Why Publishing llms.txt Does Not Prove AI Visibility

A specific pattern keeps repeating in AEO conversations: a team ships llms.txt, marks 'AI visibility' done, and moves on. The file is fine. The checkbox is the problem — it substitutes an easy, verifiable action (we published a file) for a hard, measurable outcome (engines cite us), and the substitution feels so productive that the real work stops.

What llms.txt actually is

A markdown file at a known path listing your site's key pages with descriptions — a curated map offered to language-model systems, by convention rather than standard. The analogy to sitemap.xml flatters it: sitemaps have two decades of documented consumption by named crawlers; llms.txt consumption is, at time of writing, sparsely and unevenly documented at best. Publishing one costs little and may help systems that choose to read it. That's the honest whole of the claim.

What no file can do

• Make a blocked crawler retrieve you — if the WAF 403s GPTBot, llms.txt is a menu posted outside a locked restaurant
• Make client-rendered answers visible — content that only exists after JavaScript stays invisible regardless of what any index file promises
• Earn citations — engines cite retrievable, quotable, trustworthy pages; a file listing your pages changes none of those properties
• Serve as measurement — file exists is a deploy check, not a visibility metric; the metric is citations observed on real prompts

The checklist that measures what llms.txt doesn't

If the goal behind shipping llms.txt was 'be visible to AI engines', the measurable version is: retrieval verified per crawler (the four checks in our retrievability guide), answers present in server-rendered HTML, a 20-prompt citation baseline run and logged, and movement tracked monthly against that baseline. Four items, all measurable, none satisfied by publishing a file. Do these, and llms.txt becomes what it should be — a harmless courtesy on top of a real program.

A failure worth checking

The checkbox cascade: llms.txt ships, the AEO line-item closes, and six months later someone asks why citations are zero. Investigation finds the actual blockers — edge bot-blocking, client-only rendering — untouched since before the file existed, because the file's existence answered the question everyone stopped asking. The failure isn't publishing llms.txt; it's letting an unmeasurable action retire a measurable goal.

FAQ

Should I publish llms.txt at all? — Sure — it's cheap, standards-adjacent, and might help systems that read it. Fenn's own sites publish it. Just file it under courtesy, not strategy, and never let it close a visibility ticket.

Is there any evidence engines consume llms.txt? — Consumption reports are anecdotal and engine behavior changes without notice — which is precisely why the file can't serve as a metric. Track what engines verifiably do (retrieve your pages, cite your URLs), not what they might read.

What should our AEO success metric be instead? — Citations on a fixed buyer-intent prompt set, tracked over time, with mentions counted separately — reproducible, comparable, and connected to actual discovery. Our prompt-set and evidence-log guides are the free templates.