What can llms.txt realistically do for AI visibility, and what should it not replace?

Treat llms.txt as an optional machine-readable orientation layer, not as a substitute for crawlable, authoritative web content.

If you use it, keep it concise and accurate. Point to canonical documentation, important resources, product/service explanations, policies, and other durable pages that help a machine or developer understand the site. Maintain the file when URLs change.

Do not assume that every AI system discovers, honors, or gives ranking weight to the file. Different systems have different crawling, retrieval, licensing, and source-selection behavior. There is no safe basis for treating llms.txt as a universal inclusion directive.

The fundamentals remain more important: technically accessible pages, strong information architecture, clear entities, original evidence, factual consistency, reputable external corroboration, useful answer-oriented content, and stable canonical URLs.

Also avoid placing unique critical information only in llms.txt. Anything important enough to influence understanding should exist on a normal public page that users can access and that can be cited independently.

A good implementation is low risk when it mirrors the site's real canonical resources and is maintained like a documentation index. A bad implementation is a keyword list, hidden promotional claims, or a replacement for fixing crawl/index/entity problems.

Use it as an experiment and document whether the systems you care about actually fetch it. Keep expectations proportional to evidence.