Does llms.txt Actually Work for B2B SaaS? An Evidence-Based Verdict

No, llms.txt does not currently work as an AI visibility lever for B2B SaaS. Two independent 2026 studies covering more than 400,000 domains found that 97 percent of published llms.txt files received zero requests across a full month, and that the file showed no measurable effect on AI citation frequency. It is cheap and harmless to ship. It is also not a growth tactic, so treat it as the last item on your AEO list, not the first.

We keep getting the same question from B2B SaaS marketing leads: our competitor published an llms.txt file, should we? The honest answer is that llms.txt has become the clearest example of AEO cargo culting in the market right now. It is easy to ship, it looks technical, it produces a visible artifact you can screenshot for the board, and there is almost no evidence it does anything. At DevCommX we build signal-based GTM and LLMO systems for revenue teams, and the pattern we see is consistent: teams ship the file, feel productive, and skip the work in the LLMO playbook that actually earns citations. This piece is the evidence-based verdict, not another tutorial.

What llms.txt Actually Is, and What It Was Built For

Jeremy Howard of Answer.AI proposed llms.txt on 3 September 2024. It is a Markdown file at your domain root, so example.com/llms.txt, structured with an H1 for the project name, a blockquote summary, and H2 sections listing links with one line descriptions. An optional companion, llms-full.txt, inlines the full content of those pages. The stated problem it solves is context assembly: a language model landing on a marketing site has to fight through navigation, cookie banners, and JavaScript to find the substance, and the site owner knows which pages matter.

What it is not. It is not an access control file. Robots.txt already handles what crawlers may fetch, and every major AI crawler honors it. Llms.txt grants nothing, blocks nothing, and has no enforcement mechanism. It is also not a ranking file. There is no ingestion pipeline at any major model provider that treats it as a first class input.

Read the original proposal carefully and the intended audience is obvious: developer documentation being consumed by coding assistants. It was a token efficiency idea for docs, not a search visibility idea for SaaS marketing sites. The gap between that origin and how the file is now sold is where most of the confusion lives.

The Evidence: What Crawler Logs Actually Show in 2026

This is the part of the debate that used to be opinion and is now data. In May 2026 Ahrefs analyzed 137,210 domains using bot analytics and found that 97 percent of valid llms.txt files received no requests at all during the month. Of the requests that did land, roughly 77 percent came from non-AI sources. The single largest requester category was SEO audit tools at 21.7 percent, followed by unidentified bots at 14.9 percent and general web crawlers at 13.1 percent. AI retrieval bots, the ones that would actually fetch your file to answer a live user question, accounted for about 1.1 percent. GPTBot made up 4.51 percent of requests and ClaudeBot 0.80 percent.

The most damning detail is quieter: zero AI requests targeted llms.txt files on domains that had not published one. AI tools are not going looking. They are not probing for the file and finding it missing. The demand side simply is not there.

SE Ranking ran the other half of the question across roughly 300,000 domains. Adoption sat at 10.13 percent, spread evenly across traffic tiers, with high traffic sites over 100,000 monthly visits actually adopting less, at 8.27 percent. Then they tested whether the file predicted AI citation frequency using correlation tests and an XGBoost model. It did not. Removing llms.txt from the model improved accuracy, meaning the file was adding noise rather than signal.

Meanwhile supply keeps climbing. Instances of llms.txt grew 8.8x between June 2025 and May 2026, from about 4,088 to 36,120. Llms-full.txt grew more than 100x off a tiny base. Publishing is accelerating. Reading is not.

What the Platforms Have Actually Said

Google has been unusually blunt. In July 2025 Gary Illyes confirmed Google does not support llms.txt and is not planning to. John Mueller compared it to the keywords meta tag, noted that no AI services were using it and that bots were not requesting the file, and described maintaining parallel Markdown pages for bots as a poor use of engineering time. He framed the format as a temporary crutch for saving tokens in AI coding tools parsing developer docs. Google Search AI optimization guidance published in May 2026 states plainly that llms.txt is not needed for AI Overviews, AI Mode, or any other generative Search feature.

And then Google contradicted itself. On 7 May 2026 Chrome shipped Lighthouse 13.3.0, promoting an Agentic Browsing audit category out of experimental and into the default config. One of its four audits checks llms.txt handling, alongside WebMCP integration and agent accessibility. So Google Search tells you the file is pointless while Chrome DevTools flags you for not having one. That is not hypocrisy, it is two teams optimizing for different futures. Search is describing today. Chrome is building for browser-resident agents.

The model providers are quieter than the marketing suggests. As of 2026 no major provider, including OpenAI, Google, Anthropic, Meta, or Mistral, has publicly committed to reading or acting on llms.txt in production inference. Perplexity has said it may retrieve the file to help prioritize page selection. Anthropic, Stripe, Vercel, Cloudflare, Zapier, and Hugging Face all publish one, which is the fact most often cited as proof it works. Look at what those companies have in common: they are all developer tools shipping documentation, mostly through docs platforms that generate the file automatically. That is a very different use case from a Series A sales platform hoping ChatGPT will recommend it.

Why Smart B2B SaaS Teams Keep Buying the Story

The pressure behind this is real, and it is worth naming. AI referral traffic converts extraordinarily well. Ahrefs reported a case where AI referred visitors were 0.5 percent of sessions but drove 12.1 percent of signups, a 23x conversion differential on a high consideration B2B SaaS product. Broader B2B datasets put AI referral conversion in the low teens against roughly 2.8 percent for Google organic. ChatGPT referrals convert around 15.9 percent, Perplexity around 10.5 percent. When a channel converts like that, every marketing leader wants a lever they can pull this quarter.

Llms.txt looks exactly like that lever. It rhymes with sitemap.xml and robots.txt, two files that genuinely changed how search engines saw your site. The mental model transfers cleanly and incorrectly. Sitemaps worked because Google built an ingestion pipeline for them and told everyone it existed. Nobody has built that pipeline for llms.txt.

The second driver is measurement avoidance. Shipping a file is verifiable in five minutes. Proving you moved AI visibility requires prompt sets, tracked mentions, and a baseline you can defend, which is real work. If you have not yet stood up that measurement layer, start with our guide to tracking AI visibility, because without it you cannot tell whether any AEO change worked, including this one.

What Actually Moves AI Visibility Instead

The counterfactual matters more than the critique. In May 2026 Cyrus Shepard published a meta-analysis synthesizing 54 experiments, patents, and case studies into a scored ranking of 23 citation factors. The top five by evidence strength were URL accessibility at 9.5 out of 10, classic search rank at 9.4, fan-out rank at 9.3, preview control at 9.2, and query to answer match at 9.2. Llms.txt does not appear anywhere near that list.

Brand mentions beat links. Branded web mentions correlate with AI citations at roughly 0.664 against 0.218 for backlinks, about three times stronger. Earned media dominates: one analysis found 82 percent of AI citations come from earned coverage and 94 percent from non-paid sources. That is a PR and point of view problem, not a file problem.

Structure and evidence beat volume. Adding real citations to a mid-ranked page has been measured to lift AI visibility by more than 100 percent, and adding concrete statistics by roughly 22 percent. Extractable answers, clean headings, and specific numbers are what get lifted into an answer. Platform behavior also diverges sharply: across 100,000 prompts, ChatGPT and Perplexity showed only about 11 percent domain overlap in citations, because ChatGPT leans on Bing's index, Perplexity on vector retrieval, and Google AI Overviews on its own. If you want the mechanics, our breakdown of how to get cited by ChatGPT covers the retrieval path end to end.

This is the uncomfortable DevCommX position: AEO will not fix average content. If your product pages read like every other competitor's, no root level file will make a model prefer you. Models cite the source that answers the question most cleanly and is corroborated elsewhere. There is no shortcut into that.

The Verdict, With a Decision Rule You Can Use

Tier one, ship it now. You sell to developers, you maintain real documentation, and your docs platform can generate the file automatically. Coding agents, RAG pipelines, and MCP integrations genuinely fetch it. The payoff is not search visibility, it is fewer hallucinated API calls in the tools your users already live in. That is a support cost win and it is measurable.

Tier two, ship it last. You are a standard B2B SaaS company with a marketing site and a blog. Publish the file after entity clarity, schema, extractable answer formatting, and an earned mention program are already running. It costs about sixty to ninety minutes, carries no measured downside, and buys you a small option on browser agents standardizing. Do not put it on a roadmap slide as an AI visibility initiative.

Tier three, never. Do not treat llms.txt as a substitute for the fundamentals, and do not maintain a hand-written parallel Markdown copy of your site. That is the specific failure Mueller warned about, and the maintenance debt compounds every time your positioning shifts. If you have decided to ship it anyway, follow our step by step guide to writing an llms.txt file for B2B SaaS and keep it generated, not hand-curated.

Settle the Argument With Your Own Logs in Thirty Minutes

You do not have to trust anyone's study, including this one. Pull ninety days of server or CDN logs, filter for requests to the path /llms.txt, and segment by user agent. Look specifically for GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, and Google-Extended. Most B2B SaaS teams that run this query find their only requesters are SEO audit tools and unidentified crawlers, which matches the aggregate data exactly.

Run the same query against the pages your llms.txt points at. If AI crawlers are hitting your pricing page and your comparison pages directly but never touching the index file, you have just proved that the crawl path everyone worries about is not the crawl path that exists. That is a thirty minute experiment that replaces six months of debate, and it is the kind of instrumentation we build into every LLMO engagement before we change a single page.

Find Out Where You Actually Stand in AI Search

Publishing llms.txt tells you nothing about whether ChatGPT, Claude, Perplexity, or Google AI Mode currently recommend you. Our free AI Visibility Checker does: it runs your brand against real buyer prompts and shows where you are cited, where a competitor is cited instead, and which pages are doing the work. Bring the output to a call and we will map the fixes that actually move citations, from entity clarity to earned mentions to the extractable formatting models prefer. Book a GTM strategy call and we will run your AI visibility baseline with you.

Further Reading

• Answer.AI: the original /llms.txt proposal
• Ahrefs: we analyzed 137K sites, 97 percent of llms.txt files never get read
• SE Ranking: llms.txt across 300,000 domains and its effect on AI citations

FAQ

Does llms.txt actually work for B2B SaaS in 2026?

Not as a visibility lever. Two independent 2026 studies covering roughly 437,000 domains found that 97 percent of published llms.txt files received zero requests in a full month, and that the file had no measurable correlation with AI citation frequency. It is cheap and harmless to publish, but it will not move pipeline on its own.

Does Google use llms.txt for AI Overviews or AI Mode?

No. Google has said publicly that it does not support llms.txt and has no plans to, and Google Search guidance states the file is not needed for AI Overviews, AI Mode, or any other generative feature. Confusingly, Chrome shipped a Lighthouse Agentic Browsing audit that checks for the file, so Google's own products disagree with each other.

Is llms.txt the same thing as robots.txt?

No. Robots.txt is an access control file that tells crawlers what they may and may not fetch, and every major AI crawler honors it. Llms.txt is a curation file, a Markdown index of your best pages written for language models. It grants no permissions, blocks nothing, and carries no enforcement behind it.

Do ChatGPT, Claude, or Perplexity read llms.txt at answer time?

None of them has committed to it. Perplexity has said it may retrieve the file to help prioritize pages, and Anthropic publishes one for its own docs, but no major provider has confirmed that its assistant reads llms.txt when generating an answer. Crawler logs back this up: AI retrieval bots made up about 1.1 percent of the requests that llms.txt files did receive.

Should a B2B SaaS company publish llms.txt at all?

Ship it if you sell to developers and maintain real documentation, because coding agents, RAG pipelines, and docs platforms genuinely fetch it. For everyone else, publish it last, after entity clarity, extractable answer formatting, and earned brand mentions are already in place. It should cost you about an hour, and it should never displace those fundamentals.

How do I know if my llms.txt is being used?

Pull 90 days of server or CDN logs, filter requests to the path /llms.txt, then segment by user agent for GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot, and Google-Extended. Most teams find that the only requesters are SEO audit tools and unidentified crawlers. That single query settles the argument for your specific domain in about thirty minutes.

👉 Find Out if llms.txt Works