An `llms.txt` file is a proposed Markdown-based way to declare which pages a site considers most useful to language-model systems. Its practical value is not established as a direct visibility lever. For SeoVision users, the more consequential question is whether the pages named in the file are the pages AI assistants actually cite, and whether those citations lead to brand mentions and accurate answers.

What is an llms.txt file?

An `llms.txt` file is a plain-text document published at a website’s root, commonly at `https://example.com/llms.txt`. It can describe the organization and point to selected public resources. Unlike a sitemap, it is editorial rather than exhaustive: the publisher decides which pages deserve attention.

A useful file is less a second navigation menu than a testable claim about authority. If a company lists its API reference, security documentation, original research, and product explanations, those pages should be the ones its subject-matter experts would defend as current and authoritative.

A simplified example:

```markdown

Example SaaS

Example SaaS helps software teams monitor application performance.

Product documentation

• Getting started: Setup instructions.
• API reference: API endpoints and authentication.

Guides

• Monitoring best practices: Practical implementation guidance.

```

The proposal is maintained at llms-txt.org, and its syntax and adoption may change. The terms `llm txt`, `llm text`, `llms-txt`, and `llms txt file` generally refer to the same proposed convention, but search demand for those variants is not evidence that an AI engine supports the file.

How is llms.txt different from robots.txt and a sitemap?

These files should not be treated as interchangeable diagnostics. `robots.txt` expresses crawler rules, an XML sitemap supplies URL discovery signals, and `llms.txt` curates a readable set of resources. None of those functions proves that a page will be retrieved, cited, or named in an AI answer.

Use the file as a content-prioritization layer, not as a replacement for crawl controls, canonical URLs, internal links, structured content, or an XML sitemap. If you need to understand access rules for Google’s crawler, see this guide to testing Googlebot access. More importantly, compare the file’s selected URLs with the sources returned for your target prompts. That comparison produces evidence; merely checking that `/llms.txt` exists does not.

Is LLMs.txt actually used?

No universal evidence shows that major AI assistants consistently fetch, parse, or prioritize `llms.txt`. Retrieval systems are only partly observable, and an answer can be assembled from search results, an internal index, retrieved page content, training data, feeds, or other systems. The presence of a file therefore cannot establish that a model read it, and a later citation cannot establish that the file caused the citation.

That limitation changes how the file should be evaluated. SeoVision’s AI-visibility scan corpus recorded 15,527 completed AI-assistant answers in a 28-day window across all tracked engines. Of those answers, 83% cited at least one source and 31% named the tracked brand. These figures are not an llms.txt adoption study. They show why visibility work should measure the answer and its sources directly rather than infer success from a technical implementation.

Is llms.txt worth it?

It is worth testing when the file can be maintained as part of an existing documentation or content workflow. The strongest candidates are documentation-heavy SaaS sites, developer portals, knowledge bases, and publishers with a small set of stable, authoritative resources. The weakest candidates are sites that would dump thousands of URLs into the file or cannot keep the listed pages accurate.

Treat publication as a low-cost experiment with an editorial side benefit. The file can:

1. Force agreement about which pages represent the organization accurately.
2. Expose outdated, duplicated, or weak documentation before those pages are presented as authoritative.
3. Create a defined URL set for comparing AI citations before and after publication.
4. Reveal a gap between internal priorities and the pages answer engines actually surface.

Do not use it to compensate for vague pages, unsupported claims, inaccessible content, or inconsistent company terminology. SeoVision’s audited-site data reinforces the broader point: across 1,489 audited websites, the median AI readiness score was 83/100, compared with a median SEO score of 76/100, as of September 4, 2026. Those medians do not measure `llms.txt`; they indicate that readiness is a broader condition than adding one file.

What should an llms.txt file include?

Select pages according to the questions an assistant must answer correctly, not according to keyword volume. For each URL, write a description that identifies its subject and usefulness. A page that says “learn more about our solution” gives an engine less editorial direction than one that says “documents API authentication, rate limits, and error responses.”

Prioritize:

• Product documentation and implementation guides
• Original research, methodology, or technical references
• Definitive explanations of products and services
• Pricing, integration, security, or compliance pages when relevant
• High-quality use-case and comparison pages
• Stable URLs that return successful responses without unnecessary redirects

Avoid:

• Every URL on the site
• Login pages, account areas, and private resources
• Tag archives, duplicate variations, and thin pages
• Unsupported marketing claims
• Links that are blocked, outdated, or redirected repeatedly
• Content that conflicts with the visible page or current product offering

Use the file to make an editorial choice, then test that choice. If the listed page is never cited for relevant prompts, inspect the page itself, competing sources, query intent, and engine differences before assuming the file failed.

What is llms-full.txt?

`llms-full.txt` generally describes a larger text resource containing expanded page content, while `llms.txt` is a concise guide or index. Neither should be assumed to have universal support across AI answer engines.

A combined text resource also increases the cost of stale information, duplicated content, licensing mistakes, and conflicting versions of a page. Before publishing one, confirm that the material is public, current, and consistent with canonical pages. For most organizations, a short, curated file is easier to audit and gives a cleaner test.

How do you create an llms.txt file?

Automation can collect candidate URLs, but it should not make the authority decision. Use this workflow:

1. Define the audience and test

Choose the questions the file is intended to support: implementation, integrations, API use, security, product evaluation, or another defined area. Record a fixed prompt set before publishing so later comparisons are not based on memory or cherry-picked answers.

2. Select canonical, public URLs

Export candidates from the sitemap, content inventory, or documentation system. Remove private pages, duplicates, redirects, and pages that do not answer a meaningful question. Confirm that each URL is the current source rather than an older announcement or copied summary.

3. Write a precise overview

Describe what the organization does, who it serves, and which resources are authoritative. Keep every claim aligned with the visible pages. Do not turn the overview into a slogan that an answer engine cannot verify.

4. Group and annotate links

Organize resources by user task or subject. Add one specific description per link. The description should explain the page’s scope, limitations, or evidence—not repeat promotional language.

5. Publish at the expected location

Place the file at the site root as `/llms.txt`. Confirm that the response is publicly accessible, contains the intended Markdown, and is not replaced by an HTML template, authentication wall, or unexpected redirect.

6. Validate and maintain it

Check links and page status during documentation releases, product changes, migrations, and rebrands. A file that accurately described the site six months ago may now direct an assistant to obsolete instructions.

7. Measure outcomes separately

Track a stable prompt set across relevant engines. Record brand mentions, citations, cited URLs, competitors, and material changes in answer accuracy. SeoVision provides AI visibility tracking across ChatGPT, Claude, Gemini, Perplexity, Grok, DeepSeek, Copilot, Google AI Overview, and Google AI Mode, allowing the file to be compared with observed outputs rather than treated as proof of retrieval.

Can llms.txt improve AI visibility?

It may help a compatible retrieval system discover or interpret selected content, but the effect is unproven and difficult to isolate. AI visibility also reflects page quality, authority, crawl access, entity clarity, query intent, freshness, competition, and each engine’s retrieval process.

A defensible test has a recorded publication date, a fixed prompt set, repeated runs, and a log of other changes. Compare the same or closely comparable prompts before and after publication. Look for sustained changes in citations and brand mentions across more than one run. A short-lived answer change is not enough, especially when engines update their indexes or responses independently.

What the data does not prove

SeoVision’s figures do not prove that `llms.txt` causes mentions, citations, higher rankings, or better answers. The AI-visibility corpus measures completed answers and observed outputs; it does not reveal every retrieval step or establish whether an assistant read a particular file.

The website audit corpus is a snapshot of the sites audited, not a random sample of every website. Median scores can vary by site type, industry, implementation quality, and audit coverage. The 28-day scan window captures behavior observed during that period, not permanent engine behavior. Any claim of sustained improvement would require a controlled before-and-after test with stable prompts, repeated runs, and documented content and technical changes.

Is llms.txt mandatory?

No. It is not mandatory for SEO, AI crawler access, or visibility in AI answer engines. A site can operate without it, and publishing it does not guarantee that a model will read or use it.

Prioritize crawlable pages, valid technical foundations, accurate content, clear organization, and monitoring first. If those foundations are in place, a small, accurate file is a reasonable experiment—not a compliance requirement or a shortcut to AI visibility.

What to do next

1. Audit your foundations. Check indexability, canonical URLs, redirects, status codes, internal links, and AI crawler access before creating a new file.
2. Choose high-value resource groups. Start with the pages that best answer your audience’s important product, expertise, documentation, and customer-outcome questions.
3. Remove weak candidates. Exclude private, duplicate, outdated, thin, and unsupported pages.
4. Draft `/llms.txt`. Create it manually or with an llms txt generator, then verify every title, description, and URL.
5. Test the response. Confirm that the root URL is public, returns the intended Markdown, and does not redirect unexpectedly.
6. Record a baseline. Run a fixed prompt set across relevant AI answer engines and note mentions, citations, competitors, and answer accuracy.
7. Monitor for a sustained result. Recheck comparable prompts over time and document other site changes.
8. Improve the linked pages. Strengthen direct answers, evidence, authorship, terminology, and internal linking where the baseline reveals gaps.
9. Review during releases. Remove broken links and replace pages that are no longer authoritative.

For broader context, compare this work with GEO and AEO differences, AI visibility tools, and where ChatGPT gets information.

How we measured

The percentages and answer counts in this article come from SeoVision’s AI-visibility scan corpus of daily prompt runs across all tracked engines, measured over a 28-day window as of September 4, 2026. The AI readiness figure comes from SeoVision’s audit corpus of real websites, covering 1,489 audited sites as of the same date. These corpora measure observed answers and audit conditions, not `llms.txt` adoption or causal impact.