llms.txt, GEO and AI visibility: what works in 2026 and what does not
Does your website need an llms.txt to be found in ChatGPT or Google? To answer that, we have to separate the proposed purpose of the file from its documented use.
Google states explicitly that the file does not influence its search results. Ahrefs and SE Ranking add observations on fetches and citations. None of that supports a general statement about every AI service.
This article explains the available evidence and the practical requirements for AI search. If you want the overview of the whole field: what this work looks like as a service is described in what a GEO agency is.
What is an llms.txt file?
An llms.txt is a plain text file in Markdown format that sits in the root directory of a website. It is meant to give AI systems a compact overview: what the company does, for whom, and where the most important content sits. The proposal by Jeremy Howard describes a signpost for the use of content. Unlike robots.txt, the file does not control crawler access.
The structure is deliberately simple:
# Beispiel GmbH
> Webentwicklung und KI-Integration für mittelständische Unternehmen im Rheinland.
## Kernseiten
- [Leistungen](https://beispiel.de/leistungen): Überblick aller Angebote
- [Preise](https://beispiel.de/preise): Preismodelle und Pakete
- [Kontakt](https://beispiel.de/kontakt): Ansprechpartner und AnfrageA file like this is written in ten minutes. That alone explains its popularity, and why so many agencies offer it.
Is llms.txt worth it? What the data shows
No, at least not as a lever for Google visibility. Google says explicitly that llms.txt is ignored for Google Search, including AI Overviews and AI Mode. Ahrefs and SE Ranking add observational data, but no experiments that prove an effect on mentions. Google: AI optimisation guide
What the studies show
| Source | Scope | Finding |
|---|---|---|
| Ahrefs | 137,000 websites | 97% of the llms.txt files were not fetched a single time in May 2026 |
| SE Ranking | 300,000 domains | no statistical correlation between an existing llms.txt and the frequency of AI citations |
What Google says
Google's official optimisation guide puts it directly: AI Overviews and AI Mode require no special markup, no AI files and no Markdown formats. Google Search ignores llms.txt; for other services a file can only make sense if that service explicitly uses it. Google: AI optimisation guide
The evidence in detail
| Field | Detail |
|---|---|
| Claim | For Google Search, having an llms.txt file does not increase visibility or rankings. |
| Method 1 | Ahrefs examined 137,000 domains with Ahrefs Web Analytics; around 38,000 of them had a valid llms.txt. What was measured were fetches in May 2026, not the effect on visibility. |
| Result 1 | 97% of the existing files received no fetch in the month examined. |
| Method 2 | SE Ranking compared, across 300,000 domains, the presence of an llms.txt with the frequency of AI citations; that is a correlation, not a causal test. |
| Result 2 | The authors report no statistical correlation. |
| Period covered | Ahrefs: May 2026; the details are in each original study. |
| Limits | The studies are observations of a single point in time. A fetch does not prove use, and a missing correlation does not prove that no service will evaluate the file in future. |
| Verifiable at | Ahrefs · SE Ranking |
The one fair exception
The file can make sense for services or agents that explicitly use it as a signpost. That is a different purpose from Google visibility and has to be checked service by service.
That is an important difference. Navigation is not visibility. The file may help an agent that is already there; it does not make one come.
Our recommendation
Create an llms.txt only if a service that matters to you actually uses it, or if you offer documented agent navigation. For Google Search you should not budget any visibility effort for it.
How do I get found in ChatGPT?
A mention in ChatGPT cannot be guaranteed. What helps is that your website and your company data are clear, correct and accessible to people; which sources a single system draws on is not predictable.
A correct entry in an industry directory, a specialist article or a helpful community answer can give people additional information. Do not chase artificial mentions: Google explicitly calls them no sustainable optimisation. Google: AI optimisation guide
In practice that means:
- A complete, maintained Google Business Profile. Google names Business Profiles as a source for local company information in search results and AI answers. → The complete guide to the Google Business Profile
- Keep your company data consistent. Check that name, address and phone number match factually. Different spellings are not automatically different companies.
- One short description everywhere. Use the same description of your offer in every place. Inconsistent self-descriptions dilute the brand profile.
- Directories and review platforms. Maintain your information where your customers use it to make a decision.
What that looks like step by step, sorted by field of work and with a self-test in twelve questions, is in the practical guide: getting found in ChatGPT and AI search.
And the uncomfortable truth up front: nobody can guarantee that ChatGPT names your company. The answers are not deterministic and change constantly. The only honest approach is to build the foundation.
How does AI find my company?
For Google, the search index and publicly accessible, crawlable content are the documented basis. Google can process JavaScript when it is not blocked; that says nothing about every other AI product. Google: AI optimisation guide
Serve the main content reliably and check indexing and the rendered output with Google tools. A test without JavaScript is an indicator of robustness, not a complete test for every system. If you want an outside view of whether AI crawlers can read your site, our free website check looks at exactly that.
Check robots.txt by the purpose of each identifier. Allowing training, search or user-initiated fetches are separate decisions; none of them guarantees a mention.
| Identifier | Role and decision |
|---|---|
| Googlebot | Google Search including its search features; keep it crawlable for Google visibility. |
| Google-Extended | Control for Gemini training and grounding; not a crawler of its own and not a Google Search ranking signal. |
| OAI-SearchBot | OpenAI search; assess it separately from training. |
| GPTBot | OpenAI training; allow it only if you want to allow training. |
| ChatGPT-User | User-initiated fetch; separate from automated crawler access. |
| Anthropic identifiers | Roles differ by agent; check the current documentation before you change anything. |
Googlebot controls Google Search. Google-Extended, by contrast, controls whether content crawled by Google is used for Gemini training and grounding. At OpenAI, OAI-SearchBot separates search from GPTBot for training. Google crawlers · OpenAI bots
What are GEO, AEO and LLMO, and which of them counts?
Three abbreviations, one subject. All of them describe the attempt to appear in AI-generated answers. The differences are academic, the measures largely identical. Do not let providers who turn this into three separate products unsettle you.
| Term | Written out | Focus |
|---|---|---|
| GEO | Generative Engine Optimization | Visibility in generative answers (ChatGPT, Perplexity, Gemini) |
| AEO | Answer Engine Optimization | Direct answer boxes and featured snippets |
| LLMO | Large Language Model Optimization | Mention in the model knowledge itself |
| AIO | AI Overviews / AI Optimization | Google's AI summaries in the search result |
The most important difference in practice is not between these terms, but between cited and recommended.
Being cited means: an answer points to your page as a source. Being recommended means: the system names your offer as a possible choice. Both are separate from the decision a customer makes later.
We have no robust, general weighting of the individual signals. This checklist assigns concrete tasks instead:
| Field of work | What to check |
|---|---|
| Brand mentions | Is the information correct and relevant for your audience? |
| Accountability | Is it clear who is responsible for the content and the offer? |
| Content quality | Does the page contain original, helpful and evidenced information? |
| Structured data | Do the entries match the visible content? |
| Customer reviews | Are the experiences genuine and verifiable? |
| Technical SEO | Can the relevant search system fetch the main content? |
| Links | Do they lead to helpful, fitting information? |
The points complement each other. For Google, helpful content, technical accessibility and correct information remain the documented basis.
A "best providers" list you write yourself does not prove an independent recommendation. A helpful comparison discloses its selection, its criteria and possible conflicts of interest. Even then it cannot promise a mention in AI answers.
How many brand mentions do you need?
There is no documented, robust threshold for the number of mentions. Check instead whether independent information is correct, relevant and helpful for your customers.
| Check | Goal | Question |
|---|---|---|
| Company data | Correctness | Are the details right on the important pages? |
| Independent sources | Relevance | Do they help people decide? |
| Your own content | Usefulness | Does it answer real questions? |
| Measurement | Comparability | Is the development documented over time? |
Individual mentions are not a reliable proof of success. Instead of a monthly number, track whether the information stays correct and helpful in the places that matter.
How do AI Overviews work?
AI Overviews are generative features in Google Search. Pages have to be indexed by Google and eligible for a snippet to be considered for generative search features; an appearance still is not guaranteed. Google: AI optimisation guide
A general rate for how often AI Overviews appear, or for their effect on clicks, depends on the market, the query and the measurement method. Check the change in your own search data.
Whether and when AI Overviews appear can change per query and per point in time. A small sample is not a general proof.
Google requires no special chunking for AI search. Choose your sections and your page length by what the information needs. The spam policy on scaled content, by contrast, concerns mass-produced content meant to manipulate rankings or AI answers. Google guide
7 measures for verifiable, helpful content
The GEO study compared nine interventions in a controlled experiment with model answers. It reports the metrics Position-Adjusted Word Count and Subjective Impression; the results are neither a general ranking order for websites nor a promise of business results. To the study
- Cite your sourcesBack up statements with linked, verifiable sources. Above all, that improves verifiability for your readers.
- Put statistics in contextName the source, the period and the measurement method. Example: across 137,000 domains, Ahrefs found 97% of llms.txt files with no fetch in the month it examined.
- Quote experts correctlyWith name, role and organisation. Anonymous claims do not count.
- Mark your own experienceSeparate documented facts from experience and judgement.
- Structure clearlyOne thought per paragraph. The direct answer belongs at the start of the section, not at the end.
- Use technical terms correctlyDomain language signals competence, as long as it is used correctly.
- Answer directlyAnswer the central question early in the section and explain it afterwards in plain words.
Do not write text for repetition
Google explains that its systems understand synonyms and context. Write for the question your audience has; no general percentage AI penalty can be derived from the GEO study.
For Google, llms.txt and special AI markup are deliberately missing from this list: Google Search ignores these files for visibility and rankings. Other services you have to assess against their own documentation.
Which content gets cited at all?
There is no evidence for a generally valid ranking order of content formats for AI citations. Choose the format that answers your audience's question most reliably:
| Content type | Suitable when |
|---|---|
| Specialist article | experience or a complex question is being explained |
| Guide | a process has to be reproducible |
| Product or service page | the offer, its limits and the next steps are clear |
| Overview | people have to navigate between topics |
The right form follows the question, not a supposedly preferred format.
Do I have to update my content constantly?
Update content when the facts, your offer or the guidance require it. A general age rule for AI citations is not robustly documented:
| Trigger | Update |
|---|---|
| Facts or regulations change | check and correct immediately |
| The offer or the price changes | check and correct immediately |
| New experience or insight | add it with context and date |
| No change in substance | do not move the date just for a signal |
A visible date is supposed to explain actual changes. Do not update just to move a date.
The practical consequence: update when the substance changes. Not to move a date.
How can you measure AI visibility?
For Google, the generative AI performance report in Search Console has been available since 3 June 2026; the rollout was completed on 31 August 2026. It measures Google features, not mentions in other AI products. Google: announcement and rollout
For everything outside Google that changes nothing. Anyone who wants to know whether ChatGPT or Perplexity names their brand still has to go about it differently.
Manual, free, doable right away:
- Note your 20 most important queries, including "best [your category] in [your city]" and "[your brand] reviews".
- Ask every one of them in ChatGPT, Perplexity, Gemini and Google.
- Record: are you named? Who else is? Which page is linked as the source?
- Repeat this monthly and compare.
Repeated, carefully logged questions help you spot changes. A small sample is not a complete picture of your visibility, though.
Tool-supported: specialised services exist for tracking AI citations and share of voice, among them Otterly AI, Peec AI and ZipTie. For most mid-sized companies that only makes sense once the basics are in place.
How this fits into a set of metrics that your management will read is in what remains of SEO and what works now.
Frequently asked questions
What is an llms.txt file?
A Markdown text file in the root directory of a website, proposed as a signpost to important content. It replaces neither robots.txt nor the sitemap, and it does not control crawler access.
Where do I find an llms.txt example file?
You can write one yourself in about ten minutes: create the file llms.txt in the root directory, start with an H1 heading for the company name, add a short description as a quote and list your core pages as Markdown links. The listing above in this article is a complete llms.txt example file.
Is there an llms txt seo benefit?
Not for Google Search: Google ignores llms.txt for visibility and rankings. Across 137,000 domains, Ahrefs observed 97% of files with no fetch in the month examined; across 300,000 domains, SE Ranking reports no correlation with AI citations. Those are observational data, not proof of effect.
Is llms.txt the same as robots.txt?
No. robots.txt controls which crawlers may fetch which areas, and it is actually obeyed. llms.txt is a proposal with no binding effect.
Should I allow GPTBot in robots txt?
Only if you want to allow training: GPTBot is OpenAI's training crawler. Search runs through OAI-SearchBot, and ChatGPT-User is a user-initiated fetch. At Google, Googlebot covers Google Search and Google-Extended covers Gemini training and grounding. Check the current provider documentation before you change anything.
Does llms.txt work for ai visibility?
Not for Google Search, which ignores the file. For any other service it only works if that service documents that it uses the file. What does work: helpful content of your own, evidence for your central statements and a clean technical basis. Google requires no special AI markup for this.
How do AI Overviews work?
AI Overviews are generative features in Google Search. Pages have to be indexed by Google and eligible for a snippet, and even then an appearance cannot be guaranteed.
What is the difference between SEO and GEO?
SEO covers visibility in search systems. GEO is a common term for work on generative answers; for Google the normal SEO basics remain decisive.
Does AI-generated text harm my visibility?
Not in principle. Google judges the content and its purpose. Mass-produced pages that mainly aim to manipulate rankings or AI answers and offer little use can violate the spam policies.
Do I have to show my prices publicly?
Public prices or price ranges help people judge an offer when they fit it. No automatic selection by an AI system follows from that.
How to read the evidence: Ahrefs measures fetches, SE Ranking measures correlations. The GEO study is a controlled experiment with model answers. The Google documentation describes Google products. None of it supports guaranteed mentions, rankings, leads or revenue.
Sources: Ahrefs llms.txt study · SE Ranking llms.txt study · Google: optimising for generative AI search · Google crawlers · OpenAI bots · Princeton GEO Study · llms.txt proposal · Google: GenAI performance report
