llms.txt is an illusion: proof from cats.txt and 3 figures
Summarize this article with AI
€5,400 invested in a GEO strategy: 0 citations in 6 months
€5,400.
That’s the amount one of my clients — a French B2B SaaS with 18,000 organic sessions per month — paid to a GEO-specialized agency in January 2025.
Goal: quadruple its presence in ChatGPT, Google AI Overviews, and Perplexity responses.
The plan fit in three lines: create an llms.txt file, write 4 « LLM-optimized » pages, and « quarterly monitoring ».
In July 2025, I review his metrics with him.
No additional citations in Google’s AI snippets.
Zero.
Not a single new mention on Perplexity.
Nada.
The agency’s report, though, was flattering.
It read:
- « The llms.txt file was successfully crawled by GPTBot and GoogleOther »
- « Google indexed the optimized pages in 48 hours »
- « ChatGPT mentioned your brand in a test response »
- « The file is validated by the community standard »
Four « proofs ».
Four indicators that, according to the vendor, showed the strategy was working.
Except those four arguments, a text file about a cat validated all of them.
A file named cats.txt.
Created by consultant Mark Williams-Cook and detailed in Search Engine Journal, this fictional standard declares a website’s official cats.
Cat’s name: Odd.
Breed: Tuxedo.
Position: Purr Manager.
And yet, the same four « technical proofs » worked identically.
« The four arguments used to sell llms.txt would pass for a text file about cats. Here’s why it matters for every GEO tactic you buy. » – Search Engine Journal
The GEO industry sold an illusion.
And the cats.txt experiment shows it brutally.
How a cat in a tuxedo dismantled the 4 GEO « proofs »
The cats.txt standard is a joke that tells the truth.
Let’s look at the four arguments one by one, as Williams-Cook tested them.
1. Crawling by AI bots.
When you place a text file at the root, robots from OpenAI, Anthropic, or Google come read it. That’s their job.
The cats.txt file (which contains « cat=Odd, job=Purr Manager, breed=Tuxedo, purrs_per_hour=47 ») was crawled by GPTBot in under 24 hours.
llms.txt too.
Does that change results?
About as much as the mailman walking past your mailbox proves he read your mail.
2. Indexing by Google.
Google indexed the file’s contents in its main index.
Yes, pages mentioning Odd the cat appeared in the SERPs.
But that doesn’t mean Google uses it to rank anything.
Gary Illyes from Google has repeatedly noted that mere indexation says nothing about relevance.
3. Recitation by an LLM.
ChatGPT, Claude, Gemini — all could retrieve Odd’s name, breed, and position.
They’d memorized it.
Is that surprising?
No. LLMs regurgitate accessible text. Whether they repeat a prompt or a cats.txt file, it’s the same mechanism: content availability.
Nothing suggests the file carries more weight than any other page.
4. Validation by ChatGPT.
Ask ChatGPT if cats.txt helps AI SEO, it says yes, with arguments.
But everyone knows ChatGPT can also explain why you should eat pebbles if you ask with the right words.
The test loops back on itself.
These four « proofs » — presented to you as guarantees — have no discriminating value.
If a file about a cat in a tuxedo passes them, they prove nothing.
Key takeaway
Crawling, indexing, recitation, and LLM validation are not indicators of GEO success. They are baseline properties of any accessible content.
The diagnosis is damning.
Sur 17 audits mensuels en Q3 2025, 8 sites avaient implémenté llms.txt. Le verdict est cinglant : zéro différence mesurable versus sites sans le fichier. Voici la comparaison.
llms.txt vs réalité : audit Q3 2025 sur 17 sites
Aucune corrélation entre déploiement de llms.txt et hausse de citations AI
Crawl, indexing, recitation: the correlation trap
I review 17 sites per week in audits.
In Q3 2025, 8 of them had deployed llms.txt.
None of those 8 saw citations increase in AI Overviews compared to the others.
Even on branded searches.
Even with a well-structured file.
Zero correlation.
That’s unsurprising.
llms.txt assumes LLMs read a directives file as an authority signal.
In reality, these machines work on probabilities and blend content from thousands of different sources.
A text file in their training data gets no special treatment.
As Search Engine Journal notes, we live in an age where people see patterns everywhere: it’s just noise.
Successful crawl? Just an HTTP request.
Indexation? There are trillions of them on the web.
Textual citation? It also happens with wacky poems posted on Reddit.
Jumping from « it’s present » to « it changes things » is confusion: mixing coexistence with causality.
Yet that confusion is what agencies charge thousands of euros for.
Order of magnitude: across my last 17 cases, the conversion rate « presence of llms.txt file → AI citation » is 0 for 17. The only factor that worked: semantic restructuring.
Food for thought.
Contrairement aux promesses du llms.txt, ce cas client montre qu’une vraie augmentation passe par la restructuration sémantique. Voici comment les +100% de citations se sont construits, étape par étape.
Impact réel : restructuration sémantique vs llms.txt
Décomposition des +100% de citations AI en 3 mois (cas réel 2024)
Without llms.txt, +100% AI citations in 3 months
Back to specifics.
June 2024.
An online hardware store with 350 product pages and guides contacts me.
Its organic traffic is stalled. Worse, it’s nearly invisible in Google’s AI Overviews: 2 citations all month, on ultra-long queries.
Yet its content is good.
The structure is the problem.
No llms.txt file is in place.
We decide to reorganize the information.
In 4 weeks, we group content by entities, create pillar pages, and define semantic linking around 47 key queries.
Result after 3 months:
- +100% citations in AI Overviews on core business queries (from 2 to 14 verified monthly citations)
- 27 new snippets identified via specialized tracking tool
- +43% additional organic traffic from standard SERPs (side effect)
With zero lines in an llms.txt.
None.
The only file we touched was the sitemap.
What changed everything was thematic clarity.
LLMs, like engines, prefer explicit architectures.
When you offer them a coherent semantic hierarchy, they cite you.
No need to « tell » them what to reference in a manifest file: they deduce it naturally.
Si un fichier absurde sur des chats passe tous les tests, c’est que les critères eux-mêmes sont creux. Voici comment évaluer la solidité réelle d’une tactique GEO versus ses prétendues « preuves ».
Audit des tactiques GEO fragiles : le test cats.txt
6 critères de fiabilité des techniques GEO vendues (markdown, priming, meta tags, llms.txt…)
If a cat passes the test, 90% of GEO tactics collapse
The cats.txt experiment doesn't just discredit llms.txt.
It exposes the entire current GEO machinery.
Take three common tactics sold today:
- "Optimized" markdown formatting
- "Priming" prompts to guide LLMs
- Custom meta tags for GPTBot
What do they share?
They rely on the same "proofs": crawl, indexing, recitation, LLM approval.
But if a file about cats can duplicate them, these tactics have the same logical fragility.
I reviewed 14 case studies published by GEO agencies since January 2024.
None offered controlled data with a control group.
All just showed that "optimized content was cited".
One displayed a "visibility score" percentage increase — without ever defining the score.
Like proving a diet works by photographing an empty fridge.
No stats.
No A/B testing.
Just narrative.
GEO today is a market where you're sold hope wrapped in crawl logs.
And the client pays.
A cats.txt file validates exactly the same thing.
So the problem isn't technical.
It's methodological.
The bar for proof is so low it lets any belief through.
What you must really measure (and build) to exist in LLMs
Time to change glasses.
Stop believing a text file will make you visible.
Only what you actually measure counts.
So what should you track?
1. Verifiable citations in AI responses.
Use specialized crawlers (ZipTie, Semrush AI Overviews Tracking, etc.) that query LLMs with your keywords and track your brand, page, or URL presence. Without this, you're flying blind.
2. Organic traffic growth from "AI" sources.
Check your analytics for "Google AI Overviews" or "Perplexity referral" channels. If you don't see them, you don't exist.
3. Your site's semantic consistency.
The more your content architecture is readable by machines, the more you're cited. It's not a directives file that determines this — it's how you structure entities, internal links, and pillar pages.
In 2025, AI visibility isn't decreed.
It's built.
Piece by piece.
Page by page.
Proof by proof.
I build semantic systems — silos — that organize information so LLMs understand and restate it without needing us to whisper the answer.
It's groundwork.
Nothing magical.
But it rests on facts.
And now, a question:
How many real citations has your llms.txt file brought you today?
Doubting your current GEO strategy?
I put your AI performance under the microscope in a 45-minute live audit. Does your llms.txt hold up? You walk away with facts, not beliefs.
Book a strategic call — 45 minFrequently Asked Questions
What exactly is the llms.txt standard?
<p>llms.txt is a text file placed at a site's root to tell LLMs which content to use. Created by the community, no engine officially uses it and studies showing its impact remain unconvincing.</p>
Why does the cats.txt test discredit llms.txt?
Because a file declaring imaginary cats passed exactly the same tests: crawl, indexing, LLM recitation, and ChatGPT validation. If a test accepts anything and everything, it proves nothing.
Should I delete my llms.txt file?
Deleting it isn't necessary; it doesn't hurt. But betting on it for GEO is a costly mistake. Better to invest in solid semantic architecture.
What are the real levers for AI visibility?
How you organize content by topic, your pillar pages, internal linking, and brand authority. AI citations come from context relevance, not from a declaration in a manifest file.
How do I measure the real impact of my GEO actions?
<p>Track real citations with AI tracking tools and link them to organic traffic over at least 3 months. Without measurable numbers, you're flying blind.</p>
