llms.txt: Necessary, Not Sufficient
llms.txt is worth having and not worth bragging about. In our measurements it has near-zero direct effect on whether AI platforms cite a site. We still ship it on every engagement — it is cheap, it is clean, and it has no downside — but we refuse to sell it as a ranking lever, and you should refuse to buy it as one. What actually decides citations is extractable content and third-party signals, and neither fits in a text file.
What llms.txt is — and is not
An llms.txt file is a plain-text map of a website written for AI readers: a short summary of what the site is, and links to the pages worth reading, kept free of navigation chrome and marketing layout. This site has one. The proposal behind the format is honest and modest: make it cheaper for a language model to find the right pages. Nothing about that sentence is controversial, and nothing about it promises visibility.
The controversy is commercial, not technical. Vendors sell “llms.txt optimization” as a visibility product — a line item with a price, implying that shipping the file moves a number. That is where our measurement comes in.
What we measured
This finding is ours. Across the sites we track, adding llms.txt alone did not move citation rates. AI platforms cite pages for content reasons — the passage states the fact, the fact is quotable, the surrounding page is trustworthy — not because a manifest file exists pointing at them. No buyer ever asked a question and received an answer whose citations were chosen from an llms.txt file; that is not how source selection works. The file is a convenience for readers, and the reader it conveniences is not the one picking sources.
We publish this plainly because the temptation runs the other way. A cheap intervention that shows up in a report looks like progress, and it is easier to sell “we deployed AI-ready files” than “we rewrote thirty pages and earned four third-party mentions.” One of those moves the citation rate. The other moves the invoice.
Why we still ship it
Because hygiene is worth doing precisely when it is cheap. An llms.txt takes about an hour to write well, costs nothing to maintain, breaks nothing, and may help AI readers navigate a large site — the same argument that justified robots.txt correctness twenty years ago. It also signals craft: a site that labels itself carefully for machine readers is usually a site whose pages are worth reading. But we file it under hygiene, alongside clean robots.txt and valid schema — the table stakes, not the strategy. Anyone who prices it as the strategy is selling the wrapper, not the work.
What actually decides citations
Two things, in order. Extractable content: facts stated in self-contained, quotable passages — an answer-first paragraph that a system can lift without reassembling your page. Most company sites fail here: the fact exists but is smeared across a paragraph of context, a modal, or a PDF, so the model reads it and still reaches elsewhere for a cleaner statement of the same fact. Third-party signals: independent corroboration that your facts are true and your brand is worth recommending — reviews, comparisons, directory entries, press. A vendor page that states facts nobody else confirms is a claim, not evidence, and platforms treat it that way.
Neither of those ships in a text file. They ship in weeks of rewriting and relationship work, which is why they are the actual product, and why the file is the invoice's smallest line.
The sales test
When anyone — us included, hold us to it — proposes llms.txt as a visibility line item, ask for the number it moves and the measurement behind the claim. If the answer is a citation-rate delta measured on real prompts, fine. If the answer is a paragraph about “AI-readiness,” you are being sold a file, and files are cheap.
Sources & method. The finding that llms.txt has near-zero direct effect on citation rates is from our own tracking measurements across audited sites, before-and-after deployments where llms.txt was the only change. The prescription — ship as hygiene, do not sell as a lever — is our published position, and it applies to our own practice as well as everyone else's.
© 2026 Jun Liu · JUN & GEO. This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0) — https://creativecommons.org/licenses/by-sa/4.0/. You may share and adapt it with attribution and under the same license.