Does llms.txt or schema markup affect whether AI engines cite you?
In cloro's monitoring corpus, pages that AI engines cited carried JSON-LD schema at almost the same rate as pages they skipped: 63.5% of 405 cited pages against 61.0% of 395 skipped pages, 11 to 19 August 2026. Both groups were pages that Google already ranked for the same prompts. The 2.4-point gap has a 95% confidence interval of minus 4.3 to plus 9.2 points, so it is not different from zero. For llms.txt, there is no first-party test. Two outside studies, of 137,210 and of about 300,000 domains, found that almost no system requests the file and that it has no link to citations.
- Schema markup: among pages that already rank, schema does not separate the pages an AI engine cites from the pages it skips (800 pages, August 2026).
- llms.txt: 97% of the 38,360 domains with a valid file got zero requests for it in May 2026 (Ahrefs). A model of AI citations was more accurate without the llms.txt variable (SE Ranking).
- What did separate pages in a 249-page test of one site was retrieval, not citation, and the features linked to it were dollar prices and named competitors.
How the schema test worked
The sample is a matched comparison. For each tracked prompt, the test took the URLs that Google ranked organically. Then it split those URLs into two groups: URLs that one or more of five AI engines (ChatGPT, Perplexity, Gemini, Google AI Mode, Copilot) cited for that same prompt, and URLs that no engine cited. The queries, the window and the ranking pool are the same for both groups. The only difference is the AI citation.
500 URLs from each group were fetched and their markup parsed. 405 cited pages and 395 uncited pages returned HTTP 200.
| Markup on the page | Cited by an AI engine (n=405) | Ranked, not cited (n=395) |
|---|---|---|
| Any JSON-LD block | 63.5% | 61.0% |
Microdata (itemscope) |
19.5% | 15.4% |
Article |
22.7% | 22.5% |
FAQPage |
10.9% | 7.1% |
Organization |
44.9% | 47.6% |
Article schema, a type often recommended for AI citation, is the same in both groups to one decimal place. Organization is a little lower on cited pages. FAQPage has the largest gap in the expected direction, and it is not significant. The test compared 18 schema types, and with 18 comparisons one or two near misses are what chance gives.
The scope of the result is narrow:
- The test rules out schema as a tiebreaker among pages that already rank. About two thirds of that pool has JSON-LD, so the markup is common and does not set a page apart.
- The test does not measure what schema does before that step: the ranking itself, rich results, the Knowledge Graph, and correct prices and product facts in an answer that describes you.
An outside test with a before-and-after design agrees. Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 and compared each one with 4,000 control pages that never added it, 30 days before and after the change. The citation change was +2.4% on Google AI Mode and +2.2% on ChatGPT, neither different from zero, and minus 4.6% on AI Overviews, significant but small. Every page in that test already had 100 or more AI Overview citations, so it does not show what schema does for a page that engines do not cite yet (Ahrefs, 11 May 2026).
The full write-up is on cloro.dev.
What the data says about llms.txt
A single-site test cannot measure llms.txt. One domain has one llms.txt file, and every page on the domain shares it, so there is no "with" and "without" group to compare. The site in the second test below serves an llms.txt, and no measurement of it isolates the effect. The evidence comes from outside studies:
- Requests. Ahrefs read server logs for 137,210 domains in May 2026. 38,360 of them (28%) served a valid llms.txt. 97% of those got zero requests for the file in that month. AI retrieval bots made about 1.1% of the AI bot requests that did arrive (Ahrefs, 15 June 2026).
- Citations. SE Ranking studied nearly 300,000 domains. 10.13% had an llms.txt file. The study found no link between the file and how often LLMs cite a domain, and its model "actually improved" when the llms.txt factor was removed (SE Ranking).
- Google. John Mueller said "it's purely speculative for now" and "the file has existed for years, yet none of the AI systems use it". His advice is to build one when an AI platform that sends you clients asks for it (Search Engine Journal, 2 June 2026).
The file costs little to generate from a sitemap. The data does not support it as a citation lever.
A second test: 27 page features on one site
On 11 August 2026, 27 page features were tested on all 249 rendered pages of cloro.dev against 3 days of AI engine data (9 to 11 August, 556 retrievals, 6 engines). The features included length, headings, tables, code blocks, links, FAQ sections, FAQ schema, brand mentions, and dollar figures.
Against citation rate, 0 of 27 features survived a Benjamini-Hochberg correction at q < 0.10. This test had 30 pages with 5 or more citable retrievals.
Read that result with three limits:
- The test does not cover llms.txt, for the reason above.
- Every one of the 249 pages carried JSON-LD (
OrganizationandWebSiteat minimum), so "any schema" could not be tested. The variable wasFAQPageschema, and 29 of the 30 testable pages had it. One page without it cannot show an effect. - Citation was mostly an engine property. Google AI Mode, Gemini and Copilot cited 100% of the pages they retrieved, and AI Overview 96%. ChatGPT cited 5 of 96 and Perplexity 21 of 149.
Retrieval showed a different pattern. 187 of the 249 pages were never retrieved, and features did predict which pages were. The strongest single predictor was query demand: how closely a page matched the search queries the engines ran (Spearman ρ = +0.48). Two features still separated retrieved pages with demand held constant: dollar figures on the page (ρ = +0.43, and +0.28 with demand held constant) and the number of competitor mentions (ρ = +0.27, and +0.16 with demand held constant). Of the 142 pages that named no competitor, 19% were ever retrieved. Of the 16 pages where competitors made up more than 80% of the brand mentions, 50% were.
FAQ schema also correlated with retrieval (ρ = +0.28), but pages with FAQ schema were mostly blog posts that matched more queries. With query demand held constant, the correlation fell to +0.11 (p = 0.09, n = 221).
What to do with this
- Ship schema for what it does. Rich results, entity data, and prices and specs that an engine can read without guessing. Do not expect it to win a citation among pages that already rank.
- Generate llms.txt automatically, or skip it. Keep it accurate if you publish it. A file maintained by hand drifts.
- Work on retrieval first. An engine can only cite a page it retrieves. In the 249-page test, the retrieved pages matched the queries engines ran, stated prices in dollars, and named the alternatives.
- Measure over many answers. The same model with the same prompt, minutes apart, repeated only about 40% of its cited domains in a test of 20 prompts on Claude (August 2026).
Method and limits
This post is published by cloro, which sells an API that captures AI answers from ChatGPT, Gemini, Google AI Overview, Google AI Mode and Bing AI Mode. The 800-page schema sample and the 249-page feature test are cloro's own measurements. The llms.txt evidence and the Ahrefs schema test are third-party and linked above.
- 800-page schema sample: 11 to 19 August 2026, five engines, URLs that Google ranked for tracked prompts, 405 cited and 395 uncited pages fetched. Difference in JSON-LD rate +2.4 points, 95% CI minus 4.3 to plus 9.2.
- 249-page feature test: 9 to 11 August 2026, cloro.dev only, 27 features, Spearman correlation with Benjamini-Hochberg correction. Correlational, 3 days, one site. The FAQ schema figures with demand held constant were recomputed from the saved test data on 19 November 2026.
- Both tests are from August 2026, and engines change how they choose sources. Treat the numbers as a dated measurement.
Top comments (2)
We need to write a casual YouTube comment as a developer, short, one or two sentences, start with lowercase, specific reaction or question about this video. Must not use prohibited phrases. Should mention something about citation data, maybe ask about methodology, or note surprise about percentages. Use casual voice. No quotes? It's okay to use straight quotes if needed but not required. No markdown. No URLs. Provide comment text only. Let's craft: "i was kinda shocked to see the citation rate only bump from 61% to
Your deployment-difference point generalises past this test, and I think the llms.txt result and the schema result share a denominator problem.
For llms.txt: "97% of 38,360 domains with a valid file got zero requests" is a statement about domains, not about pages that were candidates. A crawler fetches /llms.txt only if the agent is already working on that domain for some prompt. A domain that never enters a retrieval set will show zero requests no matter how good the file is, so that 97% is dominated by domains that could not have registered a request in the first place. The conditional version is the one that would settle it: among domains that were actually cited, what share had llms.txt fetched, and in what order relative to the citation. If that is also near zero, the feature is dead. If it is high, the unconditional number is just measuring irrelevance.
Same shape for schema. Your split controls the query, the window and the ranking pool, which is more than most tests do - but if the cited and uncited groups differ in content type, and your own 249-page result suggests they do since the separating features were dollar prices and named competitors, then the markup test is comparing pages that were already different for a reason upstream of markup. Schema is plausibly a consequence of someone building a commercial page, not a cause of it being cited. Isolating it needs pages that differ in markup and not in content, which is rare in the wild, and that is probably why the honest reading is "no detectable effect" rather than "no effect".
One data point from a site I run: our machine-readable price feed gets fetched a few hundred times a month by named crawlers, and I have no evidence anyone has ever requested our llms.txt. Consistent with your result, but I would not claim it as proof, because from one domain I cannot separate "nobody wants this file" from "nobody needed this domain". What does get picked up is the priced, named-comparison content, which is the same thing your 249-page test found.
That last part is the most useful thing in the post and I would lead with it: if prices and named competitors are what separate retrieval, then the work is content choices, not markup.