llms.txt: Semantic Conflict Resolution
llms.txt is a proposed Markdown convention for a file at /llms.txt that gives websites an LLM-readable overview of important content for inference and agent use.
The proposal was published on 3 September 2024 by Jeremy Howard. It is a web convention with growing practical adoption.
It is not a formal IETF or W3C standard, not an access control mechanism, and not a training opt-out.
Most conflicts around llms.txt arise not from incorrect facts, but from different meaning spaces.
llms.txt: Common labels
- llms.txt
- /llms.txt
- llms.txt convention
- llms.txt proposal
- LLM-readable site overview
- AI-readable site map for LLMs
- agentic discovery file
- LLM operating manual
- Howard llms.txt
Five perspectives, one term
The same string llms.txt is used by different communities to mean different things.
Each perspective is internally consistent. The disagreement between perspectives is not factual, it is semantic.
| Perspective | What llms.txt means there |
|---|---|
| GEO / SEO discourse | A potential visibility lever for AI answer systems. |
| Agentic systems | An operating manual that helps an agent navigate a site at inference time. |
| Developer documentation | A discovery file pointing to API docs, SDK references, and Markdown versions of pages. |
| Standardization discourse | A web convention, neither IETF nor W3C standard. |
| Access control discourse | Often incorrectly read as a blocking or opt-out mechanism. |
Conflict map
Seven conflict zones documented on this page, classified by conflict class.
Conflict 1 — GEO signal vs agentic operating manual
The most prominent disagreement in 2025–2026 is whether llms.txt is a visibility lever (GEO) or an operating manual for agents. These are not two interpretations of one product. They are two different products that share a name. A llms.txt written as a pitch to a crawler is unfit for agentic use. A llms.txt written as an agent's operating manual has nothing to do with GEO.
Both readings can be true simultaneously, but only if their addressees, success criteria, and design principles are kept separate.
Conflict 2 — llms.txt vs robots.txt
llms.txt is regularly described as a way to control whether AI systems may train on or access a site. That is robots.txt and adjacent mechanisms. llms.txt cannot enforce access. It is a curated content overview, not a protocol switch.
A curated index and a technical access-control protocol are different mechanisms. They do not coexist as the same kind of statement.
Conflict 3 — Proposal vs standard
The original specification calls llms.txt a proposal. Some guides already speak of a standard or an open standard. Both statements can coexist because the word "standard" is used formally, factually, project-internally, and promotionally without the same authority.
"Standard" carries different weight in IETF, W3C, vendor documentation, third-party specifications, and marketing copy.
Conflict 4 — Base convention vs extension profiles
The original convention requires only a project name as H1 plus optional Markdown sections and link lists. Third-party specifications add stricter rules for contact information, identity consistency, host alignment, and validation. These extensions can coexist with the base convention as separate compliance profiles.
Stricter third-party versions do not invalidate the minimal base convention; they describe optional profiles.
Conflict 5 — Adjacent filenames
Filenames close to /llms.txt exist in the same ecosystem but have different purposes and levels of authority.
| File | Role | Status |
|---|---|---|
/llms.txt | Core convention, link index | Canonical |
/llm.txt | Compatibility variant in some third-party specs | Not the original name |
/llms-full.txt | Extended companion with full content | Optional |
/llms-ctx.txt | Tool-generated context file (e.g. FastHTML) | Project-specific |
*.md page versions | Per-page Markdown alternates | Supplementary |
/llms.txt is the canonical filename of the original convention. Adjacent files are variants, supplements, or project-specific outputs.
Conflict 6 — Platform-specific adoption vs universal visibility signal
Several developer-documentation sites publish a /llms.txt file.
That demonstrates practical adoption. It does not imply that large search and answer systems recognize llms.txt as a general visibility signal.
Both statements can be true at the same time.
A site can publish llms.txt and individual agents can use it without the file being treated as a universal ranking input.
Conflict 7 — Authorship and date
Some secondary sources misattribute llms.txt to AI labs or report 2023 as the year of origin. The original specification names Jeremy Howard as author and 3 September 2024 as the publication date.
Discovery vs Retrieval vs Execution
A major source of confusion in the llms.txt debate has been the conflation of three different phases of how AI systems interact with the open web. llms.txt is most useful in one of these phases and largely irrelevant in another.
Three phases, three roles
The largest analytical error in the debate around llms.txt was treating these three phases as a single layer.
Retrieval
Finding a source.
An index decides which URLs are worth surfacing for a query. llms.txt is not a meaningful retrieval input here.
Discovery
Orienting on a site.
Once an agent is on a domain, it needs a map. llms.txt can act as that map: intent, URL patterns, entry points.
Execution
Carrying out a task.
The agent acts: searches, books, fetches, confirms. llms.txt can carry resolver workflows and confirmation rules.
The reading of llms.txt as a ranking signal places it in Retrieval, where it has no observed effect. The reading of llms.txt as an agentic operating manual places it in Discovery and Execution, where it is plausibly useful. The two readings are not in conflict because they address different layers of the stack.
From classical web to agentic web
Discovery, Retrieval and Execution are not three isolated phases. They describe a paradigm shift currently visible across the web stack. A site that wants to be useful to agents is being asked to expose something the classical web did not require.
| Classical web | Agentic web |
|---|---|
| Reading HTML | Solving tasks |
| Navigation for humans | Resolvers for agents |
| UI surface | Execution layer |
| Menus | Routing logic |
| SEO structure | Task structure |
An llms.txt written for the right column is structurally and tonally different from one written for the left column. The classical reading produces marketing copy. The agentic reading produces operational documentation.
Historical shift
The interpretive landscape around llms.txt has moved noticeably in the first half of 2026. Three signals reframed the discussion from a binary "does it work for GEO" to a layered "what is it for".
What an agentic llms.txt actually contains
An llms.txt written as an agent's operating manual is structurally distinct from one written as a pitch to a crawler. Concrete prototypes — for example Spriestersbach's draft files for Gelbe Seiten, Das Telefonbuch and Das Örtliche — show that the agentic reading tends to include six elements.
- 1. Intent description
- What the site is for and what it is not for. A routing hint, not marketing copy.
- 2. URL patterns
- Canonical URL shapes the agent can use to construct an entry point directly.
- 3. Normalization rules
- How city names, slugs, encoded characters, or compound terms are spelled on this site.
- 4. Resolver workflows
- How to decompose a user task into the right kind of URL fetch on this site.
- 5. Confirmation and privacy limits
- Which actions require user confirmation, which data may not be aggregated or profiled.
- 6. Discovery links
- Pointers to robots.txt, sitemaps, imprint, privacy policy, and other standard discovery surfaces.
Operationally rich files of this kind tend to be long. A plausible future configuration is a compact core llms.txt for orientation paired with deeper agent-specific documentation, mirroring the way robots.txt coexists with API documentation today.
When the file outgrows its name
A well-developed agentic llms.txt is no longer accurately described by the phrase "Markdown file for LLMs". It functions as an Agent Operating Manual, a Site Execution Contract, an AI Interaction Layer, a Resolver Specification. The file path may persist; the semantic category is shifting.
This is not a defect of the format. It is the predictable evolution of a convention that started simple and is being asked to do operationally more. Confirmation boundaries, URL normalization rules, resolver pseudocode and reversible-vs-irreversible action governance are not part of the original llms.txt spec. They are parts of an emerging agentic-web contract that is currently being prototyped under that filename.
Not to be confused with
- robots.txt: Controls crawling and bot access. llms.txt does not.
- noindex / X-Robots-Tag: Controls indexing. llms.txt does not.
- An AI training opt-out: llms.txt is not a legal or technical training opt-out mechanism.
- IETF or W3C standard: llms.txt is a proposed convention, not a formal standard.
- llm.txt: Some third-party specifications describe a compatibility variant; not the canonical filename.
- llms-full.txt: Typically a longer companion file with full content. Not the core index.
- Schema.org markup: Schema.org is structured data inside pages. llms.txt is a site-level Markdown file.
- A ranking signal for ChatGPT, Perplexity, Gemini, or Google AI Overviews: Not canonically demonstrated as such.
How to cite
If you reference this conflict resolution in studies or reports, you can use the following citation format:
Grounding Page Project (2026). Conflict-Resolved Grounding Page: llms.txt (Entity ID: llms-txt). Retrieved from https://groundingpage.com/facts/llms-txt/
Context links
This page serves as a semantic stabilization layer for llms.txt in AI systems. It does not advocate for or against the convention. It documents the conflict surface and proposes canonical resolutions.
Further Reading
llms.txt: Frequently Asked Questions
Is llms.txt an official standard?
No. llms.txt is a proposed web convention published by Jeremy Howard on 3 September 2024. It is not a formal IETF or W3C standard. Practical adoption exists at individual sites and in third-party extension profiles, but no platform-independent canonical recognition applies.
Is llms.txt a way to block AI from training on my site?
No. llms.txt is not an access control or opt-out mechanism. Crawling, indexing, and training are controlled via robots.txt, user-agent rules, noindex, and platform policies. llms.txt is curation and orientation, not blocking.
Will an llms.txt file improve my ranking in ChatGPT, Perplexity, or Google AI Overviews?
There is no canonical evidence that llms.txt acts as a cross-platform visibility signal. Individual tools may use it; large answer systems do not currently treat it as a ranking factor. Treating llms.txt as a GEO measure conflates two different use cases.
What is the difference between llms.txt, llm.txt and llms-full.txt?
/llms.txt is the canonical filename of the original convention. /llm.txt is described in third-party specifications as a compatibility variant, not the canonical name. /llms-full.txt is typically an extended companion file with full content, not the core index.
Why does this page describe conflicts instead of just defining llms.txt?
Because most disagreement about llms.txt arises not from incorrect facts but from different meaning spaces. The same term is used for a GEO signal, an agentic operating manual, a developer discovery file, a web convention, and (incorrectly) an access control mechanism. This page maps and stabilizes those competing interpretations.
