Many companies know the situation: their own website ranks solidly on Google but barely appears in the answers from ChatGPT, Gemini or Copilot. That seems contradictory, but there is a comprehensible reason for it. AI assistants work differently from a classic search engine that returns a list of results for a search term.
An analysis by Ahrefs of over 15,000 search queries shows the gap clearly: only 12 per cent of the URLs that these assistants cite also appear in the Google top 10 for the same question. For Google's own AI overviews, the company's statement that good SEO is sufficient largely applies. For the other systems, this logic only applies to a limited extent.
One question becomes several searches
The decisive difference lies right at the start of the processing chain. An AI chatbot generally does not send the user's question verbatim to a search engine, but derives several more targeted search queries from it. Google calls this approach Query Fan-out, and OpenAI describes the same principle for ChatGPT search.
An example illustrates this: if an office manager asks about height-adjustable desks for an office with 50 workstations, the system breaks this question down. It searches, for example, for ergonomic desks, for the load capacity for multiple monitors and for volume discounts for business customers. A manufacturer that ranks second on Google for the main keyword can still come away empty-handed if there is no suitable page for these sub-questions.
The figures support this explanation: according to Ahrefs, around 80 per cent of the pages cited by AI systems do not rank at all for the original question. They are found exclusively via the derived sub-searches. For content planning, this means: it is not the one contested top keyword that decides, but the coverage of the specific sub-questions that actually concern customers.
Which index sits behind which assistant
Each of these sub-searches runs against a search directory, i.e. an index. And this index differs depending on the provider. Gemini and Google's AI overviews draw on Google. According to OpenAI, ChatGPT uses search partners such as Microsoft's Bing as well as its own search bot. Copilot relies on Bing. According to research by Techcrunch, Claude draws on the Brave search engine, and Perplexity operates its own index.
This gives rise to a practical consequence: anyone who has so far looked only at Google may be overlooking Bing. Registering in the Bing Webmaster Tools takes manageable effort and creates a data basis.
Usage figures also support these two indexes. According to Bitkom, 71 per cent of AI users in Germany use ChatGPT, 50 per cent Gemini and 43 per cent Copilot. Perplexity and Claude are well below that. A clean setup on Google and Bing therefore covers the majority of usage.
Access and technology are rarely the problem
For an AI system to be able to use a page, the search bots must be allowed to retrieve it. In a test of the homepages of 309 medium-sized companies and family businesses, a good four out of five allowed all the AI bots tested. Where blocks were in place, they mostly affected only training bots, which collect texts for future models and play no role in search.
JavaScript is also rarely an obstacle. According to a Vercel analysis, the bots from OpenAI, Anthropic and Perplexity do not execute it. However, most websites deliver their text directly in the HTML. It only becomes critical where content is loaded later in the browser, for example in product configurators or dynamic price overviews. Such areas remain invisible to these bots.
Measure instead of guessing
No provider discloses the criteria by which the systems ultimately select. However, there are ways to at least partially track your own presence.
- According to OpenAI, ChatGPT automatically appends the parameter utm_source=chatgpt.com to links in its answers. In Google Analytics, this shows how many visitors come from ChatGPT.
- Since February 2026, Microsoft has offered the AI Performance report in the Bing Webmaster Tools. It shows how often a website appears in Copilot and in Bing's AI answers and via which search terms it was found.
- Specialised platforms for AI visibility regularly put typical customer questions to several AI systems and evaluate how often a company is mentioned.
With such platforms, it is worth taking a look at the methodology. Because AI answers fluctuate from run to run, the number of questions tested and the frequency of repetitions are decisive for how meaningful the results are.
Assessment for web and software projects
Good SEO remains the foundation but is losing its sole dominance. Visibility in AI answers arises where a separate, up-to-date and cleanly marked-up page exists for every relevant sub-question — this shifts the focus from keyword optimisation towards content structure and editorial planning.
Technically, this brings topics to the fore that are part of the duty of care in CMS projects anyway: server-side delivered content, traceable crawler rules and an information architecture that gives sub-topics their own entry points.
Anyone who also measures which content is actually cited can target budget precisely where it has a demonstrable effect.