Does LLMs.txt improve AI search visibility?

The evidence does not show that an LLMs.txt file improves AI search visibility on its own. In our two ecommerce studies, brands with shopping-agent instructions were recommended more frequently when mentioned, but they did not consistently appear earlier in answers. The observed differences do not prove that the files caused them.

That distinction matters. Being recommended once an assistant has named your brand is not the same as being more likely to appear across all its answers. A file can be useful documentation without being a demonstrated ranking tactic.

What is LLMs.txt?

LLMs.txt is a proposed Markdown overview of a website, intended to help language models and agents find useful information. The familiar /llms.txt path can contain a short description and links to important pages. /llms-full.txt can provide a fuller content version.

Commerce documentation can serve a different purpose: telling a shopping agent where to find products, policies or supported shopping capabilities. We call that content shop-style agent instructions. Its presence does not tell us whether an assistant actually retrieved or followed it.

Platform behaviour also matters. Shopify documents generated agent-discovery URLs, with /agents.md as its canonical location and LLMs.txt paths offering compatibility. By default, the three paths can return the same content. An absent root /llms.txt is therefore not sufficient evidence that a store has no agent documentation elsewhere. Our audit did not check /agents.md.

What we examined

In October 2026, Citations.io checked /llms.txt and /llms-full.txt on 1,560 verified brand websites. Of those, 339 (21.7%) returned one or more files. 253 (16.2%) contained shop-style agent instructions. Those 253 websites are a subset of the 339, not another group to add to the total.

At 454 websites (29.1%), both checked paths were confirmed absent. Inaccessible or unresolved checks were not counted as absent. These were checks of file availability, not measurements of engine use.

We then compared file cohorts with brand observations in two existing ecommerce studies. We examined recommendations, position and sentiment separately, rather than merging them into one claim about visibility.

The comparison is observational. AI answers were collected in July and September 2026, before the October file audit. We do not know whether the files were available when those answers were generated. That timing limits every conclusion below.

Are brands with files recommended more often?

Among observed brand mentions, the shop-instruction cohort had a higher recommendation rate in both studies. A recommendation means the answer suggested the brand as a choice, rather than merely naming it.

In Ecommerce study 01, brands with shop-style instructions were recommended in 1,509 of 2,411 mentions: 62.6%. The confirmed-no-file group was recommended in 15,674 of 25,816 mentions: 60.7%. The difference was 1.9 percentage points.

In Ecommerce study 02, the corresponding figures were 934 of 989 mentions: 94.4%, against 2,079 of 2,440: 85.2%. The difference was 9.2 percentage points.

These denominators are brand mentions, not all AI answers. They answer: “When these brands were mentioned, how often were they recommended?” They do not answer: “How often did these brands appear in the first place?” An answer can mention several brands, and a brand can recur across answers.

The larger difference in study 02 is not evidence of a stronger file effect. Its brands, prompts, category and engine coverage differed. These are descriptive cohort comparisons, not controlled estimates or a promise of an uplift for your store.

Do they rank higher in AI answers?

Not in these comparisons. Among mentions with a recorded rank, the mean answer position for shop-style instructions was 3.30, compared with 3.01 for confirmed-no-file brands in Ecommerce study 01. In Ecommerce study 02, it was 3.42, compared with 3.24.

Lower means earlier in the answer, so the file cohort appeared slightly later on average in both studies. These are positions within sampled answers, not Google rankings.

Position is conditional on having a recorded rank. Unranked mentions and brands absent from an answer do not enter that mean. It cannot tell us a brand’s overall probability of appearing. Nor does it prove that a file made a brand appear later: the same differences between cohorts could influence position.

What about sentiment?

Among brand mentions with known sentiment, positive sentiment was 65.9% for shop-style instructions versus 57.1% for confirmed-no-file brands in Ecommerce study 01. Ecommerce study 02 recorded 94.8% versus 83.4%.

These are positive-label rates among mentions where sentiment was known. Unknown sentiment is excluded; the recommendation denominators above must not be reused for this measure. Positive wording also does not necessarily mean an explicit recommendation.

The association is worth investigating, but sentiment can reflect product category, brand reputation, prompt wording and the sources available to an assistant. It does not demonstrate that instructions changed how an engine felt about a brand.

Why correlation is not causation

Brands that maintain agent documentation may differ from those that do not. They might use different commerce platforms, have more complete product information, attract different third-party coverage or belong to different market segments. Any of those factors could be associated with the observed outcomes.

The two studies also involved different prompts, brands, engines and seasonal retail categories. We did not statistically control these differences or randomly assign files to otherwise comparable stores. Repeated mentions are not independent experiments.

Most importantly, we have no historical baseline of file availability. Finding a file in October cannot establish that it existed during a July or September answer collection, let alone that an engine read it. We cannot infer a causal effect from this design.

A stronger test would record documentation before answer sampling, track changes over time, keep the prompt panel and engines stable, and compare suitably matched brands. Even then, a change in results would require careful interpretation.

Should ecommerce brands create these files?

Treat accurate agent documentation as one possible technical input, not a guaranteed ranking service. If your platform already generates it, inspect the content before paying someone to create another file. Check product links, policies and capabilities, and keep them correct as the store changes.

Do not prioritise a file over accessible pages, clear product facts or useful third-party sources. Documentation cannot compensate for contradictory availability, unclear returns or content an agent cannot reach.

For Google specifically, its June 2026 guidance, recorded in Search Central updates, says these files neither help nor harm Google Search visibility or rankings because Google Search ignores them. That is guidance about Google, not proof of how every other assistant operates.

For an ecommerce team, the practical order is simple: make core information accessible, make it accurate, understand the sources used in your category, and then maintain any relevant agent documentation. Our ecommerce overview explains the broader measurement context.

What marketers should measure

Start with what AI visibility means, and keep the following questions separate:

• Visibility: in what share of sampled answers does your brand appear, including answers where it is absent?
• Recommendations: when mentioned, is it suggested as a useful choice or merely listed?
• Position: where does it appear among mentions with a recorded rank?
• Sentiment and accuracy: how is it described, and are product facts correct?
• Citations and competition: which sources are cited, and which competitors appear alongside you?

Use repeat sampling of relevant buyer questions. Keep prompts and engine coverage comparable over time, retain the answer text, and report denominators. Read our measurement methodology before treating a short-term change as meaningful.

A documentation change is something to record alongside your measurements, not something to assume caused them. Answer-level evidence gives marketers a better basis for decisions than a checklist of fashionable files.

Methodology in brief

The audit covered 1,560 verified brand websites at the two LLMs.txt paths in October 2026. File-returning and shop-instruction cohorts overlap; confirmed absence requires both paths to be absent. Unresolved access is not absence, and /agents.md was not checked.

Ecommerce study 01 concerned UK retail and gifting, using ChatGPT, Claude, Gemini and Perplexity. Ecommerce study 02 concerned US beauty and skincare, using ChatGPT and Gemini. The original studies recorded 9,957 and 4,903 usable AI answers respectively, 14,860 in total, during July and September 2026.

Across both studies, 481,654 extracted components comprised 112,583 identified brand references, 112,583 recommendation labels, 112,573 sentiment labels, 64,899 recorded answer-position items and 79,016 citation-source occurrences. This is 480,000+ extracted datapoints, not independent AI answers: several components can come from the same answer.

Recommendation rates use observed mentions in each compared cohort. Mean position uses ranked mentions only; positive sentiment uses mentions with known sentiment only. The figures are descriptive, rounded as shown, not statistically controlled or causal. Different prompt panels and later file checks prevent attributing outcomes to files.

Frequently asked questions

Does Google use LLMs.txt to rank results?

Google says it does not use these files for Search visibility or rankings, including its generative AI capabilities. Creating one is not a Google ranking requirement.

Is LLMs.txt the same as robots.txt?

No. LLMs.txt is a proposed information format. Robots.txt communicates crawler access rules. Neither should be mistaken for a guarantee that content will appear in an answer.

Do all Shopify stores need a separate file?

Not necessarily. Shopify provides agent-discovery documentation by default, with compatibility paths. Check what your store serves and whether it is accurate before adding another solution.

Does a file guarantee an AI assistant will read it?

No. Availability is not proof of retrieval, use or recommendation. Our checks establish what the website returned at the checked paths, not what an engine consumed.

Track how AI engines mention, recommend and cite your brand. See your brand’s AI visibility, explore plans or request a demo. Start with evidence you can inspect, not a promised ranking shortcut.