What llms.txt actually does in 2026 (and what it doesn’t)

What is llms.txt, in one paragraph?

llms.txt is a proposed convention: a Markdown file placed at the root of your domain (/llms.txt) that gives large language models a curated, machine-readable map of your most important content. The idea is reasonable. When an AI tool fetches one of your pages, it does not read it the way a person does; it receives raw HTML and has to extract meaning from a document full of navigation, cookie banners, scripts and footer links. A clean Markdown index, the proposal goes, lets a model skip the noise and find what matters. The disagreement is not about whether that is a nice idea — it is about whether the systems you care about actually use the file. That is an empirical question, and in 2026 we finally have enough data to answer it.

Does llms.txt help you get cited in AI search?

On the evidence available right now, no — not in any measurable way. Three independent lines of data point the same direction, and it is worth seeing them together because each on its own gets hand-waved away.

First, the crawlers do not read it. One analysis of more than 515 million AI bot visits over a 90-day window — filtering specifically for the user agents that drive citations, including GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot and Google-Extended — found that only 408 of those requests touched /llms.txt. That is not a small effect; it is statistically indistinguishable from zero. The file that is supposed to guide the search and answer bots is, in practice, almost never fetched by them.

Second, Google has said no, on the record. In mid-2025, Google’s Gary Illyes confirmed that Google does not support llms.txt and is not planning to, and John Mueller compared it to the long-discredited keywords meta tag. Google’s official AI optimization guidance is explicit that the file is not needed for Search, and that AI Overviews and AI Mode draw from the same index and the same quality signals as organic results. Two Google product teams sent mixed signals in passing during 2026, but the documented stance has not changed: it is not a ranking or citation factor.

Third, adoption is low and flat. A study of 300,000 domains by SE Ranking put adoption at roughly 10% — about one site in ten, after eighteen months of industry conversation. A standard that genuinely moved citations would not sit at one in ten and stall; the market tends to chase things that work. So if someone is selling you llms.txt as a way to rank or get cited in ChatGPT, Perplexity, Claude, Gemini or Google AI Mode, they are selling you something the data does not support. That is the blunt half of the answer.

So where does llms.txt actually do real work?

In a different layer of the stack entirely: the agentic web, where AI agents act on a user’s behalf — fetching context, choosing tools and completing tasks. This is the part most coverage misses, because it looks at the file through a search lens and concludes it is useless. Viewed as an agent-to-agent signal, the picture inverts, and the evidence is concrete rather than aspirational.

Anthropic recommends llms.txt in its guidance on writing for agents. OpenAI maintains llms.txt files for its Agents SDK and for the Agentic Commerce Protocol. Perplexity has been observed retrieving and using llms.txt to prioritize which pages it reads, independent of standard retrieval. Mistral reads it where present, with the spec support still maturing. Each of those is a major AI company actively using the file in an agent or agent-to-agent workflow — regardless of whether Google Search touches it. That is why the cleanest framing in 2026 is the one almost nobody uses: llms.txt is a business-to-agent signal, not a search lever.

There is even a cultural tailwind worth noting. At Google I/O 2026, Logan Kilpatrick remarked that “the hottest new programming language is markdown” — a half-joke that points at something real, which is that agent workflows increasingly run on clean, structured Markdown context. A curated llms.txt is a small instance of that broader shift toward making sites legible to machines that act, not just machines that rank. The correlation data fits this reading: brands that publish a well-curated file see a modest but measurable uplift in citation rates on Anthropic and Perplexity specifically, strongest for sites with sprawling navigation that benefits from explicit curation. Modest and specific — not the universal win the hype promised, and not the nothing the skeptics claim.

What is worth watching through the rest of 2026 is whether the agentic use deepens enough to matter commercially. OpenAI’s Agentic Commerce Protocol and the wider move toward agents that buy, book and transact on a user’s behalf are exactly the workflows where a machine-readable map of your site stops being decorative and starts being functional. If that future arrives at scale, llms.txt graduates from a nice-to-have to part of being transactable by agents — still not a search lever, but infrastructure for a separate channel. The sensible posture: keep the file current, ready if that channel matters to your buyers, without pretending it does something it does not.

Why are both common positions wrong?

The “llms.txt is a ranking factor” camp is wrong because it confuses a plausible mechanism with a demonstrated one. It is easy to argue that a clean index should help a model understand your site; it is another thing to show that the search bots fetch it and that citations move as a result. Neither holds up: the bots do not fetch it, Google disclaims it, and adoption has stalled. Selling it as a citation shortcut is, at best, optimism dressed as strategy.

The “llms.txt is useless” camp is wrong because it judges the file only by the search lens and never checks the agentic one. “The crawlers ignore it, therefore it does nothing” skips the question of which crawlers, doing what. The answer and search bots ignore it; the agent frameworks do not. Dismissing the file outright means missing a genuine, if narrow, use that three of the largest AI companies have already built on. Both camps are confident, both are tidy, and both are wrong for the same underlying reason: they treat “AI” as one thing when it is at least two — the systems that answer search queries, and the systems that act as agents — and llms.txt lands very differently in each.

Should you create an llms.txt file?

For most small B2B SaaS sites, developer-led products and documentation-heavy sites, the honest answer is yes — with the right expectation attached. It is low cost to produce, it cannot hurt you, and it positions you for the agent workflows that several major AI companies already run on the file. If your buyers or their tools are starting to use agents that fetch and act on web context, agent-readiness is worth the afternoon it takes.

But create it as agent-readiness, not as a citation play, and keep it in its place. It should be near the bottom of your AI-visibility list, not the top. If you have a choice between writing an llms.txt file and fixing the three things that actually move citations — clean, server-rendered content; a coherent, verifiable entity; and earned mentions on the third-party platforms the engines read — do those first, every time. The file is a finishing touch on a house with foundations, not a substitute for the foundations. The most expensive mistake in 2026 is treating it as the shortcut that lets you skip the hard work.

How do you write a good llms.txt file?

As a curated map, not a dump. The two failure patterns are predictable. The first is treating the file as a second sitemap — listing every URL on the site with no descriptions — which gives an agent volume instead of guidance and defeats the point of curation. The second is writing vague link labels like “click here” or “read more” that tell a machine nothing about what it will find. Both produce a file that is technically present and practically worthless.

A good file does the opposite. It points to your most important pages — your entity definition, your core product, your real documentation — and gives each a short, specific description of what an agent will find there. Think of it as a concise table of contents written for a reader that must choose what to load within a limited context window. Lead with the page that defines who you are and what you do, because that is the page an agent most needs to anchor on. Keep it short, keep it current, and treat every line as guidance you are giving a machine that is deciding, in that moment, what to read about you.

What should you do instead, if your goal is citations?

Spend the effort where the evidence says citations are decided. Three things carry most of the weight, and none of them is a root-level Markdown file. The first is making your content machine-readable: clean, server-rendered HTML, because AI crawlers receive raw markup and structural noise competes directly with your content inside a fixed context window. If your important content only appears after JavaScript runs, or only reveals on scroll, a crawler may never see it — a failure we treat as seriously as a broken page, because to an engine it is one.

The second is a coherent entity: a consistent, verifiable description of who you are across your own site and the wider web, so an engine can attach what it reads to a clear “you” rather than confusing you with a similarly named competitor. The third is earned presence on the third-party sources the engines actually cite — community discussions, review platforms, comparison pages — because the majority of AI citations point off your own domain. Research on the underlying mechanics keeps confirming the same direction: adding verifiable statistics and clearly attributed quotations to content measurably improves how often it gets cited, while old-school keyword stuffing performs worse than doing nothing. The work that compounds is the unglamorous work — and llms.txt, useful as it is in the agent layer, is not a way around it.

It is worth being concrete about that last claim, because it is one of the few things in this field with a controlled study behind it. The foundational GEO research that gave the discipline its name tested specific content changes against a benchmark of generative-engine answers. Two interventions stood out: adding relevant statistics, and adding quotations from credible sources. Together they improved visibility by figures in the range of a fifth to over a third, depending on how it was measured — while the SEO-era reflex of stuffing keywords actually reduced it. That is the empirical backbone under the advice to write answer-first, evidence-dense pages: it is not an aesthetic preference, it is what the engines measurably reward. A curated llms.txt does none of that work; it points at pages, it does not make the pages worth citing.

And if you do build the file, measure honestly rather than assuming it worked. The cleanest signal most teams ignore sits in their own server logs: the request frequency of GPTBot, PerplexityBot and ClaudeBot on a page tends to rise a few weeks before any citation lift, which makes crawl frequency a usable leading indicator — and, incidentally, a way to see directly whether those bots are touching your llms.txt at all. Most will find they are not. That is not a reason for despair; it is a reason to put the file in its proper place and spend the freed-up effort where the logs and the studies agree it pays. The file is not the enemy; treating it as the strategy is. Build it if it fits your stack; just do not expect it to do a job in search that it was never doing, and do not let the afternoon it takes stand in for the months of real work that being cited actually requires. The brands that compound AI visibility through the rest of 2026 are not the ones that shipped a Markdown file and waited; they are the ones producing non-commodity content, earning third-party citations on the platforms the engines read, and tracking citation share across those engines to see what is working. Put llms.txt on that list.

If you want to know where your brand actually stands across the five engines before you spend on any of this, that is exactly what our AI visibility audit measures, and the method behind it is published in full. llms.txt has a place on the checklist. It is just much further down than the noise suggests.

llms.txt in 2026: quick answers

A note on sources and certainty

Everything above is dated for a reason: this is a fast-moving area, and a confident claim from a year ago is often wrong now. The figures here come from public 2026 analyses — the 515-million-visit bot study, SE Ranking’s 300,000-domain adoption data, Google’s own optimization guidance and on-record comments, and the published agent documentation from Anthropic and OpenAI. Where the data is correlational rather than causal, we have said so; the citation-rate uplift from a curated file is a correlation, not a guarantee. If the major engines change their stance — if Google reverses, or the search bots start fetching the file at scale — we will update this page and date the change. That freshness discipline is not a flourish; it is the same thing the engines reward, applied to our own writing.

The AC Group has spent 27 years earning attention online by being the substantive source rather than the loudest one. This note is that principle applied to a topic where the loud takes are mostly wrong in both directions. If it helped you decide what to ship, it did its job — whether or not you ever talk to us.