Should a SaaS team spend time creating and maintaining another machine-readable file? The short answer to “does llms.txt matter” is that no broad increase in AI-search citations has been established. A file may have a narrower use if a known documentation or coding agent reads it in a user-directed workflow, but that is different from improving unsolicited search visibility. Treat it as an optional experiment, not a ranking shortcut, and prioritize accessible, accurate public documentation first.
What is llms.txt?
llms.txt is a proposed, site-hosted text file intended to point language-model tools toward selected public resources. Publishing the file does not show that a search system fetched it, used its contents, or cited the site. It is distinct from robots.txt, which communicates crawler access preferences; neither file should be treated as a guaranteed citation lever. A comparison of llms.txt vs robots.txt should focus on their different purposes, especially actual access controls versus a proposed resource index.
File or content | Purpose | What you can verify | What it does not prove |
|---|---|---|---|
| Propose a concise index of selected resources for tools that choose to use it | Whether the file is available and whether a request for it appears in server logs | That a particular tool supports it, processed the contents, or cited a page |
| Communicate crawler access instructions | The published directives and, where available, crawler requests | That every crawler follows those directives or that a page will be cited |
Indexed HTML and documentation | Make public content available through ordinary web pages and search eligibility processes | Indexing and search visibility signals available in relevant tools | That an AI answer will retrieve or cite a page |
Implementation details should be based on observed behavior from the particular consumer being tested. The existence of a file is not evidence that a provider has adopted a common convention.
Does llms.txt increase AI-search citations?
No reliable, general citation uplift is established. SE Ranking’s domain-level analysis examined nearly 300,000 domains and found llms.txt on 10.13%; its statistical analyses and machine-learning model found no measurable relationship between file presence and AI citation frequency. This vendor study is a snapshot, so it does not prove that the file can never help a specific workflow. But it does not support treating it as a general visibility tactic.
Source and date | Observation | Design | Valid inference | Invalid inference |
|---|---|---|---|---|
SE Ranking, November 7, 2025 | Nearly 300,000 domains examined; 10.13% had llms.txt; no measurable relationship with AI citation frequency was found | Cross-sectional vendor analysis using statistical analyses and a machine-learning model | The study did not detect a measurable association in its analysis | The file can never help any site or workflow, or the study proves a particular causal mechanism |
Contentful’s evidence review, June 3, 2026 | Evidence is described as emerging; the review calls for reproducible independent testing and cautions about before-and-after observations | Editorial analysis, not a controlled causal result | Early logs and case examples need careful interpretation because platform changes and concurrent site work can confound them | A reported change after publication establishes that llms.txt caused a citation increase |
The distinction matters for a marketing team. A file can appear on a site while citation frequency changes for unrelated reasons, including platform behavior or other site changes. The absence of a measured broad relationship also does not rule out a limited use for a particular agent. Keep both points in view: neither an observed association nor the absence of one establishes a universal ranking factor.
Is llms.txt required for Google AI features?
No. Google’s Search AI-feature guidance says pages must be indexed and eligible for a Search snippet for its AI features, and it describes no additional technical requirements for those features. The guidance does not identify llms.txt as a requirement. This is specific to Google Search AI features and should not be generalized to every model vendor or tool.
When might an llms.txt file still be useful?
A concise index might help in a user-directed documentation or coding workflow if the agent is known to request and use it. That is a possible workflow utility, not evidence of an increase in general AI-search citations, as distinguished in Contentful’s review of the proposed file and its evidence. Verify the behavior of the individual consumer before investing in a format or naming a provider as a supported user.
For example, a B2B SaaS company could propose listing its public API documentation and canonical integration overview for a known agent task. That is an illustrative use case, not a measured success. The list should not expose private tenant documentation or promote stale pricing. The team would need an owner for the file and a process to remove or update entries when the underlying public content changes. The aim is to help with a defined task, not to create an unmaintained second copy of the documentation.
If the underlying pages are hard to reach or render, a text index may not address the problem. Review the structure of your public product documentation and determine whether the relevant content can be accessed through the actual workflow being tested. Whether AI crawlers can render JavaScript is a separate accessibility question.
Should your SaaS team implement it now?
If core pages are inaccessible, inaccurate, or not indexed, fix those issues first. If the basics are sound and a known agent consumes the file, a low-cost, public-only experiment may be reasonable, provided someone owns its accuracy and maintenance. It is not a required generative engine optimization (GEO) task.
Condition | Recommended priority | Evidence to collect | Maintenance owner |
|---|---|---|---|
Important public pages are inaccessible, inaccurate, or not indexed | Fix the underlying access, accuracy, or indexing issue first | Relevant page status, content accuracy, and indexing information | Documentation or web owner |
No known agent or workflow is expected to use the file | Defer implementation; do not add work on the assumption of citation gains | A clearly defined prospective workflow or observed consumer behavior | No owner needed until a test is approved |
A known agent may use a concise index and upkeep is inexpensive | Run a bounded, public-only experiment | Requests for the file, agent task outputs, and maintenance effort | Named documentation or web owner |
The file would duplicate changing content without a review process | Do not publish until ownership and update rules are set | Canonical destination list, change triggers, and broken-link checks | Named maintainer |
Keep the file limited to public, accurate destinations. Use canonical pages, define who updates the list when documentation changes, and review for broken links. A link to a page is not a substitute for making that page accessible or maintaining it.
How can you test whether llms.txt helps?
Separate two outcomes: verified use in a known agent workflow and changes in citations or recommendations across repeated AI-search prompts. Server logs may show a request for the file, but they cannot establish that the returned content influenced an answer. Before-and-after changes without a comparison are also inconclusive when platforms and other site work change at the same time.
Define agent-task success and AI-search citation outcomes separately. Decide what counts as a successful agent task and what counts as a citation or recommendation.
Record the baseline prompt, platform, model, observation date, and page group. Keep prompt wording and observation conditions consistent where feasible.
Publish a file that lists the selected test pages, while leaving similar comparison pages out of the index. Both groups may still be accessible through the same site-hosted file or other retrieval paths, so account for that contamination risk. This design reduces some ambiguity but does not guarantee causal identification.
Verify genuine requests for the file in server logs and examine the relevant agent outputs. A logged fetch shows a request, not that the content shaped the result.
Repeat prompt observations and compare citation or recommendation rates and qualified visits across the groups. Treat these outcomes separately from agent-task performance.
Document platform changes, concurrent site work, and other plausible confounders. Stop if maintenance exceeds observed utility.
The test should answer a defined operational question, such as whether a known agent can locate a canonical public integration guide more reliably when given the index. Do not present that result as proof of improved search citations. For a visibility test, predefine the prompts and observation window, record each platform and date, and report uncertainty rather than treating a small number of changing answers as a stable trend. A measurement plan should focus on AI citation rate measurement as an outcome, not treat implementation itself as success.
Keep the measurement record interpretable. Save the exact prompts, note when an answer cites the company or a specific page, and distinguish a citation from a mention or an unlinked recommendation. Record the denominator as well as the number of observed citations so repeated samples can be compared. These observations describe the tested prompts and conditions; they should not be generalized to all buyers, queries, or AI systems. Track qualified visits separately because a change in citations does not automatically imply a change in commercially relevant demand.
Can server logs prove that an AI answer used the file?
No. A request in server logs proves that a fetch was recorded, not how the returned content was processed or whether it caused a citation. Pair logs with controlled task outputs and repeated answer observations, while acknowledging that those observations still may not reveal what influenced a system’s response.
Conclusion
llms.txt is optional and is not a proven AI-citation lever. Prioritize crawlable, accurate public documentation, then test any known agent use separately from search visibility. If you run a test, define the outcome, record fetches and answer observations, and account for concurrent changes. For the distinct access-control question, see llms.txt vs robots.txt.
Frequently Asked Questions
Is llms.txt effective?
Available research does not establish that llms.txt broadly increases AI-search citations. It may still be worth testing for a specific agent workflow if that agent is known to request the file and the team can maintain an accurate, public-only index.
How Nortav Can Help
For SaaS teams deciding whether to maintain llms.txt, Nortav can help prioritize accessible, accurate documentation and define a test that separates agent workflow outcomes from AI-search visibility.
Our approach starts with understanding your current search visibility, identifying gaps against competitors, and developing a strategy focused on driving qualified traffic, generating leads, and delivering measurable business value.
Book a Discovery Call →
Sources
Google Search Central, “AI features and your website.”
Contentful, “Do llms.txt files actually improve AI search visibility?” June 3, 2026.
SE Ranking, “LLMs.txt: Why Brands Rely On It and Why It Doesn’t Work.”
