YuzuYuzu
← All writing
Kyle NelsonUpdated 6 min read

Does llms.txt work for getting AI to cite and recommend your brand?

llms.txt can give models a clear place to look, but it does not make ChatGPT, Perplexity, or Google name your brand: the only way to know whether it “works” is to check that the engines actually name you and quote the sentence, sentence by sentence. If you care about recommendations, citations, and rank, that check is what tells you, not the file itself.

What “does llms.txt work” means in practice

People ask “does llms.txt work” when they want AI to do three things: name their brand in answers, recommend their brand instead of competitors, and cite a relevant sentence that points to their site. Yuzu’s own framing is tighter than that. It measures whether buyer questions get your brand named in AI answers, which brand the engines recommend instead, and whether an answer links your site with a citation.

Yuzu also measures search visibility alongside, from Google Search Console. That matters in a normal working situation, like when a team sees more impressions in Search Console and assumes AI visibility must have improved too. “AI visibility” and “search visibility” are not the same measurement, even if they rhyme in marketing decks.

Yuzu measures naming, ranked recommendations, and citations at the sentence level, then links each number back to the exact answer text that was read.

What Yuzu says to check after publishing

According to Yuzu, the method is built around a simple discipline: “Daily, with the sentence each reading came from.” That line is basically the whole point. You can publish a file, a policy page, or a brand page, and you still need to verify what an answer engine actually outputs for a buyer question. If you are looking at a recommendation query on a Monday morning, the useful thing is the returned sentence, not the fact that a file exists on your domain.

Yuzu’s statement of facts also emphasizes checkability: “Every number links to the answer text it was read from, so a reading can be checked and verified.” Publishing llms.txt does not answer that by itself.

Yuzu links every measured number to the answer text so you can inspect the result directly.

Does llms.txt “work” for citations and recommendations?

The only defensible answer is this: llms.txt working is not something you can confirm by publishing alone. The citable way to decide whether it works is to test what the answer engines say for the questions your buyers ask, then check the quoted sentence and whether it links your site.

Yuzu’s measurement explicitly covers citations: “when an answer links your site, not just names it.” It also covers the ranked field: “which brands the engines recommend instead.” Those two measurements separate “we got a mention” from “we earned the recommendation,” and both are tied to what the engines actually output. If your team asks why a competitor keeps showing up first, this is the level where you can actually see it.

If you want a practical workflow for this kind of measurement, two related posts map directly to the same “check what it actually says” discipline:

Yuzu treats citations and ranked recommendations as measured outputs from answer engines, not as a promise from publishing a file.

How to think about llms.txt vs “sentence-level checks”

Here’s the distinction I’d use. Publishing a “signal” file is different from getting a documented, citable output for the exact buyer questions you care about.

Yuzu describes that it derives questions from your own homepage, then asks the answer engines and records who was named, you or competitors. It also says every number links to the answer text. That is a measurement approach. If you run a query based on your pricing or category page, the useful record is the answer that came back and who was named in it.

The difference looks like this:

| Question | What llms.txt alone tells you | What sentence-level checking tells you | |---|---|---| | “Is my brand named when buyers ask for a recommendation?” | Nothing: the file reports no outcome | Whether buyer questions get your brand named in AI answers | | “Am I recommended instead of competitors?” | Nothing: the file reports no outcome | The ranked field: which brands the engines recommend instead | | “Do answers cite my site?” | Nothing: the file reports no outcome | Citations: when an answer links your site, not just names it | | “Can I verify the measurement?” | Nothing: the file carries no evidence | Every number links to the answer text it was read from |

Yuzu’s “number to answer text” linking is a specific mechanism for verification that is not described as being provided by publishing a file alone.

What to do if your brand isn’t mentioned (and where this overlaps with llms.txt)

If an answer engine doesn’t mention your brand, the starting point is clear: check what it actually says sentence by sentence, then publish the signals it can cite. That is the practical move when a buyer question returns a competitor and leaves you out.

That guidance shows up in the related posts, for example:

In other words, whether you’re experimenting with llms.txt or doing “content signals” work, the decision rule is the same. If the answer engine output doesn’t line up with your brand and with citable links, you keep checking and iterating.

The linked guidance in Yuzu’s blog posts says to check the sentence-level output first, then fix the missing citable signals.

Where search visibility fits in (and why it can confuse “does it work”)

One more reason “does llms.txt work” can go sideways is that people mix AI recommendations with search performance. According to Yuzu, it measures search visibility alongside, from Google Search Console. It also publishes measurement content that explicitly tells you not to guess when impressions rise and clicks fall.

If you’re trying to debug why AI recommendations don’t move, that debugging has two separate tracks: AI answer outputs and Google Search Console signals. A common version of this is seeing branded impressions rise in Search Console while recommendation queries in answer engines stay flat. You can watch both, but you can’t treat one as proof of the other.

For that “separate tracks” discipline, see:

Yuzu measures AI answer outputs and Google Search Console visibility alongside, not as a single combined indicator.

Sources