagnosis:

Not found. The page does not appear in the sub-searches or appears too far down, it is not in a system's index at all, or its content is not delivered without JavaScript. That is a search, indexing or technical problem. It belongs to SEO work, with the addition that the relevant searches are observed rather than derived from keyword tools, and that Bing and Brave count too.

Found, but not chosen. The page is retrieved or even cited, but the recommendation still goes to another provider. That is a question of the content, data and signals the system finds when making its selection. What helps here is a comparison with the providers that actually are chosen: what is unambiguous for them, and what is not for you?

Anyone who combines both problems in one metric ends up not knowing where to start.

What you can do now

• Check without JavaScript: open a product page with JavaScript disabled. If name, price and availability are not in the HTML, many AI crawlers do not see them.
• Check indexing: in Google Search Console and Bing Webmaster Tools, check whether your most important product, category and terms pages are indexed. Set up IndexNow.
• Align headings with questions: give pages on shipping, returns, payment and product suitability headings that match how the question is asked, not how the internal navigation is labelled.
• Short and structured rather than long: our recommendation, not a finding of the study: on product pages, lists and tables for features, suitability and conditions. Very long texts did not perform better in the AirOps data, at any rate.
• Observe sub-searches: during measurements, log which sub-searches the systems trigger for your most important purchase questions, and check where you stand for exactly these searches.
• Evaluate found and chosen separately: first establish whether you are found. Then whether you are chosen. The measures are different, and often so are the people responsible.

Limits of the studies cited

The AirOps data comes only from ChatGPT, over three runs, with no language or country reported. It measures citations, not recommendations, and shows associations, not what results in what. The ALM analysis is based on a different, earlier AirOps dataset; its figures cannot be directly combined with those of the April report. That domain strength and backlinks show no positive association with citations in the AirOps data applies only to pages that have already been retrieved. Whether a page is retrieved at all may still depend on authority. The Vercel data on JavaScript rendering dates from the end of 2024; the crawlers may have changed since. AirOps itself sells content software, Peec AI sells software for measuring AI visibility. To our knowledge, there is as yet no comparable public analysis for German-language queries in Switzerland.