ChatGPT Does Not Search Through Google: 97% of Its Sources Are Missing From the Top 50 Results

If you have spent years investing in SEO, I have bad news for you: ChatGPT pays little attention to your Google ranking. For three months, researchers asked ChatGPT, Gemini, and Perplexity the same questions and analyzed what actually shapes their answers. The result is a useful map of how GEO works—that is, optimization for generative search results.

A Search Engine That Is Not Google

The study's headline figure is that 97% of the pages found by ChatGPT are absent from Google's top 50 results for the same queries. ChatGPT searches independently, and its search works radically differently from conventional search results. That is why the strategy of “first rank highly on Google, then appear in an AI model's answer” does not work automatically.

There is another important feature. The model inserts familiar brands into the search query itself. If your product was not present in the training data, the model will most likely never even consider searching for it. This gives large companies an advantage and creates a barrier for smaller ones.

How Smaller Companies Can Break Through

External roundups help overcome that barrier. Appearing in lists on sites such as Reddit has a very strong impact on the outcome, including for major brands, so this is not a myth. For smaller companies that were absent from the training data, links like these increase the chances of being mentioned by tens of times.

The mechanism is typical and fairly simple. First, the model finds an external list, such as a roundup of the best tools in a particular niche. It then runs a separate search about the companies included in that list. If you are not on the list, that second step will never reach you.

The practical conclusion is that managing mentions on third-party platforms is now part of product distribution. You need to identify which roundups exist in your field and earn inclusion honestly: through real users, reviews, and discussions rather than artificial manipulation.

Anticipate the User's Question

On your own website, matching the wording of the question works best. A person asks, “Which refrigerator makes the least noise?” That means you need a page with a similar headline or a section in the middle of the text. “Quiet Refrigerators” will work. “Important Advantages of Our Refrigerators” will not: it was written for a marketer, not for the query.

Interestingly, AI does not appear to consider much of the standard toolkit on its own. This includes FAQs, schema.org markup, a summary at the beginning of an article, an author box, and loading speed. There is an important caveat, however: if an FAQ contains real questions phrased naturally, it helps because it matches the query—not because it is a formal page element.

What to Do About It

• Collect real customer questions from support, sales, and chats, and turn them into page headlines.
• Find external roundups in your field and check whether you are included.
• Do not expect a single markup change or faster website to get you into an answer: these are not what causes inclusion.
• Set up regular checks of what ChatGPT, Gemini, and Perplexity say about you and your competitors.

Limitations

The results change very quickly. The authors cite a figure of 86% over 10 weeks, but the author of this summary considers that metric somewhat odd and would not draw conclusions from it without reading the study itself. This is one set of queries measured over one period, while the behavior of the models and their search systems is constantly being updated. Many of the findings are best treated as working hypotheses to test against your own queries.

The conclusion is simple: GEO is not an extension of SEO but a separate discipline with its own sources. The winners will be those who write for real user questions, appear where models look for roundups, and regularly check what the models say about them. The primary source is available as arXiv preprint 2610.15033.