Google’s May 2026 AI optimisation guide says llms.txt, chunking and special schema aren’t needed. Has anyone got data contradicting it for other AI platforms?

1 reply

• 

  adminAdmin

  Original poster 1 day ago

  The short answer is no: I have not found credible data showing that llms.txt or “special” AI schema improves visibility on ChatGPT, Perplexity or Copilot. The better datasets published since late 2025 point the same way as Google’s guide. Chunking is the one area with a genuine argument, and most of that argument is about what the word means rather than conflicting data.

  The important qualification is scope. Google’s guide speaks for Google Search only, and other platforms retrieve content differently (live fetches that often skip JavaScript, Bing’s index, third-party search providers). So “Google doesn’t need it” does not automatically mean “nobody uses it”. But the burden of proof sits with the tactic, and the evidence for other platforms is currently null, tiny, or correlational.

  What Google’s guide says, and what it deliberately leaves out

  Google’s guide to optimising for generative AI features (May 2026) says you do not need llms.txt or other AI text files, Markdown versions of pages, artificial chunking, special schema, or AI-specific rewording. It still recommends normal structured data for rich result eligibility. Search Engine Journal’s coverage makes the obvious point that this does not settle the question for non-Google platforms, which may weight signals differently.

  There is also a small internal inconsistency worth knowing about. Lighthouse 13.3 added an experimental Agentic Browsing category that checks llms.txt validity if the file exists. That is an audit for AI agents operating a browser, marked as under development, not a Search ranking or citation signal. SEJ covered the mixed messaging, but it is not evidence that llms.txt affects AI answers.

  The evidence outside Google Search, tactic by tactic

  Tactic	Best evidence for non-Google platforms	Evidence type	What it supports
  llms.txt	SE Ranking (about 300k domains) and Ahrefs server logs (137k domains)	Large datasets, correlational plus direct request logs	No measurable citation link; AI retrieval bots rarely request the file
  Schema	Microsoft statement (2025); Ahrefs difference-in-differences test; two small direct-fetch tests	Platform statement, quasi-experiment, single-page tests	Standard schema may aid understanding on Bing; no measured citation lift on ChatGPT
  Chunking	Mike King’s retrieval argument; Kevin Indig’s ChatGPT citation analysis	Technical reasoning, large observational dataset	Clear passage-level writing is associated with citation; no evidence that fragmenting pages helps

  llms.txt: the bots mostly are not asking for it

  A file can only influence a system that fetches it. That makes llms.txt unusually easy to check, and the checks so far are not encouraging.

  SE Ranking analysed roughly 300,000 domains and found about 10% had an llms.txt file, with no measurable relationship between having one and AI citation frequency. Removing the llms.txt variable actually improved their model’s accuracy.

  The Ahrefs study is stronger because it uses request logs rather than correlation. Across 137,210 domains using Ahrefs Web Analytics in May 2026, 28% published an llms.txt file, and 97% of those files received zero requests that month. Of the requests that did arrive, SEO audit tools made 21.7%, while retrieval bots made 1.1% and training crawlers such as GPTBot and ClaudeBot made 5.3%. Ahrefs also found zero AI bot requests for llms.txt files that did not exist, which suggests the bots are not even probing for it. Ahrefs notes its customer base skews technical, so treat the adoption figure as an upper bound.

  I could not find documentation from OpenAI, Anthropic, Perplexity or Microsoft stating that llms.txt is used for search retrieval or citation selection. The one place it has a plausible job is developer documentation read by coding agents. In the Ahrefs data, AI agents and agentic infrastructure were the largest AI category at 10.5% of requests, which fits that use case.

  Schema: a real Microsoft statement, but no measured citation lift

  This is the closest thing to a platform-level contradiction. At SMX Munich in March 2025, Microsoft’s Fabrice Canel said schema markup helps Microsoft’s LLMs understand content. Two caveats matter. He was talking about ordinary schema.org markup, which Google also recommends, not a special AI vocabulary. And “helps understand” is not the same claim as “increases citations”.

  When citations were measured, the effect was hard to find. Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026, matched each against three control pages, and compared 30 days before and after. AI Overviews showed a small but statistically significant 4.6% decline relative to controls. AI Mode (+2.4%) and ChatGPT (+2.2%) were indistinguishable from noise. The limitations are real: only pages already heavily cited by AI (100+ citations), all schema types pooled, a 30-day window, and Perplexity and Copilot were not included. So for Copilot specifically, the honest position is “plausible, unmeasured”.

  The small direct-fetch tests disagree with each other, which is informative in itself. SearchVIU (late 2025) found none of five AI systems extracted a price that existed only in JSON-LD, and only Gemini rendered JavaScript during live retrieval. Mark Williams-Cook then showed ChatGPT and Perplexity pulling an address out of invalid JSON-LD, treating it as plain text on the page. Both are single-page tests, so neither generalises far. The overlap between them is the useful part: when an AI fetcher reads your markup at all, it appears to read it as text, not as a trusted structured record.

  Chunking: a dispute about definitions more than data

  Google’s position is explicit. Danny Sullivan said on Search Off the Record in January 2026 that Google does not want content broken into bite-sized chunks for LLMs, and the May guide repeats it.

  The counter-argument is technical rather than empirical. Mike King argues that retrieval systems work at passage level, and demonstrates that splitting a multi-topic paragraph improved cosine similarity for a target query by 19.24% in his example. That is a retrieval proxy on a constructed example, not a measured change in citations.

  The largest observational data comes from Kevin Indig’s analysis of 3 million ChatGPT responses, narrowed to 18,012 verified citations: 44.2% came from the first 30% of the content, and cited passages used definitional language and question-style headings more often. That tells you what cited passages look like. It does not show that restructuring an existing page changes whether it gets cited.

  Separate two things people call “chunking”:

  Writing sections that each answer one clear question under a descriptive heading is good for readers and for any passage-based retrieval system. Fragmenting a page into thin standalone snippets, or producing a second machine-oriented version, is what Google is warning against. The evidence supports the first. Nothing I have found supports the second on any platform.

  Why these tactics would rarely move citations much

  It helps to picture where each tactic could act in a typical AI answer pipeline:

  Discovery: the platform learns the URL exists (search index, links, sitemaps)

  ↓

  Fetch: an index copy or a live request, often without JavaScript rendering

  ↓

  Passage retrieval: which parts of which pages match the prompt

  ↓

  Citation selection: which sources get named in the answer

  ↓

  Answer synthesis

  llms.txt could only act at discovery or fetch, and only if requested. Schema acts at fetch and understanding. Chunking acts at passage retrieval. None of them controls the gate before all of that: whether the page is in the index the platform actually queries. OpenAI documents separate crawlers for training (GPTBot) and ChatGPT search (OAI-SearchBot), and its ChatGPT search help page says ChatGPT search sometimes partners with third-party search providers. Bing reports Copilot citations and the grounding queries behind them in its AI Performance report, which points to Bing’s index as the gate there. That is why ordinary crawlability and ranking carry over to these platforms far more than any AI-specific file does.

  How to get your own data, and what would change my mind

  If you want to test this rather than take anyone’s word for it, run the cheap test first.

  1. Check your logs for llms.txt requests. Filter for requests to /llms.txt over 60 days and verify the user agents against each platform’s published crawler details. If no verified retrieval bot requested it, the file cannot be influencing that platform’s answers on your site.
  2. For schema or restructuring, use a matched test. Pick 30 to 50 pages on one template as the test group and a matched set on the same template as control. Change one variable only. Ahrefs noted that pages often shipped content and link changes alongside schema, which contaminates the result.
  3. Measure citations, not impressions alone. Run a fixed prompt set repeatedly on each platform, because answers vary between runs. Use Bing’s AI Performance report for Copilot citations (it has no click data), and GA4’s AI Assistant channel for referrals, remembering that visits without a referrer still land in Direct.
  4. Run it for 6 to 8 weeks and log confounders. Model changes, core updates and SERP feature changes all move citation patterns independently of your change.

  Evidence that would change my view	Evidence that would not
  Verified retrieval bots requesting llms.txt, then citing URLs that appear only in that file	A screenshot of one ChatGPT answer citing a site that has llms.txt
  A test-versus-control lift that repeats in a second time window and exceeds the control group’s normal week-to-week movement	A single-site before-and-after chart with no control group
  Platform documentation stating the file or markup is used in retrieval or citation	A tool that flags a “missing llms.txt” as an error

  Where I would spend the effort instead

  • Confirm the retrieval crawlers can reach you. Blocking GPTBot for training is a separate decision from blocking OAI-SearchBot for ChatGPT search, and CDN bot rules often block both by accident.
  • Server-render the content that matters. Most live AI fetchers in the SearchVIU test did not execute JavaScript.
  • Put the facts you want quoted (prices, specs, dates, locations, definitions) in visible HTML text, and keep any schema consistent with that text.
  • Write sections that answer one question each under a clear heading, with the answer near the top of the section, without splitting the page apart.
  • Keep llms.txt only if it costs you nothing or you publish developer documentation. Do not report it to stakeholders as an AI visibility win.

  My practical position: treat Google’s guide as correct for Google and as the best available default elsewhere, because nothing measured so far contradicts it. Spend five minutes on the log check for llms.txt, run one properly controlled test if schema or restructuring is a live debate on your team, and put the rest of the effort into crawl access, visible content and ranking in the indexes these platforms actually query.

  Need help with this on your own site?

  Niraj Raut works with businesses in Nepal, Australia, the UK, Europe and the US on technical SEO, ecommerce SEO, local SEO and AI search.

  See SEO services