Do Markdown Files and llms.txt Help You Get Cited? What the Data Says in 2026

Key takeaways

• No evidence shows llms.txt improves citations in ChatGPT, Claude, Perplexity or Google. Google says Search ignores it, and John Mueller compared it to the keywords meta tag.
• An Ahrefs study of 137,000 sites found 97% of llms.txt files received zero requests in May 2026. Only 1.1% of the requests that did arrive came from AI retrieval bots.
• Markdown files were just 0.05% of AI search citations in a Promptwatch Data snapshot, against 99.94% for HTML. Markdown mirrors of blog and marketing pages are not a citation strategy.
• Markdown and llms.txt do have a real audience: coding agents and developer documentation, usually when someone points the agent at the file.
• If you want citations, fix crawlability, page structure and content quality first. Publishing an llms.txt is cheap, so do it only after that.

Why everyone is asking this question

Every few weeks a new thread claims that a small markdown file at the root of your domain will get you into ChatGPT answers. It sounds plausible. LLMs read text, markdown is clean text, so give them a tidy index and they will cite you more. Simple.

The data does not back this up. But the story is not entirely one-sided either, which is why the debate never dies. Chrome's Lighthouse now has an llms.txt check, coding agents really do read these files, and the file costs almost nothing to publish. This guide separates what is demonstrated from what is hoped for.

Two different things get mixed up in this conversation, so let me split them:

1. llms.txt, a proposed index file at /llms.txt that lists your important pages.
2. Markdown versions of pages (.md mirrors), which some sites serve alongside their normal HTML.

The evidence differs for each, so I'll take them in turn.

What llms.txt is (and isn't)

The format was proposed by Jeremy Howard of Answer.AI in September 2024. It is a markdown file at the root of a domain. The only required element is an H1 with the name of your site or project. After that you typically add a short summary and sections of links to your most useful pages, ideally pointing to markdown versions of detailed content.

The spec says it uses markdown rather than XML because "we expect many of these files to be read by language models and agents." Think of it as a sitemap written for a model instead of a crawler.

It is not a control mechanism. Ahrefs described it well in its 2026 study: it controls nothing and blocks nothing. There is no allow, no deny, no crawl-delay. If you want to manage AI crawler access, that is still robots.txt and your server or CDN rules. Anyone selling llms.txt as a way to restrict AI bots has the mechanism wrong.

A variant, llms-full.txt, puts the full content of your key pages into one file. The theory is that it removes a retrieval step for agents working in real time.

What the data says about llms.txt

Almost nobody requests the file

Ahrefs' June 2026 study analyzed about 137,000 domains and found that 97% of llms.txt files received zero requests in May 2026. Not zero AI requests. Zero requests from anything, bots or humans.

Of the small share that did get hits, most came from bots that were not AI bots. Of the requests that did come from AI-related bots, only 1.1% came from the retrieval bots that fetch pages to answer live user queries. Ahrefs also reported that no AI bot requested an llms.txt file on a site that did not have one, so the bots are not speculatively probing for it.

One caution on the numbers. Different write-ups cite both 137,000 sites and roughly 38,000 valid llms.txt files from this study. These are most likely different things, sites studied versus files found, but check the Ahrefs study directly if you plan to quote either figure.

No correlation with citations

SE Ranking looked at around 300,000 domains to see whether having an llms.txt file correlated with being cited more by AI. It found no statistically significant relationship. When researchers removed the llms.txt variable from their citation prediction model, accuracy improved. The file was adding noise, not signal.

Adoption in that dataset was 10.13% of domains. Among the 50 most AI-cited domains, only one had an llms.txt file. The sites AI cites most are mostly not bothering with it.

The Ahrefs analysis reached the same conclusion: no correlation between having the file and how often a domain gets cited.

What Google and Mueller have said

Google's own guidance says llms.txt will neither harm nor help your visibility in Google Search, because Search ignores it. John Mueller was blunter in an April 2025 Reddit thread, as reported by Search Engine Journal:

"AFAIK none of the AI services have said they're using LLMs.TXT (and you can tell when you look at your server logs that they don't even check for it). To me, it's comparable to the keywords meta tag."

There is a fair counterargument. The keywords meta tag made claims nobody could verify, while llms.txt lists real URLs that either exist and deliver or don't. And robots.txt is not strictly necessary either. Those are reasonable points about mechanism. None of them is evidence of a citation lift, and no such evidence exists yet.

It is also worth noting what the big AI platforms' documentation does not say. None of it states that llms.txt is required, recommended, or used to decide what gets cited.

So why does Lighthouse check for it?

Chrome's Lighthouse added an llms.txt check inside its Agentic Browsing audits, reasoning that without the file, agents may spend more time crawling to understand a site's structure. People have misreported this. The audit only fails on a server error. A missing file is marked Not Applicable, because providing it is optional for now.

That tells you where the real use case is. Search ignores it. Agents that browse and act on a site may use it.

What the data says about markdown files

Promptwatch Data looked at the format of everything cited by AI search. In the Markdown in AI Search report, covering the 7 days before its February 19, 2026 publication across ChatGPT, Claude, Perplexity and Google AI Overviews, the split was:

That is roughly 2,000 HTML citations for every markdown citation. This is a single one-week snapshot from February 2026, so treat it as a strong directional signal and not a permanent constant. The data comes from real-UI monitoring of ChatGPT, Gemini, Perplexity, Claude and AI Overviews, part of Promptwatch's 26B+ data points.

The practical read: building .md mirrors of your blog and marketing pages to win AI search visibility is effort spent in the wrong place.

Where markdown does matter

The same report shows that markdown is not irrelevant. It shows crawler requests to docs .md paths from Claude (anthropic-claudebot) and OpenAI (openai-searchbot). The top path had 3,241 requests, the next 2,870, then 1,934. The paths are partly masked and the table has no stated time window, so don't over-read the exact counts.

The explanation is coding agents. Claude Code, Codex and similar tools fetch docs, READMEs and API references directly because markdown is cheaper to parse and keeps the context window clean. Agent-native search engines such as Exa, which OpenCode integrates, fit the same pattern.

So markdown serves AI agents doing a task. It does not serve AI search answering a question. Those are two different systems with different retrieval behavior.

Who actually reads llms.txt

Coding agents, and mostly when told to. Ahrefs' reading is that agents fetch llms.txt when directed, not speculatively. A developer pastes your docs URL into Cursor or Claude Code, and the agent grabs the file. Nobody discovers it by wandering in.

That makes the decision fairly clear:

If you do ship one, keep it honest and small. A short summary, 20 to 90 well-chosen links, accurate descriptions, and .md versions of docs pages where you can generate them. One site owner who hand-wrote a 14KB file with ninety curated links admitted he could measure nothing from it in AI search, and keeps it anyway. That is about the right attitude. The cost is an hour, the downside is close to zero, and the evidence of upside is thin.

One small caution: it is a public file. Anyone, including competitors, can read exactly which pages you consider most valuable.

What actually drives AI citations

Citations come from the same inputs that drive organic visibility: a crawlable, technically sound site, authority signals, and content that answers specific questions clearly. Google's own AI features guidance says they pull from the same index and ranking systems as regular Search, using retrieval-augmented generation and query fan-out. That is a crawlability and quality problem, not a file-format problem.

Promptwatch Data points the same direction on what gets cited:

• Product pages were the most cited content type in ChatGPT at 22.4% of citations from Sep 28 to Oct 4, 2026, followed by landing pages and listicles. See the content type citation report.
• Landing pages and docs stayed cited longest, while news lost half its pages within 36 days, according to the ChatGPT content half-life report.
• Almost half of ChatGPT responses cite 10 or more unique domains (Sep 28 to Oct 4, 2026), per the unique domains report. There is room to earn a slot without any special file.
• Mid-authority sites (DR 46-75) earned almost half of ChatGPT citations in August 2026, according to the domain rank report. You do not need to be a giant.

In all of that, there is no llms.txt variable and no markdown format advantage. Everything cited is ordinary HTML on ordinary pages.

One agency case reinforces it. The HOTH tracked its own AI-referred traffic from January 2025 to February 2026 and saw sessions grow 354%, with ChatGPT referrals up 502%. The inputs were digital PR link building, ongoing content on topics where AI tools recommend solutions, and a consistent brand presence on social channels. Those visitors converted at 22.79% against 2.45% for standard organic. Another site that had been invisible to AI tools picked up 54 AI citations after being made crawlable, not after adding a new file.

A better checklist than llms.txt

If you have a day to spend on AI visibility, this is the order I would work in.

1. Confirm AI crawlers can reach your pages

Check robots.txt, CDN bot rules and firewall settings. A surprising number of sites block the very crawlers they want citations from. Server logs show which AI bots actually visit and what errors they hit. Tools like Promptwatch include IP-verified AI crawler logs that filter out spoofed bots, which is handy for this check.

For a manual route, a desktop crawler such as Screaming Frog helps you audit what bots see.

2. Fix rendering

Heavy client-side JavaScript can leave retrieval systems with an empty page. Promptwatch's report advises fixing this first, because a page the retrieval system cannot parse loses citations. Prerendering services exist for sites that cannot easily change their stack.

3. Structure the HTML

Use a clear heading hierarchy, semantic markup and direct answers near the top of sections. This is the same advice as a decade ago, and it works for the same reason.

4. Measure whether it moved anything

Track which prompts mention you, which pages get cited, and whether AI-referred visits grow. A free scan is a fine start, and you can browse more options in the GEO software directory at bestgeosoftware.com. Use Google Search Console alongside, since AI features lean on the same index.

5. Then, optionally, publish llms.txt

If you run developer docs, add it, link to markdown versions of key pages, and keep it updated. Check what you publish by opening it in any text editor. Do not expect a citation bump, and do not report one to your boss.

Common mistakes to avoid

• Treating llms.txt as a ranking factor and promising clients results from it.
• Generating hundreds of .md mirrors of blog posts that duplicate your HTML.
• Stuffing the file with self-promotional claims. Models cannot trust self-reported superiority, and neither should you.
• Using it to block crawlers. It cannot.
• Skipping the real work (crawl access, rendering, content quality) because the file feels like a quick win.

The bottom line

llms.txt is low cost and fine to have, especially for documentation. It is not a citation lever. Markdown matters to coding agents and almost not at all to AI search, where HTML made up 99.94% of citations in the snapshot we have. Spend your time making your real pages easy to crawl, easy to parse and worth citing, then check the logs and citation data to see whether it worked. For ongoing numbers on how AI engines pick sources, the Promptwatch Data hub is updated regularly.