What does publishing llms.txt prove about AI visibility?
Publishing llms.txt establishes that you served a file; it does not establish that an AI engine used it or changed its answer. Housecall Pro was named in 108 of 150 answers and own-domain cited in 43 of 150, captured on 22 September 2026. Neither count identifies the file's role, so it cannot be used as evidence of a gain caused by publication.
What we measured
Measured. Housecall Pro's September record covers ten fixed field service questions, three repeats on ChatGPT, Claude, Perplexity, Google AI Overviews and Google AI Mode. For naming, the observed rate was 72.0 percent and its Wilson 95 percent interval ran from 64.3 to 78.6 percent. The separate own-domain citation rate was 28.7 percent, with a Wilson 95 percent interval of 22.0 to 36.4 percent. The source does not record an llms.txt intervention or compare treated and untreated pages.
The two events share a valid base but mean different things. A citation observation is not a measurement of every file fetched, nor evidence that a source guide caused the vendor to be named.
What this means for you
Reasoned. Match each conclusion to its receipt. A successful public fetch can demonstrate the file's content and delivery. A verified crawler log may demonstrate a particular request under the log's identification rules. A retained answer can demonstrate what the engine said and cited at that time. These records should not be substituted for one another.
The same limit applies to llms-full.txt and other full-text exports. More text in a served artifact does not turn its publication into evidence of retrieval, comprehension or recommendation. Review the exported material for current facts, useful structure and permission to publish it.
If you later observe a changed answer, preserve the question, engine, date, answer text and sources. Other page changes, engine changes and ordinary answer variation can coincide with publication. A before-and-after difference alone does not isolate the file as the cause.
Accept an implementation project using the deliverable you can inspect: the agreed content, working destinations and served version. Keep any unmeasured effect on future AI answers unknown. That gives both publisher and buyer an honest record of completed work.
- For a publisher: record the served file separately from a later answer.
- For a buyer: ask which claimed effect has its own evidence.
- For an agency: report delivery and observed AI outcomes in separate fields.
Also asked as
Does llms.txt improve AI visibility?
Do llms.txt and full-text variants improve AI citations?
Related
Next step
Use the Free 10-question check (one engine). No email is needed; runs are limited per network and per day, and by the site's daily allowance. Save the result as one dated answer rather than attributing it to a recently published file.
For one supported page change with an inspectable served result, Fix one answer, $490.
Every figure comes from the published method, v1.1.
Claiming or correcting a vendor record is free and never depends on a purchase.
Prepared by the Broadcastwell team.
Page owner: Rohan Iyer, AI Engineer. Corrections: hello@broadcastwell.com