Research · October 8, 2026
Does llms.txt work for AI citations? A study of 3,048 websites
400 Dutch-language questions without a brand name, 3 runs per chatbot, measured through the ChatGPT and Gemini APIs. We checked 3,048 websites for an llms.txt: the websites cited as a source, and the websites from the Google top 10 that were not cited.
Matt Timmermans, Timmermans Media. Measured: September 26, 2026. Preregistration: doi.org/10.5281/zenodo.22974863. Data: doi.org/10.5281/zenodo.23010967.
Does llms.txt work for AI citations? In our study of 3,048 websites, the websites ChatGPT and Gemini cite did not have one more often. We found an llms.txt on 22.4% of the checked websites that ChatGPT or Gemini cited as a source. Among the websites from the Google top 10 that were not cited, the share was 25.4%. We also found no such file on 16 of the 20 most-cited websites, including Consumentenbond (the Dutch consumer association), ANWB (the Dutch motoring association), and Rijksoverheid (the Dutch central government).
In other words, websites get cited all the time without us finding an llms.txt. Whether the file changes the chance of being cited is something we did not measure. The results apply to the APIs on September 26, 2026.
The findings
llms.txt study: three figures on cited websites
22.4% versus 25.4%:
the share of checked cited websites where we found an llms.txt, versus the uncited websites from the Google top 10.
16 of the 20:
the number of most-cited websites where we found no llms.txt, not even in four other common locations.
151 of the 610:
the number of files we found that carried a recognizable mention of a WordPress SEO plugin, most often Yoast SEO (supplementary).
Definition
What is an llms.txt file?
An llms.txt is a text file in the root of a website, available at /llms.txt. It contains a short summary of the website and links to its most important pages, in a simple format (Markdown). The idea is that it helps language models see more quickly what a website is about.
The format was proposed by Jeremy Howard of Answer.AI in September 2024. It is not an official standard. In its guide to generative AI features in Google Search, Google states that Google Search ignores llms.txt files: an llms.txt neither helps nor hurts visibility in Google Search. You are free to create the file for other services that use it. That statement is about Google Search, not about the Gemini API we studied.
An llms.txt is often recommended as a way to get cited by AI. We looked at whether cited websites have the file more often.
What we count as “llms.txt present”: the website returns a text file at /llms.txt, not a web page, with at least one heading or link. Many websites return a regular web page for any unknown address; those do not count.
Comparison
Do cited websites have an llms.txt more often than uncited ones?
Not in this measurement. Among the checked cited websites, we found an llms.txt on 22.4%; among the uncited websites from the Google top 10, on 25.4%. The intervals overlap. This was preregistered as a descriptive question, without a statistical test. Put the other way around: on 78% of the checked cited websites (1,033 of 1,331), we found no file with recognizable content at /llms.txt.
| Cited by ChatGPT or Gemini | In the Google top 10, not cited | |
|---|---|---|
| Websites | 1,585 | 1,463 |
| Could not be checked (blocked, connection dropped, timeout) | 254 | 237 |
| Checked | 1,331 | 1,226 |
| llms.txt present | 22.4% (20.3 to 24.6%) | 25.4% (22.9 to 28.0%) |
| Not found (error code) | 69.7% | 66.4% |
| Web page instead of llms.txt | 7.3% | 7.8% |
| Present, but without a heading or link | 0.1% | 0.0% |
| Empty file | 0.5% | 0.3% |
Percentages of the checked websites. Each website counts once, no matter how often it was cited. In parentheses, the 95% interval; it does not account for clustering of websites within the same question and is therefore descriptive only. Differences were calculated from unrounded figures.
By chatbot: of the checked websites ChatGPT cited, 20.6% had an llms.txt (122 of 592); for Gemini, the share was 22.9% (202 of 883). A website cited by both counts in both groups (supplementary, not preregistered).
We do not conclude that cited websites are less likely to have an llms.txt. The difference is small and untested, and the two groups consist of different types of websites. Had the uncited group had the same mix of website types as the cited group, its share would have been 23.8% instead of 25.4% (supplementary, not preregistered). That shrinks the gap with 22.4% from 3.1 to 1.4 percentage points. Part of the difference, then, comes from how the groups are composed. Two cited websites were left out of the reweighting, youtube.com (video) and tweakers.net (mixed: forum, news, and price comparison), because those types did not occur among the checked uncited websites.
Figure 1. Share of the checked websites where we found an llms.txt: cited by ChatGPT or Gemini 22.4% (298 of 1,331, 95% interval 20.3 to 24.6%) and in the Google top 10 but not cited 25.4% (312 of 1,226, 22.9 to 28.0%). Checkable websites only; each website counts once. The interval is descriptive. Measured September 26, 2026.
Top 20
Which of the most-cited websites have an llms.txt?
We found one on four of the twenty: KVK (the Dutch Chamber of Commerce),e-Boekhouden.nl and Moneybird (both accounting software), and Univé (an insurer). On the other 16, we found no llms.txt, not even in four other common locations.
| Rank | Website | Answers citing it as a source | llms.txt |
|---|---|---|---|
| 1 | consumentenbond.nl | 180 | No |
| 2 | anwb.nl | 136 | No |
| 3 | rijksoverheid.nl | 121 | No |
| 4 | voedingscentrum.nl | 76 | No |
| 5 | google.com | 68 | No |
| 6 | belastingdienst.nl | 65 | No |
| 7 | independer.nl | 47 | No |
| 8 | kvk.nl | 47 | Yes |
| 9 | overheid.nl | 42 | No |
| 10 | thuisarts.nl | 35 | No |
| 11 | overstappen.nl | 34 | No |
| 12 | trustoo.nl | 30 | No |
| 13 | youtube.com | 29 | No |
| 14 | tweakers.net | 26 | No |
| 15 | werkspot.nl | 25 | No |
| 16 | ikea.com | 22 | No |
| 17 | e-boekhouden.nl | 21 | Yes |
| 18 | moneybird.nl | 21 | Yes |
| 19 | unive.nl | 21 | Yes |
| 20 | geld.nl | 18 | No |
Number of baseline answers from ChatGPT and Gemini combined in which the website appears as a source. “No” means we found no recognizable llms.txt in any of the checked locations: /llms.txt, the www address, the /.well-known/ folder, the /nl/ subfolder, and a docs subdomain. Not every location was reachable; most sites had no docs subdomain at all.
Consumentenbond, ANWB, and Rijksoverheid are cited as a source 121 to 180 times, yet we found no llms.txt on any of them. That shows the file is not a prerequisite. It does not show that an llms.txt can have no effect: for that, you would need to compare the same websites with and without the file.
SEO plugins
How many llms.txt files come from Yoast and other SEO plugins?
In 151 of the 610 files we found (24.8%), we recognized a fixed mention of a WordPress SEO plugin. Files without such a recognizable mention fall outside this count.
| Plugin | Files | Share of the 610 |
|---|---|---|
| Yoast SEO | 96 | 15.7% |
| Rank Math | 42 | 6.9% |
| All in One SEO | 13 | 2.1% |
| At least one plugin | 151 | 24.8% |
Supplementary, added after the fact and not preregistered. We counted the fixed phrase “Generated by ...” or its Dutch equivalent “Gegenereerd door ...” in the first 2,000 characters of the file; a passing mention of a plugin name does not count. No file mentions more than one plugin. The count was not checked by hand.
A plugin-generated file was slightly more common among uncited websites: 85 of the 312 websites with an llms.txt (27.2%), versus 66 of the 298 cited ones (22.1%).
The mention tells you which program most likely created the file, not whether the owner deliberately switched it on.
For SEO specialists
Is llms.txt worth it for AI visibility?
Not as a must-have, based on this study. Do not make an llms.txt a requirement in advice on AI visibility, and do not present it as a proven way to get cited more often. Whether the file helps is something we did not measure. This study therefore provides no evidence for or against an llms.txt.
| Area of work | How to apply this study |
|---|---|
| Client advice | Do not present an llms.txt as a requirement or a proven way to earn more citations. On the most-cited websites, we mostly did not find the file. |
| Technical audit | Check whether an SEO plugin is already generating an llms.txt, and whether its content is correct. Nearly a quarter of the files we found mentioned a plugin. |
| Competitor analysis | Do not use the presence of an llms.txt to explain a competitor's citations. |
| Reporting | Measure citations per chatbot directly, rather than using an llms.txt as a proxy. |
This applies to Dutch-language questions through the APIs, measured on a single day.
To see which websites ChatGPT and Gemini do use as sources, read our study of Reddit as an AI source. For how often a page from the Google top 3 comes back as a source, read our study of Google rankings and AI sources. Whether websites with little Google traffic get cited too is covered in our study of the Google traffic of AI sources.
Deep dive
Deep dive: llms.txt by type of website
Brands and online stores most often have an llms.txt (27 to 28%), government websites and news media the least (4 to 9%). Within each type, the shares for cited and uncited websites are usually close together.
| Type of website | Cited: with llms.txt | Not cited: with llms.txt |
|---|---|---|
| Brand or online store | 27.4% (843) | 28.4% (799) |
| Comparison or affiliate site | 23.4% (192) | 26.5% (151) |
| Specialist blog | 14.1% (64) | 25.2% (163) |
| News media | 8.9% (56) | 4.0% (25) |
| Government and institutions | 4.3% (164) | 4.5% (67) |
Exploratory. Website type assigned by a language model, without human review. Percentages of the checked websites of that type; in parentheses, the number of checked websites. Together, the five types cover 1,319 of the 1,331 cited and 1,205 of the 1,226 uncited websites. The remaining 12 cited websites are forums (2), review platforms (4), other (4), video (1), and tweakers.net (1); the remaining 21 uncited ones are forums (13), social media (3), review platforms (2), reference works such as Wikipedia (2), and other (1). They were included in the reweighting, except the video website and tweakers.net; social media and reference works do not occur among the cited websites and therefore carry no weight.
The largest differences are in specialist blogs and news media, but those involve small numbers. At ChatGPT, government websites and institutions account for the most citations: 48% of all citations, including Consumentenbond and ANWB. As distinct websites they form a smaller group: 164 of the 1,331 checked cited websites. We found an llms.txt on 4.3% of them.
Methodology
Method: checking more than 3,000 websites for an llms.txt
What was preregistered
This part was included in the preregistration as a descriptive question, without a testable hypothesis (DOI 10.5281/zenodo.22974863). Fixed in advance were the check itself, the six outcome classes, the comparison group, and an extra check of the 20 most-cited websites without an llms.txt.
Deviations, per-website check, false positives, and limitations
The plugin count is a supplementary analysis added after the fact. The counting rules were fixed before the script ran, but after we had looked at the files during an earlier check. Rules, script, and results are in the data deposit.
The preregistration called for a manual check of 20 websites. We replaced that with a fixed automated check of four additional locations. We selected the 20 most-cited websites that had no file at /llms.txt; websites that returned a web page, an empty file, or no usable response also counted as “no file.” Sixteen of them are in the top 20 above. This substitution is logged as a deviation. For one of the 20, facebook.com, the additional locations also mostly returned no response, so for that site we did not establish that there is no llms.txt.
After the deposit, two rules were amended on September 26, 2026, both before any analyses were run. Amendment 1 (a second page of Google results for 88 search terms) helps determine which websites fall into the Google top 10 comparison group. Amendment 2 concerns Ahrefs data and does not affect this publication. Both are in the amendment log in the data deposit.
How we checked each website
| Element | Details |
|---|---|
| Which websites | All websites ChatGPT or Gemini cited as a source in the baseline measurement (1,585), and all websites from the Google top 10 for the 400 search terms that ChatGPT and Gemini did not cite (1,463). 64 of those 1,463 were cited in later repeat measurements; the grouping follows the baseline measurement |
| Request | One request per website to https://[website]/llms.txt, identifying itself recognizably as Timmermans Media, at most 5 redirects, 10-second timeout |
| Timing | September 26, 2026 (19:41 to 20:10 UTC), additional checks on September 27, 2026 |
| Outcome | Six classes: present with recognizable content, present without recognizable content, empty, web page (HTML), absent, could not be checked |
Ruling out false positives
We wanted to rule out websites that return the same text at every address and would therefore wrongly count as “present.” So for every website with an llms.txt in the full measurement (1,630 websites, including the 610 in this comparison), we also requested an address that cannot exist. Not a single website returned the same content there. This check is supplementary and not preregistered.
Websites without a usable response
About 16% of the websites (491 of 3,048) gave no usable response, mostly because of security that blocks automated requests (error code 403) or drops the connection. Those websites are excluded from the percentages. A recheck with a different program on 30 websites with a dropped connection gave the same error for 24 and a response for 5. For the thirtieth, no connection was made within 10 seconds (timeout). We did not separately check the 403 blocks, the largest group.
Limitations of the llms.txt measurement
- Descriptive, not causal. We measure how often an llms.txt occurs, not whether the file influences citations.
- Different groups. Cited and uncited websites differ in type, size, and topic. A difference in llms.txt may be related to that.
- Website level, ChatGPT and Gemini combined. Each website counts once, no matter how often it was cited.
- Standard location only. For all websites, we checked /llms.txt. Only for the 20 most-cited websites without a file there did we check four additional locations.
- A single point in time. Websites may have added or removed an llms.txt after September 26.
- Websites that could not be checked. 16% could not be checked. If those are more or less likely to have an llms.txt, the percentage changes.
- Website type assigned automatically. By language models, without human review.
Open data
Data and citation for the llms.txt study
The preregistration (10.5281/zenodo.22974863) and the data deposit (10.5281/zenodo.23010967) are on Zenodo. The per-website check results and the analysis scripts are in the data deposit.
The supplementary analyses on this page are in the data deposit as well: the plugin count, the breakdown by chatbot, and the reweighting by website type.
The data are licensed under CC BY 4.0, the analysis scripts under MIT.
Preregistration and data on Zenodo
Cite this study
Timmermans, M. (2026). Does llms.txt work for AI citations? A study of 3,048 websites. Timmermans Media. Preregistration: doi.org/10.5281/zenodo.22974863. Data: doi.org/10.5281/zenodo.23010967. License CC BY 4.0. timmermansmedia.com/research/llms-txt-ai-sources/
This study is available under the CC BY 4.0 license.
Frequently asked questions
Questions about llms.txt and AI citations
Does llms.txt work for getting cited by AI?
This study cannot show that it works or that it does not: we did not measure the effect of adding the file. What we did find is that websites ChatGPT and Gemini cited did not have one more often. We found an llms.txt on 22.4% of the checked cited websites and on 25.4% of the uncited websites from the Google top 10.
Do I need an llms.txt to be cited by AI?
No. On most of the websites ChatGPT and Gemini cited in this measurement, we found none, including 16 of the 20 most-cited. Whether an llms.txt increases the chance of being cited as a source is something we did not measure.
Should I remove my llms.txt?
This study gives no reason to remove your llms.txt. We did not examine whether adding or removing one affects citations.
Is llms.txt actually used by AI systems?
We cannot say from this data. We measured whether cited websites have one, not whether an AI system reads the file.
Why do so many websites have an llms.txt?
Partly because of WordPress SEO plugins that can generate the file. In 151 of the 610 files we found, there was a recognizable mention of Yoast SEO, Rank Math, or All in One SEO.
Can you measure whether my website shows up as a source in AI?
Yes. Using the same measurement design, we can measure whether your website appears as a source in ChatGPT, Gemini, and Perplexity. See our GEO service.
Related research
Research · same measurement
How consistent are AI brand mentions? A study of 60 buying questions
The same measurement: 60 buying questions, each asked 8 times, to see how much the brands mentioned vary from one answer to the next.
Research · same measurement
Does ChatGPT cite small websites? A study of 400 questions
The same 400 questions, with the estimated Google traffic of every cited website. Many of these websites get cited, but each in only a handful of answers.
Research · same measurement
Does ChatGPT use Google results? What 1,200 answers show
The same 400 questions, set against the Google results from the same day.
Research · same measurement
Does ChatGPT use Reddit as a source? What 3,600 AI answers show
The same measurement, focused on Reddit and on sources found versus sources cited.
This page is the English edition of a study first published in Dutch. Read the original Dutch version.
Glossary
Dutch terms on this page
The study used Dutch questions and Dutch websites. These are the Dutch names and words that come up.
- ANWB
- Dutch motoring and travel association with roadside assistance and consumer advice, comparable to AAA.
- Belastingdienst
- Dutch tax authority, comparable to the IRS.
- Consumentenbond
- Dutch consumer association that tests products and services, comparable to Consumer Reports.
- e-Boekhouden.nl
- Dutch online bookkeeping software for small businesses; boekhouden means “bookkeeping.”
- Gegenereerd door
- Dutch for “generated by.”
- Geld.nl
- Dutch personal finance comparison site; geld means “money.”
- Independer
- Dutch comparison site for insurance, mortgages, and energy, comparable to NerdWallet.
- KVK
- Dutch Chamber of Commerce (Kamer van Koophandel), where every business registers.
- Moneybird
- Dutch online invoicing and accounting software for small businesses, comparable to QuickBooks.
- Overheid.nl
- Dutch government portal for laws, regulations, and official publications.
- Overstappen.nl
- Dutch comparison site for switching energy, internet, and insurance; overstappen means “to switch.”
- Rijksoverheid
- Dutch central government's website for rules and public services, comparable to USA.gov.
- Thuisarts
- Dutch health information site written by family doctors; the name means “home doctor.”
- Trustoo
- Dutch review platform for local service providers, comparable to Yelp.
- Tweakers
- Dutch technology news and review site with a price comparison tool and a large forum.
- Univé
- Dutch cooperative insurer for health, home, car, and other consumer insurance.
- Voedingscentrum
- Netherlands Nutrition Centre, a government-funded source of official healthy eating advice.
- Werkspot
- Dutch marketplace that connects homeowners with tradespeople, comparable to Angi or Thumbtack.
Does your website show up as a source in AI?
Measure your visibility in AI answers
We measure for your industry and region whether your website appears as a source in ChatGPT, Gemini, and Perplexity, and which websites the chatbots cite instead.