Structured data vs llms.txt: a vocabulary engines read, a file agents might

What is the difference between structured data and llms.txt?

Where each lives and who reads it. Structured data is markup inside each page that describes what the page is. Google defines it as a standardised format for providing information about a page and classifying its content, recommends JSON-LD where a site can use it, and reads it to decide whether a page is eligible for rich results. The vocabulary is schema.org, which describes itself as shared vocabularies the major search engines understand and names Google, Microsoft, Yandex and Yahoo.

llms.txt is a single markdown file at your site root, or at a sub-path for the pages beneath it, proposed by Jeremy Howard in September 2024. It lists the pages an AI agent would find most useful, with links and short descriptions, and points to clean markdown versions of them. It describes no page in detail. It is a reading list.

So one is a description attached to each page, in a vocabulary with documented consumers. The other is a curated index for a reader whose use of it the answer engines have not documented.

What does each one do for AI answers?

For Google's AI features, neither is required. Google states that there is no special schema.org structured data you need to add to appear in AI Overviews or AI Mode, and that you don't need new machine-readable files, AI text files or markup. A page needs to be indexed and eligible to be shown with a snippet. Structured data still helps Google understand the page and can make it eligible for rich results in ordinary search, which is reason enough to have it.

For engines that fetch pages themselves, there is no published statement from OpenAI or Perplexity that either structured data or llms.txt changes which pages they cite. What they document is robots.txt for their crawlers. Google's John Mueller has gone further on llms.txt, calling it purely speculative for now and noting that the file has existed for years without the AI systems using it. Anything beyond that is inference from practice.

What do AI answers themselves say about the two?

They mention structured data far more often, and they name specific types. We searched the newest stored answer to each of our 61 tracked prompts, one per engine, asked between 2 and 4 October 2026: 121 answers in all. Twenty-four mentioned structured data, schema markup or JSON-LD. Six mentioned llms.txt. Most of the schema mentions sat in lists of practical advice: add Organization, Product and FAQ markup, keep business details consistent, use clear headings. The llms.txt mentions clustered on prompts about developer documentation, MCP servers and tools that make site changes, which is where the file was designed to matter. One of the six, the ChatGPT answer above, told the reader not to over-invest in it.

Two cautions. These answers were assembled from third-party guides written by vendors and agencies, not from any engine's documentation, so they reflect what the industry writes rather than what the engine does when it chooses sources. And a mention is not a measurement of effect. Still, if the question is which of the two the engines' own answers steer people towards, the readings are clear.

What should you spend your time on?

Pages first, structured data second, llms.txt third if it is free. A page that answers a buyer's question directly, near the top, with its numbers and their sources inside the passage, does more for AI answers than any markup. Then add structured data generated from the page's own data: Organization on the about page, Product or SoftwareApplication on product pages, Article on articles, FAQPage where the page really shows questions and answers, BreadcrumbList everywhere. Google requires every required property of a type before the markup makes a page eligible for an enhancement, so partial markup can be worth nothing.

If your platform generates llms.txt automatically, publish it. If it would mean hand-maintaining a list of URLs, skip it until someone documents a reason. Documentation platforms increasingly generate the file, and markdown versions of each page, without being asked, which removes the cost argument for docs sites but not the evidence question for everyone else.

Why does structured data still matter when the reader is a model?

Because a model that fetches a page reads its HTML, and JSON-LD is part of it. A product's price in markup, an FAQ's questions and answers in markup, an article's author and date in markup: these are unambiguous statements a model can lift without parsing layout. Nobody has published proof that this raises citation rates, and we will not claim it does. But it costs little when generated from page data, it helps Google, and it removes one source of ambiguity for any reader.

Do agents that read documentation behave differently?

Probably, and this is where llms.txt has its strongest case. Coding assistants and agents that work with a product's documentation often fetch documentation pages directly and benefit from a clean index and markdown versions, which is the use the proposal was written for. Its site says AI labs publish llms.txt files for their own developer docs and that Chrome's Lighthouse checks for one. Our readings agree: half of the llms.txt mentions came from prompts about developer docs. When we asked ChatGPT itself how to get docs cited, its 12 answers over 30 days drew all 337 of their citations from OpenAI's own domains, mostly the help centre: the engine answers that question from its vendor's documentation, not from the guides that sell either format. If your product has developer documentation, an llms.txt for the docs is a reasonable addition. If your site is marketing pages, the case is weaker.

How can you tell whether either one made a difference?

By reading answers before and after, not by checking the markup. Pick the prompts the page should win, record how often each engine names you and cites the page over a few weeks, add or fix the markup, and keep reading. A green tick on a validator says the markup is well formed; it says nothing about citations. A difference that does not clear the noise is no difference, and on a small prompt set that usually means waiting longer than anyone wants to.

If someone tells you markup or a file raised their AI citations, ask how many answers the before and after figures rest on, which engines they read, and what else changed on the page in the same weeks.