Many companies rank high on Google and still don't show up in ChatGPT. Lucas Würms, founder of growResonance, traced the reason in his article for t3n (in German): it comes down to the path a question takes before an AI chatbot cites a source. Here we summarize the four stages, put the studies behind them into context and show you how to check your own website stage by stage.
Key takeaways
- Only about 12% of the pages that AI assistants such as ChatGPT, Gemini and Copilot cite rank in Google's top 10 for the same question. That is the finding of an Ahrefs analysis of 15,000 search queries.
- The reason lies in the path to the answer: the AI splits the question into sub-searches, queries a different index depending on the provider, has to be able to read the page technically and finally selects by page type, freshness and third-party sources.
- So what matters is not the one top keyword, but whether there is a relevant, up-to-date page for every sub-question your customers have, and whether you measure how much of it makes it into the answers.
Why a Google ranking is not enough for ChatGPT
Google's position is that website owners don't need to do anything special for AI answers: good SEO is enough, t3n reported (in German). For Google's own AI Overviews, that is largely true. Ahrefs found that 76% of the citations in AI Overviews come from Google's top 10. So AI Overviews follow the classic search results.
The picture is different for the other assistants. For an analysis from August 2025, Ahrefs asked 15,000 questions on Google, Bing and the AI assistants and compared the cited pages.
| AI system | Cited pages in Google's top 10 |
|---|---|
| Google's AI Overviews | 76% |
| ChatGPT, Gemini, Copilot | about 8% each |
| Average of all assistants analyzed | 12% |
Source: Ahrefs (2025), AI Overviews figure from an earlier Ahrefs study.
An example from the t3n article makes this concrete: an office furniture maker with 400 employees ranks second on Google for "höhenverstellbarer Schreibtisch Büro" (height-adjustable office desk). An office manager asks ChatGPT which height-adjustable desks suit an office with 50 workstations. The answer names three competitors, but not the manufacturer. That isn't necessarily down to the quality of its website. Usually, the content the AI looks for on its way to the answer is simply missing. The four stages show that path.
Stage 1: One question becomes many searches (query fan-out)
An AI chatbot rarely passes a question to a search engine word for word. It turns it into one or more targeted searches. Google calls this "query fan-out", and OpenAI describes the same principle for ChatGPT search. The desk question, for example, turns into sub-searches like these:
| Sub-search | Page that could answer it |
|---|---|
| ergonomic desks for the office | Guide to workplace ergonomics |
| load capacity for multiple monitors | Product page with technical specs |
| volume discounts for business customers | Page for business customers with terms |
This explains the low overlap with Google. According to Ahrefs, about 80% of the cited pages don't rank for the original question at all. They are found through the sub-searches. If you only optimize for the one top keyword, you don't cover these sub-searches.
How to check: Collect the questions customers ask before a purchase decision and break each one down into its sub-questions: criteria, technical details, prices, use cases. For each sub-question, check whether your website has a page that answers it directly. The gaps in this list are your content agenda. The article on query fan-out shows how AI search breaks a question into sub-searches and what that means for your content.
Stage 2: Each provider searches a different index
Each sub-search runs against a search directory, known as an index. Which one depends on the provider.
| AI system | Index | What you can do |
|---|---|---|
| Gemini, Google AI Overviews, Google AI Mode | classic SEO, Google Search Console | |
| ChatGPT | search partners such as Bing, its own search bot | Bing Webmaster Tools, allow OAI-SearchBot |
| Microsoft Copilot | Bing | Bing Webmaster Tools |
Bing in particular is underestimated. According to Bitkom (in German), 71% of AI users in Germany use ChatGPT, 50% use Gemini and 43% use Copilot. So if you are well positioned on Google and Bing, you reach most of them. Yet many companies have never paid attention to Bing.
How to check: Add your website to Bing Webmaster Tools. It takes a few minutes, and you can import a site that is already verified in Google Search Console. Then check whether your most important pages are in the Bing index.
Stage 3: Search bots must be able to read your page
For the AI to use a page, its search bots have to be able to fetch it. Usually that is not a problem. For the t3n article, Lucas Würms checked the homepages of 309 mid-sized and family-owned companies: just over four in five let in all the AI bots tested. Where bots were blocked, it was mostly training bots.
This distinction matters, because providers send out different bots. Training bots collect text for future models and play no role in search. Search bots build the index the answers come from. If you block a search bot, you disappear from that provider's answers.
| Provider | Search bot (for answers) | Training bot |
|---|---|---|
| OpenAI | OAI-SearchBot | GPTBot |
| Googlebot | Google-Extended (a control token in robots.txt, not a separate bot) |
Be careful with Google-Extended: according to Google, the token controls not only the training of the Gemini models but also whether content may be used for grounding in the Gemini apps. It has no effect on Google Search. There are also fetches triggered by a specific user question (such as ChatGPT-User at OpenAI).
JavaScript is rarely an obstacle either. According to an analysis by Vercel, the bots from OpenAI and several other AI providers don't execute it, but most websites deliver their text directly in the HTML. It gets tricky with content that only loads in the browser, for example in product configurators.
How to check: Open your-domain.com/robots.txt and look for the search bots from the table. If there is a Disallow: / under one of them or under User-agent: *, that bot is locked out. For JavaScript, the page source is enough: open it in your browser (Ctrl+U, on a Mac Cmd+Option+U) and search for a sentence from your most important content. If it is missing there, the bot can't see it either.
Stage 4: What the AI ends up citing
No provider discloses the criteria its systems use to select sources. But large analyses show three clear patterns.
Page type matters
Wix Studio's AI Search Lab analyzed more than one million citations from AI search tools such as ChatGPT and Google AI Mode, Search Engine Land reports. Lists and comparisons accounted for 21.9% of citations, guides for 16.7% and product pages for 13.7%. Which type won depended mainly on search intent: if you ask about providers, you get comparisons. If you want to solve a problem, you get guides.
Freshness matters
According to an Ahrefs analysis of about 17 million citations, the content AI assistants cite is on average 25.7% newer than the pages in Google's classic results. ChatGPT favors new content the most. Outdated prices, model names or years are therefore a real disadvantage. But changing the date alone doesn't help: revise the content of the page.
Third-party sources matter
Besides manufacturer websites, the AI often cites comparison portals, trade media and industry directories. If you appear there, you end up in the answer indirectly, even if your own website isn't cited. A simple test shows which sources count in your market: ask the typical customer questions yourself in ChatGPT, Gemini and Copilot and note which pages are named.
Measure instead of guessing: how to see what gets through
According to OpenAI, ChatGPT adds utm_source=chatgpt.com to links in its answers. In Google Analytics, you therefore see these visits under the session source "chatgpt.com". Since February 2026, Microsoft has offered the "AI Performance" report in Bing Webmaster Tools: it shows how often your website is cited in Copilot and in Bing's AI answers and which search terms the AI used to find it.
| Method | What it shows | What it doesn't show |
|---|---|---|
Web analytics with utm_source=chatgpt.com | Visits via links from ChatGPT | Mentions without a click, other assistants |
| "AI Performance" report in Bing | Citations in Copilot and Bing, search terms used by the AI | Clicks, ChatGPT and Gemini |
| Your own spot checks | Mentions and sources in individual answers | Fluctuations and trends |
| AI visibility platform | Mention rate, competitors and sources over time | Visits (use web analytics for that) |
Spot checks have a limit: AI answers fluctuate. The same question returns different brands and different sources tomorrow. The picture only becomes reliable when many questions are asked regularly and repeatedly. That is exactly what specialized AI visibility platforms do. When choosing one, check how many questions are tested and how often they are repeated.
growResonance runs your customer questions in ChatGPT, Gemini and Google AI Mode every day on all plans and shows how often your brand is mentioned, which competitors appear and which sources the answers use. This lets you see whether new pages for your sub-questions actually make it into the answers.
Checklist: check your website in seven steps
- Cover the sub-questions: Collect typical customer questions, break them down into sub-questions and check whether there is a matching page for each one.
- Choose the right page type: Offer comparisons and overviews for provider questions, guides for problem questions.
- Stay up to date: Review prices, models and data regularly and actually revise pages, don't just change the date.
- Maintain third-party sources: Find out which portals and media the AI cites in your market, and be present there.
- Don't forget Bing: Add your website to Bing Webmaster Tools.
- Run a quick technical check: Don't lock out search bots, and check whether your most important content is in the page source without JavaScript.
- Measure: Track visits from ChatGPT in your web analytics, use the "AI Performance" report in Bing and measure mentions regularly.
Conclusion: Visibility comes from answering the sub-questions
Google is right: good SEO remains the foundation. But AI assistants don't search for one keyword. They look for answers to many sub-questions, in different indexes, and they favor current, well-structured content. If you deliver this content and measure what gets through, you are not optimizing blindly.
Read the full article at t3n (in German): GEO- und Website-Check: Wie sich der KI-Chatbot Infos sucht und was Unternehmen daraus lernen können ("GEO and website check: how the AI chatbot looks for information and what companies can learn from it").
About the author
Lucas Würms is the founder of growResonance. Companies and agencies use the software to measure how visible their brand is in the answers of ChatGPT, Gemini and Google AI Mode, and to improve their presence in AI search. You can find his articles for t3n on his author page (in German).
Frequently asked questions
Is good SEO enough to be mentioned in ChatGPT?
Good SEO is the foundation, but it isn't enough. According to Ahrefs, only about 12% of the pages that AI assistants such as ChatGPT, Gemini and Copilot cite rank in Google's top 10 for the same question; for ChatGPT alone, it is about 8%. The AI finds sources through sub-searches, partly in other indexes such as Bing, and favors current content. Google's AI Overviews are different: there, 76% of citations come from the top 10.
What is query fan-out?
Query fan-out means breaking a question down into several more targeted searches. Google uses the term for its AI features, and OpenAI describes the same principle for ChatGPT search. A question about desks for an office with 50 workstations, for example, turns into searches about ergonomics, load capacity and volume discounts. According to Ahrefs, about 80% of cited pages don't rank for the original question at all; they are found through these sub-searches.
Which search engine does ChatGPT use?
According to OpenAI, ChatGPT uses search partners such as Microsoft's Bing and its own search bot called OAI-SearchBot. Gemini and Google's AI Overviews draw on the Google index, Copilot on Bing. If you are well positioned on Google and Bing, you cover most AI usage in Germany.
Should I block AI bots in robots.txt?
Distinguish between training bots and search bots. If you block a training bot such as GPTBot, you exclude your content from the training of future models, but you can still appear in answers via the search bots. If you block a search bot such as OAI-SearchBot, you disappear from that provider's answers. Be careful with Google-Extended: according to Google, it also controls use for grounding in the Gemini apps.
How do I measure whether my website appears in AI answers?
For initial numbers, three free methods are enough: visits with utm_source=chatgpt.com in Google Analytics, the "AI Performance" report in Bing Webmaster Tools and your own spot checks. Because AI answers fluctuate, a reliable picture requires many questions that are repeated regularly. The free AI visibility check offers a quick start.
Further reading
- GEO and website check at t3n (in German): the full article by Lucas Würms.
- Query fan-out: how AI search breaks one question into many searches and what that means for your content.
- AI visibility with growResonance: daily measurement in ChatGPT, Gemini and Google AI Mode.
