llms.txt and Local SEO: What the Evidence Actually Shows

A local business owner or agency told to add an llms.txt file so AI assistants can read their site should not spend the hour. Three independent studies reported in our discourse corpus concluded it has no measurable effect, and the practitioner room has moved from debating it to using it as a credibility test.

We also checked the local search industry itself. Across 1,921 published URLs from the seven local rank-tracking vendors in our competitor corpus, scanned August 2026, exactly one publishes an llms.txt file — and it is the only one of the seven that mentions the format anywhere on its site.

  LLMS.TXT ACROSS SEVEN RANK TRACKERS
  ┌──────────────────────────────────┬───────┐
  │ vendors publishing an llms.txt   │ 1 / 7 │
  │ vendors mentioning it at all     │ 1 / 7 │
  │ published URLs scanned           │ 1,921 │
  └──────────────────────────────────┴───────┘
  our own corpus · scanned August 2026

What It Is Supposed to Do

llms.txt is a proposed convention: a markdown file at the root of a site listing the pages an author considers most useful, so that a language model reading the site has a curated map rather than a crawl.

The idea is reasonable and the analogy to robots.txt is what sells it. The difference is that robots.txt is honoured because crawlers were built to read it. llms.txt is honoured only if the systems it addresses chose to implement it, and that is an empirical question rather than a design one.

It is worth being clear about who is claiming what here. No vendor selling local search tooling has published evidence that it works, and the vendors are otherwise happy to publish — the same seven-vendor corpus is the one we used to show that none of them publishes a cost per scan either.

What Three Studies Reported

  WHAT THREE STUDIES REPORTED
  ┌──────────────────┬─────────────┬─────────────┐
  │ study            │ scale       │ verdict     │
  ├──────────────────┼─────────────┼─────────────┤
  │ 137k-site scan   │ 137,000     │ 97% unread  │
  │ SE Ranking       │ 300,000     │ no effect   │
  │ Rankability      │ adoption    │ unchanged   │
  └──────────────────┴─────────────┴─────────────┘
  reported in our discourse corpus, not read by us

Three separate analyses, at large scale, reaching the same conclusion by different routes: one measuring whether the files are fetched at all, one measuring whether publishing one changes anything, one measuring adoption.

We have not read these studies directly, and we are not going to present their figures as ours. What our corpus holds is how they were reported and discussed by practitioners, which establishes the state of the professional consensus rather than the underlying measurement. If you need the numbers for a decision that matters, read the studies rather than this page.

What the consensus establishes is enough for the practical question. Three independent attempts to find an effect, none finding one, is the shape of a tactic that does not work.

What the Room Actually Says

This is a different corpus from the local-search one used elsewhere on this site, and it is worth naming precisely: our AEO tool-landscape research, a set of 253 Reddit threads about AI-search tooling, each tagged by the friction it expresses. Of those, 22 are about llms.txt, and the single most common friction in them is does_it_work — not which generator to use, not how to implement it, but whether the thing does anything at all.

The titles carry the verdict without needing to be opened:

• "Is llms.txt file a scam?"
• "Stop wasting your time on fake AEO/GEO Tricks (Yes, including llms.txt)"
• "How to spot an SEO noob… they're telling you to implement llms.txt"
• "Did llms.txt actually change anything, or is it a 20 minute feel-good task?"

That last one names the real cost. Twenty minutes is cheap, which is exactly why the tactic survives: it is affordable enough that nobody demands evidence before doing it. The expensive part is that recommending it now marks the person recommending it — it has become a negative credibility signal in the practitioner community rather than a neutral one.

Why It Persists Anyway

Two forces keep it alive, and neither is evidence.

1. Vendors ship generators for it. Two AEO vendors in the same tool-landscape research offer a free llms.txt generator, and one of the seven local rank trackers publishes its own file alongside a support page explaining the format.
2. People search for it. The keyword llms txt generator draws roughly 1,900 searches a month in India against 590 in the United States — figures from the keyword research behind that same tool-landscape work. Real demand exists for something the informed half of the market has already written off, which is precisely the condition that keeps a free tool profitable to build.

There is a small irony worth noting from our scan. The one vendor publishing an llms.txt also lists that file in its XML sitemap — submitting a file written for language models to the search index, which is not what either format is for.

What to Do With the Hour Instead

The underlying goal — being usable by an AI assistant that reads your page — is legitimate. The things that serve it are boring, textual, and have evidence behind them.

  WHERE THE EFFORT ACTUALLY PAYS
  ┌──────────────────────────┬──────────────────┐
  │ llms.txt                 │ no measured lift │
  │ citing your sources      │ measured lift    │
  │ numbers as page text     │ readable at all  │
  │ answer in first 100 words│ reaches the model│
  └──────────────────────────┴──────────────────┘
  the last three are cheap and they are text

The second row is not a guess. A 2024 study presented at KDD tested nine content tactics across roughly 10,000 queries against a generative search engine and found citing sources, adding quotations and adding statistics each lifted visibility by around 30 to 40 percent — while keyword stuffing was the only tactic tested that made things worse.

The third row is the one local businesses fail most often. A ranking figure, a review count or a service area that exists only inside an image cannot be read by a system that answers questions about you. That single constraint is why our own posts draw their figures as text rather than as graphics, and it applies to any page describing Google Maps ranking factors.

For Local Businesses Specifically

Nearly everything an assistant knows about a local business comes from sources you do not host. Your Google Business Profile, your reviews, your categories and your hours are read from Google, not from a file on your server — which caps what any root-level text file could achieve for a local business even if it were honoured perfectly.

The work that moves local visibility is the same work either way: an accurate and complete profile, a category that matches what people search, reviews that keep arriving, and pages that state facts as text. That order of work is in how to rank in the Google map pack, and measuring whether any of it worked is what geo-grid rank tracking is for.