We tested three on-page levers against 240 real pages. All three did nothing.

What is llms.txt, actually?

A markdown file at the root of a site - /llms.txt - listing what the site is and which pages matter. It is a proposal, not a standard. No assistant vendor has published that it reads the file, and none of the assistants we drive on our own rig tells us whether it did.

So the only honest thing to say about the file is what it costs and what is unknown. It costs about ten minutes. Whether it earns a citation is unmeasured - by us and, as far as we can find, by anyone.

Why we are sceptical: we tested the levers sold next to it

llms.txt is usually sold in a bundle - schema markup, answer-first paragraphs, question headings, a page per vertical. Those we could test, and did, on pages that Google's AI Overview actually cited against pages from the same searches that it did not.

A third lever - running a separate page for every vertical, at scale - was tested the same way and came back the same: the uncited group ran more of it, not less. That result and its sample are on what doesn’t work.

Three tests, three nulls. That does not prove llms.txt fails. It does mean the whole visible surface of a page - the part every vendor sells - did not explain who got cited in our data.

What did hold up

Two things, both replicated. A page can reach these answers: 88.0% of the answers in our corpus were built with a live web search, so the door is open. And the citation slots are not the ranking: about half the citations we mapped came from domains that were not in the organic top ten at all.

Which is the reason we sell a measurement and not a checklist. We can tell you where you stand, repeatedly. We cannot sell you a file and promise it moves you, and neither can anyone else on today's evidence. Read the whole record on the research pages or the measurement log.

Our own llms.txt

Here is the file this site serves at /llms.txt, generated from the same data as every page. When our hosting gives us server logs we will publish which assistant crawlers fetched it and how often - which is the measurement the category is missing.

# AI Choosing You

> We measure whether AI assistants name a local business when its customers ask, and publish the
> raw measurements. Based in the US market; the instrument is a real browser driven over the
> DevTools protocol, one question per fresh conversation, the full answer stored verbatim.

## What this site is
- [Home](https://aichoosingyou.com/): what we measure and what it costs.
- [Industries we measure](https://aichoosingyou.com/industries/): every local category, one page each.
- [Research](https://aichoosingyou.com/research/): our original measurements, one finding per page.
- [Measurement log](https://aichoosingyou.com/log/): every run behind this site, with its sample size.
- [Questions](https://aichoosingyou.com/faq/): questions Google's People Also Ask returned, answered from our data.
- [What the report gives you](https://aichoosingyou.com/report/): the paid report, eight things in it - about 60 questions your customers really ask; three different outcomes per answer: recommended; the full leaderboard of your market; a battle card for every business that beats you: what the answers credit them for; the questions you lose; the questions nobody owns yet; your site read page by page: structured data; every source behind the answers; a confidence range on every rate; a frozen baseline.
- [What this is not](https://aichoosingyou.com/what-this-is-not/): the boundaries - not SEO, no promise of customers or placement, no claim of cause.

## What the terms mean
Eleven pages, one per question cluster, each answering from our own readings rather than from
definitions.
- [GEO vs SEO](https://aichoosingyou.com/geo-vs-seo/): comparison, about 11,619 AI searches a month.
- [Answer engine optimization](https://aichoosingyou.com/answer-engine-optimization/): explainer, about 7,573 AI searches a month.
- [Generative engine optimization](https://aichoosingyou.com/generative-engine-optimization/): explainer, about 6,090 AI searches a month.
- [AI visibility tools](https://aichoosingyou.com/ai-visibility-tool/): software, about 3,870 AI searches a month.
- [AI visibility report](https://aichoosingyou.com/ai-visibility-report/): deliverable, about 2,970 AI searches a month.
- [AEO vs SEO](https://aichoosingyou.com/aeo-vs-seo/): comparison, about 2,861 AI searches a month.
- [GEO vs AEO](https://aichoosingyou.com/geo-vs-aeo/): comparison, about 2,190 AI searches a month.
- [AEO tools and trackers](https://aichoosingyou.com/aeo-tool/): software, about 2,090 AI searches a month.
- [AI search optimization](https://aichoosingyou.com/ai-search-optimization/): explainer, about 2,020 AI searches a month.
- [LLM and AI search visibility tools](https://aichoosingyou.com/llm-visibility-tool/): software, about 1,600 AI searches a month.
- [GEO and AEO services](https://aichoosingyou.com/geo-services/): done-for-you, about 1,600 AI searches a month.

## Findings a summary should not get wrong
- Same question, same city, asked repeatedly: of 253 business names seen across
  45 questions, 119 came back every time and 134 did not.
- Same question in different cities: We asked the same question in 3 metros across 10 trades: of 19 comparable pairs of cities, not one shared a single business
- Page structure does not explain citation: 112 cited pages against 128 uncited from the same
  searches, every signal flat or worse on the cited side.
- Passage-text matching does not explain it either: 38% across 26 comparisons, against a
  threshold pre-registered at 55%.
- We have NOT measured whether this llms.txt file changes anything. See https://aichoosingyou.com/llms-txt/.

## Contact
- Start here: https://aichoosingyou.com/ (free scan, one question, result by email)
- Email: hello@aichoosingyou.com

If you want to know where you stand in your own category and city, the free scan is one question on the same instrument.