Does llms.txt Do Anything? Four GEO Claims, Checked Against the Sources
Most GEO advice is repeated, not verified. Here are four claims you have probably read this month, checked against the primary sources, including two we had to correct on our own site.
- 01No major AI search system reads third-party llms.txt files. Publish one for optionality, not for citations.
- 02Google's own position is that optimising for AI search is still SEO, and that AEO and GEO are rebranded labels for the same work.
- 03Across 1M+ citations, blog-format content (listicles plus articles) is about 38%, nearly 3x product pages at 13.7%.
There is a lot of confident advice about AI search right now, and very little of it cites anything. Most of it is one blog post repeating another blog post. So we went and checked four claims you have probably read this month against the primary sources, and two of them turned out to be wrong on our own site too.
Claim 1: "Add an llms.txt and AI will cite you"
This one is everywhere, and the evidence against it is unusually clear.
John Mueller of Google has said publicly that no AI system currently uses llms.txt, and compared the file to meta keywords, the tag search engines abandoned decades ago. Gary Illyes, also of Google, said at Search Central Live in July 2025 that Google has no plans to support it.
The field data agrees. SE Ranking looked at 300,000 domains and found that among the 50 most AI-cited domains, exactly one had an llms.txt file. OtterlyAI audited server logs and found that 0.1% of AI-bot traffic touched the file, 84 requests out of 62,100.
Anthropic, Stripe, Cloudflare and NVIDIA all publish one. None of them has stated that their own crawlers consume anyone else's.
Where it does help: AI coding agents. Cursor, Claude Code and similar tools read llms.txt when loading library documentation, and Mintlify auto-generates the file for thousands of developer-docs sites. If you sell developer tooling, publish it, because it helps agents quote your docs correctly.
What the field looks like: when we scanned 42 Singapore business websites already ranking on page one for buyer queries, 74% had no llms.txt at all. They rank anyway.
What we do: rankr publishes an llms.txt at rankrhq.com/llms.txt, and the file itself says it is there for optionality, not because it is a citation lever. Our free checker gives it the lowest weight of any check and labels it emerging and unproven. That is not hedging, it is what the evidence supports.
Claim 2: "GEO is a new discipline, separate from SEO"
This is the claim the entire GEO consulting market rests on, so it is worth being honest that Google disagrees.
Google's own Search Central documentation states that optimising for generative AI search is still SEO, and that AEO and GEO are rebranded labels for the same work. Their stated position is that there is no separate AI index. A page has to be indexed and eligible to appear with a snippet in Google Search before it can appear in any AI feature.
We sell AI-search work, so read this as against our own interest: the technical work is largely the same work. Clean titles, real answers to real questions, structured data, a site that loads and can be crawled. Anyone selling you GEO as a brand new discipline with brand new deliverables is selling you SEO with a markup.
What genuinely is different is narrower than the marketing suggests. Measurement is different, because rankings do not tell you whether an AI answer named you, and you have to check by asking. Formatting matters more, because a model extracts passages rather than ranking pages. And the surfaces differ, which is why being indexed in Bing is worth ten minutes of your time even though it does nothing for Google.
Claim 3: "Product and landing pages are over 30% of what LLMs cite"
We saw this one repeated recently and went looking for the source.
The best public dataset we found is Wix Studio's AI Search Lab research, covering 75,000 AI answers and more than a million citations. The actual breakdown:
| Content type | Share of citations |
|---|---|
| Listicles | 21.9% |
| Articles | 16.7% |
| Product pages | 13.7% |
Product pages are real and underused, so the underlying advice is not silly. But 13.7% is not "over 30%", and the implication that you should stop writing blog content is backwards: listicles and articles together are about 38%, nearly three times the product-page share.
The honest takeaway is boring. Build both. Comparison and listicle content is the single most-cited format, and a service or product page is the thing that converts the visit once the citation sends someone your way.
Claim 4: "Rank number one on Google and AI will cite you"
This one used to be closer to true and is drifting away fast.
Ahrefs found that only about 12% of AI-cited pages rank in Google's top 10. More recent data shows AI Overview citations drawn from Google's own top 10 results falling from 76% to 38% in a short window. The two systems are separating.
Meanwhile, what does correlate is not what most people optimise for. Across 75,000 brands, Ahrefs found branded web mentions correlate 0.66 to 0.71 with AI visibility, while backlinks correlate under 0.2. Being talked about beats being linked to.
There is a related finding worth knowing if you optimise for ChatGPT specifically. Bing is the index OpenAI documents, and Seer Interactive found about 87% of ChatGPT citations matched Bing's top results in 2024. Since then, independent testing has shown ChatGPT also pulling from Google's index without OpenAI disclosing it, and a teardown of ChatGPT's backend API traced results to Bright Data, a retrieval layer running hundreds of millions of residential proxies. We had that wrong on this blog until recently and have corrected it.
What we would actually do
Stripped of the acronyms, the list is short and unglamorous.
- Get indexed in Bing as well as Google. Ten minutes, and almost no Singapore business has done it.
- Answer real customer questions as headings, and answer them with specific numbers and named sources. Extractable beats clever.
- Keep meta descriptions under about 150 characters, because that snippet is often what a model reads rather than your full page.
- Publish comparison and listicle content, because it is the most-cited format, and back it with pages that convert.
- Earn mentions, not just links. A named mention in a place people already read is worth more than a directory backlink.
- Publish an llms.txt if you like. It costs nothing. Do not expect it to do anything.
Frequently asked questions
Does llms.txt actually work?
There is no public evidence that any major AI search system reads third-party llms.txt files. Google's John Mueller has compared it to meta keywords, SE Ranking found only one of the 50 most AI-cited domains had one, and server-log audits put AI-bot traffic to the file at around 0.1%. It is genuinely useful for AI coding agents reading developer documentation. For a normal business site, publish it for optionality and do not count it as a citation lever.
Is GEO just SEO with a new name?
Largely, and Google says so directly: optimising for generative AI search is still SEO, and AEO and GEO are rebranded labels for the same work. The real differences are in measurement, formatting for passage extraction, and which surfaces you check. The underlying technical work is the same work.
What content gets cited most by AI?
Across 75,000 AI answers and over a million citations, listicles were 21.9%, articles 16.7% and product pages 13.7%. Blog-format content is the largest share, and product pages are the most underused of the three.
Does ranking on Google get you cited by ChatGPT?
Less than people assume. Only about 12% of AI-cited pages rank in Google's top 10, and AI Overview citations sourced from Google's top 10 have fallen from 76% to 38%. Rank well anyway, but treat AI citation as a separate thing you have to measure separately.