llms.txt Experiments: What the AI Visibility Evidence Shows

Do llms.txt experiments show a benefit for AI visibility? Four public reports test different stages: file requests, page discovery and citations. Separating those outcomes explains why a successful guided demo and an absence of ordinary crawler requests can both be true.

This review summarizes each report, shows what they have in common and lists what they cannot establish. It ends with a test protocol you can run on your own site. Named practitioners published these reports; SEOVision did not run them, so treat every figure as the author's claim.

The lesson in one line

Test whether the file is requested, whether it helps discovery and whether visibility changes, and treat them as three separate outcomes. A result on one stage says little about the other two.

What four llms.txt experiments reported

Each report below covers a different setup and a different stage. Read them as four narrow observations rather than one verdict.

Lee Davies: explicit instructions worked, ordinary discovery did not

Lee Davies describes building a path from llms.txt through a manifest to structured answers. When he told an assistant to follow that path, it could use it. However, when asked an ordinary customer question without instructions, it fetched the homepage and stopped.

So a guided demonstration was different from spontaneous discovery. The report shows that the file can work as an instruction aid, not that assistants look for it on their own. Read the original post by Lee Davies.

Zoë Elizabeth Blogg and Reboot: three months without discovery

Zoë Elizabeth Blogg reports a two-domain test by Reboot. The team listed new pages only in llms.txt, with no internal or external links pointing to them. Over three months, logs showed bots visiting the sites but not requesting the file or its listed pages.

That finding concerns one discovery setup. It does not test sites where the listed pages are also linked normally, or agents that a user directs to the file. Read the original post by Zoë Elizabeth Blogg.

GO MO Group: no measurable lift over six months

GO MO Group reports a six-month experiment across clients in different industries. Their conclusion is that implementing the file produced no measurable AI-visibility benefit.

The public post gives the conclusion and links to a longer write-up. Still, it does not include enough detail to estimate what effect size the test could detect, or to conclude that every AI platform ignores the file. Read the original post by GO MO Group.

Hristo Stanchev: a positive result with another likely cause

Hristo Stanchev reports roughly three times as many AI citations for a dynamic setup as for a static one across two sites. That sounds like a win for the file at first glance.

Yet he says explicitly that nothing proves the file caused it. The same pipeline also improved sitemaps, stale-URL handling and crawlable product content, so it is a bundled freshness change rather than a clean test. Read the original post by Hristo Stanchev.

The four reports at a glance

Side by side, the reports clearly test different stages. Use the table to see which question each one can answer.

What the experiences have in common

The shared lesson is that existence, access, discovery and selection are different stages. A successful fetch establishes access. It does not show that an assistant would find the file unprompted, use its links or recommend the business. Equally, no request during one test cannot prove that no agent will ever use the format.

The positive report sharpens the question rather than resolving it. When several publishing changes ship together, a second test must isolate the one that mattered. Otherwise a convenient filename gets credit for ordinary improvements elsewhere on the site.

For Google specifically, use the platform guidance below rather than interpreting all assistants as one system.

What these llms.txt experiments cannot establish

These are small or incompletely specified observational tests, with different bots, prompts and endpoints. In addition, server user-agent strings alone do not prove which crawler made a request. So the reports support neither a guaranteed visibility boost nor a universal claim that the file is useless.

Set the platform scope before spending time on a test

For Google Search, the current AI optimization guide says Google ignores llms.txt and that it does not help or harm visibility. This is a platform-specific statement. It does not describe every assistant or user-directed agent.

The Reboot experiment methodology describes two existing domains and four isolated pages on each, monitored for three months. With no ordinary links to those pages, the setup tests file-led discovery. It cannot estimate the contribution of llms.txt when normal links already provide another discovery route.

A test you can run: proposed protocol

If you want your own answer, run a test that keeps the three questions apart. These five steps are a starting design:

1. One hypothesis: test discovery of a controlled page set, agent task completion or a change in citations, but never all three as one score.
2. A baseline: record logs and repeated answers for a fixed prompt set, and separate crawler requests, user-triggered fetches and monitoring tools.
3. Stable conditions: keep HTML, sitemaps and publishing frequency stable while you add the file, and label any freshness change as a separate intervention.
4. Separate tracking: count file requests and later page access apart from citations and referrals, and keep guided prompts in their own group.
5. Honest reporting: report a null result with its site count, duration and limits, and treat a positive change as provisional until repeated.

Before the first log entry, make sure the file itself is valid. A test of an unreadable file measures nothing.

Conclusion

These llms.txt experiments do not establish a dependable citation or traffic lift. Guided use, unprompted discovery and downstream visibility are different claims, and the positive report bundled other changes. Google's current guidance also removes the case for treating the file as a Google Search requirement.

Keep a low-cost file if you have a specific agent use case, but set a time budget and a measurable endpoint. Validate access, separate guided requests from ordinary traffic, and log concurrent changes. Prioritize accurate, accessible pages; stop investing in the file when your chosen test shows no useful benefit within its documented limits.