GEO myths: what primary sources say and what remains unanswered

Several common GEO recommendations have no backing in primary sources. The Common Crawl documentation consulted does not mention Internet Archive as a discovery route; Google states that no special files or markup are needed for its AI features; and vendor studies do not show that llms.txt, the top 10 or text length explain citations on their own.

This guide is written for those who must decide where to invest time, not only for those who implement. Technical terms are explained in plain words and the detail goes under the label "For the technical team". CookingMetrics offers audits and support to development teams, and does not sell visibility tools. Two of the cited sources come from companies that do sell SEO software; this is declared in each case. This guide is not a complete list of myths: it includes only those we could check against a primary source.

In summary

• A claim is called a "myth" here only if a primary source contradicts it or if there is no evidence in its favor. If the evidence is mixed, it is presented as uncertainty.
• Myth 1: saving pages in Internet Archive is not listed in the Common Crawl documentation as a way for it to crawl them. There is no official backing; that does not prove it is impossible.
• Myth 2: the fact that Common Crawl knows an address does not imply that it downloads it. Its own documentation says it archives a sample of the web.
• Myth 3: Google states that there is no need to create new files, AI text files or special markup to appear in AI Overviews or AI Mode. This applies to Google's features, not to other assistants.
• Myth 4: a vendor study of about 300,000 domains found no correlation between having llms.txt and being cited. There is no evidence in favor, and no definitive proof against either.
• Mixed evidence: that being in the top 10 "ensures" citations and that long text is cited more are neither proven myths nor truths: the available studies come from vendors, with methods that change between versions.
• Distinguishing between model training and real-time page retrieval avoids most of the confusion.

How each claim was classified

Each claim carries an evidence class, so it is clear how much weight what is said carries:

• Official documentation: the service's own page on how it works.
• Study with a published method: an analysis that explains its sample and its method. If it is published by a company that sells related products, this is stated.
• No source consulted: the claim circulates, but we did not find a primary source that supports it.

No class proves causality on its own. A correlation between two things, for example having a file and being cited, does not show that one causes the other.

Exclusive content

For signed-in readers

A claim review sheet

Distinguish supporting, contradictory, and insufficient evidence with an example of a universal citation promise.

Sign in to access this section →

We work to offer the best possible content, which is why some of these explanations are not public.

Before starting: training is not retrieval

Much of the confusion comes from mixing two different processes:

• Training: models learn from large collections of pages, which are filtered and processed beforehand. Common Crawl is one of the public collections of that kind.
• Retrieval: when an assistant with search answers, it consults pages at that moment and may cite them. It is a different mechanism, with different robots and different rules; see how to measure traffic from AI assistants.

Being in a training collection does not on its own explain a citation in an assistant with search, and the sources consulted do not say otherwise.

Myth 1: "saving pages in Internet Archive makes Common Crawl pick them up"

• What the documentation says: the Common Crawl FAQ, consulted on November 12, 2026, states that its crawler uses the sitemaps announced in the robots.txt file and that it can find pages by following links from other sites. It does not mention Internet Archive.
• What it does not say: it does not say that this list is exhaustive or that the practice is impossible. That is why this guide does not claim that the practice does not work: it claims that there is no official documentation that backs it.
• Evidence class: official documentation, by absence. No source consulted for the claim.
• What is advisable instead: declare the sitemap in robots.txt and keep links from other sites pointing to the important pages. They are the two routes the documentation mentions.

Myth 2: "if Common Crawl knows an address, it downloads it"

• What the documentation says: the same FAQ states that it does not usually archive a whole site, but a randomly chosen subset, and that the dataset is a sample of the web.
• What the study says: a research paper, Host-Aware Web Crawling for LLM Pretraining Data (Malte Ostendorff, October 9, 2026), whose author is listed as affiliated with Common Crawl Foundation, analyzes how crawling is distributed across sites. With data from a 2024 crawl:
  • the correlation between how much of a site is crawled and its "quality density" is −0.08 across 5,654,186 sites; that is, there is almost no relationship;
  • the authors estimate, for that crawl, about 58 million high-quality pages among the known addresses that were left undownloaded. It is an estimate that assumes the undownloaded pages resemble the downloaded ones;
  • an alternative policy would achieve, in simulation, up to 7.8 times more high-quality pages per downloaded page with small budgets, and 1.7 times as the budget grows.
• What "quality" means here: the prediction of a classifier on its usefulness for training a model. The paper itself treats it as an approximation, not as quality in an editorial sense.
• What it does not say: that Common Crawl has changed its crawler, or that it will. The proposal was tested in simulation. It does not explain citations in assistants with search.
• Evidence class: official documentation and study with a published method. The limitations section of the paper was not read; nothing is claimed about it.

For the technical team. The correlation cited is Spearman's, a measure of how similar the order of two lists is, between −1 and 1; a value close to 0 indicates little relationship. The study uses a sample of WARC files (the format in which Common Crawl stores pages) and an index of sites for the full analysis.

Myth 3: "you have to create a file or special markup to appear in Google's AI features"

• What the documentation says: the Search Central page "AI features and your website" states that there is no need to create new machine-readable files, AI text files or markup, and that there is no special structured data format to add. To appear, the page must be indexed and be eligible to be shown with a text snippet. Appearing is not guaranteed. The page shows no update date.
• Scope: it applies to AI Overviews and AI Mode. It says nothing about other assistants, which have their own documentation.
• Evidence class: official documentation.
• What is advisable instead: the usual SEO practices that this same page recommends; see what GEO is.

Myth 4: "llms.txt makes assistants cite the site more"

• What llms.txt is: a proposal by Jeremy Howard, published in September 2024 and updated in August 2026 at llmstxt.org, to offer agents information that helps them use a site. Its author says he expected it to be useful mainly when the agent consults the site, rather than for training. The page claims no effect on rankings, citations or traffic.
• What the companies say: we did not find official Google documentation saying that it reads it, and the Google page from myth 3 states that it is not needed. OpenAI's page on its crawlers does not mention llms.txt when describing them. We did not consult Anthropic documentation on the subject.
• What the only study consulted says: SE Ranking, a company that sells SEO software, published on October 15, 2026 a study of about 300,000 domains. Only 10.13% had the file. The study reports that there is no correlation between having it and being cited, and that removing that variable from its model improved its accuracy. The page does not state which assistants it consulted, how many answers, on which dates or the values of the correlations, and its authors warn that the results depend on the model and on the dataset.
• What can be concluded: there is no evidence in favor of llms.txt increasing citations, and the only study consulted found no relationship. Nor is it definitive proof against: the study comes from a vendor, with an incomplete method on the page, and assistants change. The file may have other uses for those who develop agents, as its proposal indicates. Cloudflare, in its agent readiness evaluation, treats it as optional; see what isitagentready.com measures.
• Evidence class: proposal by the format's author, absence from official documentation and a vendor study.

Mixed evidence 5: "being in the top 10 ensures you get cited"

This claim is neither a proven myth nor a truth. What exists are two Ahrefs studies, a company that sells SEO software, on Google AI Overviews:

• July 2025: 76.10% of the cited pages were among the first ten results, out of 1.9 million citations from 1 million AI Overviews, taking the first three citations of each answer.
• March 2026: 37.9% of the cited addresses were among the first ten, out of 863,000 queries and 4 million addresses. The article itself compares "~76% versus ~38%", but states that it improved its analysis method; that is why it is not a direct comparison. The data come from the vendor's own visibility tool.

Two corrections to the reading:

• It is a proportion of the cited pages, not of the top 10 pages. That a share of the cited pages is in the top 10 does not say how many top 10 pages are cited.
• Both figures are from AI Overviews. They do not apply to ChatGPT or AI Mode. In the guide on query fan-out another vendor study is discussed, according to which about two thirds of the cited pages were in the top 10 neither for the main query nor for any sub-query.

The prudent statement is that ranking well helps in being considered, without a guarantee, and that the percentages vary with the method.

Mixed evidence 6: "long content is cited more"

An Ahrefs study from December 3, 2025 analyzed 560,346 AI Overviews and 1,677,876 cited addresses, and obtained a Spearman correlation of 0.04 between word count and citation: practically none. The article uses the same figure for being cited and for the position of the citation. The average cited page had 1,282 words; 53.4% had fewer than 1,000.

A limitation pointed out by this guide, not by the study: the sample includes only cited pages. It does not compare with pages that were not cited, so it does not allow concluding whether length influences being cited or not. It is a vendor study, on AI Overviews, and there are no length data for ChatGPT in the materials read. The evidence is "no relationship observed in those data", not a rule.

Exclusive content

For signed-in readers

A lab for three recommendations

Evaluate three fictional statements and turn each into a decision with explicit limits.

Sign in to access this section →

We work to offer the best possible content, which is why some of these explanations are not public.

What to do instead

CookingMetrics recommendation, not a guarantee:

• Secure the basics: that pages are indexable, with content in the initial HTML and accessible to the robots you want to allow. See what to look for in a web design and development provider.
• Measure your own data: visits, crawling and citations of your own site, before relying on third-party figures.
• Always ask about the source: who says it, with what sample, what they sell, and whether it is official documentation or a vendor study.
• Treat mixed evidence as uncertainty: test one change at a time and record the date, without attributing the result to it.

Exclusive content

For signed-in readers

A reviewable decision log

Document when to adopt, test, defer, or reject a practice and what evidence could change the decision.

Sign in to access this section →

We work to offer the best possible content, which is why some of these explanations are not public.

What not to conclude

• That a practice "does not work" because we found no official backing: it is an absence of documentation, not proof of the contrary.
• That what is said about Google's AI features applies to other assistants.
• That a vendor's correlation demonstrates causality.
• That this list exhausts the GEO myths.

What to do next

For common claims about conventional search, see what Google says about SEO myths.

Before implementing a GEO recommendation, write in one line which source backs it and of what class it is. If you would like us to review which recommendations on your site have backing and how to measure their effect, learn about our SEO for AI search service: we review and support your team, and we do not sell visibility tools.

Frequently asked questions

What is Common Crawl and why does it matter in GEO?

It is a public collection of web pages that is used, among other things, to train models. Its documentation describes a crawl of a sample of the web. Being in that collection does not on its own explain a citation in an assistant with search.

Do I need an llms.txt file to appear in AI Overviews?

No. Google states that there is no need to create new files or AI text files to appear in AI Overviews or AI Mode. The file is a proposal designed for agents that consult a site.

Does being in the top 10 guarantee that Google cites me in AI Overviews?

No source guarantees it. The Ahrefs studies measure what proportion of the cited pages was in the top 10, with results of 76.10% and 37.9% depending on the version and the method, and not the probability of being cited when in the top 10.

Is longer text cited more?

An Ahrefs study on AI Overviews found no relationship between word count and citation (correlation of 0.04). It analyzed only cited pages, so it does not allow concluding about being cited or not.

How do I tell a myth from a recommendation with backing?

Ask who states it, with what sample, whether they sell something related and whether there is official documentation. If there is no primary source, it is a hypothesis; if the evidence comes from a vendor, it is an indication, not proof.