← The journalTECHNICAL / FIELD GUIDE

llms.txt, Markdown, and schema: choose the right job for each

Evaluate machine-readable formats without mistaking an implementation choice for a guaranteed visibility signal.

Museum architecture with directional camera-motion blur.TECHNICAL / SUNDIAL
Photography: Harrison Lin / Unsplash ↗ · Sundial motion treatment
THE PRACTICAL TAKEAWAY

Improve access and clarity first; test optional formats against a stated purpose.

Define the problem you are solving

A format is useful when it makes a task easier. Structured data can express certain facts for supported consumers. Markdown can offer a simpler representation of documentation. An llms.txt file can organize links for systems that choose to use it. None of those purposes, by itself, establishes a universal AI ranking advantage.

Write down the intended consumer, the information it needs, and the failure you observe today. If your content is inaccessible or contradictory, adding another file may duplicate the problem rather than solve it.

Follow documented requirements

Use the relevant provider’s current documentation for required formats and controls. Google says new AI text files and special schema are not requirements for its AI search features. Treat that as a specific statement about those features, not proof about every other assistant.

Reference: Google’s documented AI search requirements ↗

Maintain one source of truth

If you provide multiple representations, keep them synchronized with the visible page. A Markdown version with old pricing or a structured record with obsolete availability can create ambiguity. Assign ownership and generate secondary representations from the authoritative source when practical.

Do not put material facts only in image metadata or a hidden machine-oriented block. A buyer should be able to inspect the information and its qualifications. Keep markup truthful and consistent with what the page actually offers.

Test the consumer, not the label

Verify that the intended system can access and use the representation under its documented behavior. Observe the actual returned content and any downstream answer. A successful file request does not establish that the file influenced a recommendation.

Record the implementation cost and maintenance burden alongside the outcome. If the format improves a documentation workflow, that may justify it even without a visibility change. If the only benefit is a hypothesis, keep it labeled as a test instead of turning it into a requirement for every page.