What does machine-readable mean?
Machine-readable means a computer can read and process the data without a person interpreting it first and without losing its meaning. For a web page that means more than text in a browser: headings in real HTML, structured data such as JSON-LD, a sitemap, and an llms.txt index that an AI crawler or agent can parse directly.
Part of Glossary: the terms behind MCP, coding agents and shared pages, defined plainly.
A person reads a page by looking at it. A program reads the bytes, so the structure has to be in the bytes: real headings, lists and tables, not styled divs.
A format with a published schema, such as JSON, CSV or XML, lets a program parse the data without guessing. That is the W3C definition, and the one in US law.
On a web page, structured data (JSON-LD) states what the page is: an article, a defined term, a FAQ. Search engines and agents read it without parsing prose.
Site-level files point programs at the content: robots.txt says who may crawl, sitemap.xml lists every page, llms.txt summarises them in Markdown for agents.
The test is simple: could a script, with nobody watching, pull out the facts and keep their meaning? If not, the page is not machine-readable yet.
Machine-readable versus human-readable
The W3C's Data on the Web Best Practices, a Recommendation from 2017, defines machine-readable data as data in a standard format that a computing system can read and process automatically, and lists XML, JSON, HDF5, RDF and CSV as examples. US law goes one step further. The definition added to 44 U.S.C. 3502 by the OPEN Government Data Act says the format must be processable by a computer without human intervention "while ensuring no semantic meaning is lost".
That last clause is the detail people miss. A scanned receipt is human-readable only. A PDF of the same receipt is readable by a computer in the sense that the bytes are there, but the program has to guess which figure is the total. A CSV of the receipt is both: a person can open it, and a script gets every value with its label intact. Machine-readable is about keeping the meaning, not just the characters.
Example: one page, read three ways
Take a glossary page like this one. A person sees a heading, a definition and a diagram. A search crawler reads the title, the heading outline and a JSON-LD block that says, in schema.org terms, that this is a DefinedTerm with a FAQPage attached, so it never has to infer what kind of page it found. An AI agent asked to look something up may fetch the site's llms.txt first, a Markdown index with one line per page (title, link, summary), pick this page from the list, and only then load the HTML. intraface.cc publishes all three layers: semantic HTML, JSON-LD on every generated page, and an llms.txt plus a sitemap index.
The same content serves all three readers because the structure lives in the markup and the data files rather than in the styling.
Why it matters for AI search
The llms.txt proposal from Jeremy Howard makes the case plainly: HTML built for people carries navigation, ads and scripts that are hard to convert cleanly into text, and an agent's context window cannot hold a whole site, so every wasted token costs time and money. Agents want concise information in a predictable place. Google defines structured data as a standardized format for providing information about a page and classifying its content, and recommends JSON-LD because it is the easiest format for site owners to implement and maintain at scale.
Now the trade-off. Google's own guidance on AI Overviews says there are no additional requirements and no special schema.org markup needed; the same SEO fundamentals apply, and the one concrete instruction is to keep important content in text form. So machine-readable is not a trick that ranks you. It is what lets every other reader, the crawler that builds an AI index, the agent a developer runs from a terminal, parse your page correctly the first time.
Common mistakes
- Content that only exists after scripts run. A crawler or agent that does not execute JavaScript sees an empty shell. Render the text into the HTML.
- Text inside images or canvases. Pixels carry no meaning to a program. Put the words in the document and use alt text for the picture.
- JSON-LD that says something the page does not. Google says not to add structured data about information the page does not show. Generate both from the same source so they cannot drift.
- Treating llms.txt as a ranking file. It helps agents that fetch your site. It does nothing for Google, by Google's own account.
- Shipping a PDF when a table would do. If the data matters, publish it in a format with a schema.
FAQ
- Is a PDF machine-readable?
- Only in the weakest sense. A program can open it and pull out characters, but a table in a PDF loses which number belongs to which column, so the meaning goes. Under the US legal definition, which requires that no semantic meaning is lost, it does not count. The same table as CSV or JSON does.
- Is HTML machine-readable?
- It can be. HTML with real headings, lists and tables parses cleanly, and a JSON-LD block makes the page's type and facts explicit. A page whose text only appears after scripts run, or that puts its content in images, is readable by a browser and not by much else.
- Does a machine-readable site need llms.txt?
- Not for Google. Its documentation says no new machine-readable files or AI text files are needed for AI Overviews. llms.txt is for agents that fetch your site directly and have a small context window to spend. It is cheap to publish; expect no ranking effect from it.
Stop sending HTMLover Slack.
Create an account in a minute, connect your coding agent, and make your first board.