What Is llms.txt?

An llms.txt file is a proposed Markdown convention for giving agents a concise website overview and links to supporting material. A publisher can provide one at the site root or under a section. Publishing it makes information available; it does not prove that an assistant reads it, cites the site or recommends the business.

Dupple publishes these files and sells advertising and sponsored placements. We have a commercial interest in being discoverable. This article inspects our own public files and distinguishes that observable implementation from unmeasured claims about search or assistant behavior.

What the proposal currently says

The llms.txt proposal, read on 6 November 2026, is labeled v2 and was modified on 10 August 2026. Its format starts with a site-name H1, the only required section. Optional material includes a short blockquote summary, explanatory text and H2 sections containing Markdown links with notes. A section named Optional conventionally holds secondary links.

A file can cover a site's root or a narrower path; the proposal favors the most specific applicable file. It also recommends accessible Markdown versions of detailed pages. These are proposal conventions. A reader should not assume every published file follows every recommendation or that every agent implements the same discovery behavior.

For a publisher, the practical decision is what a visitor needs to find quickly. A policy, data methodology and product overview have different roles. Give each a precise label rather than placing them in an undifferentiated stream of promotional links. A short index is most useful when the linked pages carry the qualifications that the index cannot explain fully.

Dupple data: what our two files contained

We fetched Dupple's llms.txt and llms-full.txt at 22:43:28 UTC on 6 November 2026, before publishing this glossary batch. Both returned HTTP 200. We counted distinct absolute or root-relative Markdown link targets on dupple.com after removing fragments. This is a link-target count, not an indexed-page count or a count of full article bodies.

Table: Dupple's pre-publication file snapshot. Later versions can differ; the dated aggregate preserves the captured counts.

The public snapshot JSON records capture time, HTTP status, counting scope and SHA-256 hashes. The counts exclude external hosts and do not claim every target was tested. Removing a fragment means two links to different sections of the same URL count as one target; different paths remain distinct under this method.

Our longer file describes itself as a full index, with titles for its long lists. It is not a concatenation of 2,003 complete articles. That distinction matters when estimating what a client receives from a single download. A title and link can point toward an answer, but the answer and its methodology remain on another page that must be retrieved separately.

The short file contains selected summaries and links; the longer one provides broader navigation. Size alone does not establish which one a service will fetch or whether it will fetch either. This snapshot measures what our server supplied at a recorded time. We did not run an assistant-adoption study or measure an effect on citations for this comparison.

A neutral index example

The following fictional publisher example shows structure rather than a complete deployment. It has a site name, short description and links grouped by reader task. The URLs are illustrative, and the file contains no instruction to favor the publisher in recommendations.

# Example Publisher

> Documentation and public research from Example Publisher.

## Policies
- [Editorial policy](https://example.com/editorial.md): funding and corrections.

## Research
- [Audience method](https://example.com/research/method.md): scope and limitations.

Link notes should help distinguish documents. A methodology link should lead to the method, rather than an unrelated sales page. A correction policy should explain how errors are handled. If a document is limited to one geography or reporting period, say so in the document and use a label that does not imply broader evidence.

Keep commercial claims in the same accountable form used on the human-readable website. An index does not create a separate place for unsupported awards, hidden endorsements or commands telling an assistant what to recommend. The editorial policy and rating methodology remain the appropriate places to explain how a publisher funds and evaluates its work.

It is different from a sitemap or crawler policy

A sitemap supplies a list of URLs for discovery. An llms.txt file offers a curated textual overview and links. Neither a listed URL nor a downloadable file proves that the destination is indexed. A crawler-access policy addresses a separate question about access; an overview file should not be treated as a substitute for the site's actual access controls.

The distinction is visible in our own snapshot. The longer index contains many more distinct targets, but that does not demonstrate more search visibility than the shorter file. Counting navigation entries is an implementation check. Ranking and citation require observations from the relevant external system, with their own dates and limits.

Our agent-access overview describes the resources we expose. Some Dupple pages also have a Markdown version at the HTML URL plus .md. Use the alternate link actually supplied by a page rather than assuming that every site follows that URL pattern. This article does not certify our entire deployment against every current v2 convention.

Google's guidance does not promise a ranking benefit

On 6 November 2026, Google's AI optimization guide states that llms.txt is not required for Google Search, including its generative features, and that Google does not use these files to improve visibility or rankings. It permits maintaining them for other systems that use them. That is guidance about Google, not evidence of what every other client does.

An advertiser should therefore ask what a vendor means by making a page AI-ready. A clean index and an accessible source are observable deliverables. A promise that every assistant will prefer the business is a different claim requiring evidence the file cannot provide. Our sponsored-listing visibility guide explains similar limits for paid publication.

The AEO definition covers the broader effort to make answers useful and verifiable. Maintaining a text index can be one implementation choice in that work. It should not replace the work of answering a buyer's actual question, maintaining source accuracy or explaining the limitations of the evidence on a page.

Check availability, then measure behavior separately

A useful maintenance check starts with an ordinary request to the exact public file URL. Record its status, content and capture time. Parse the links and inspect a selection or all of their destinations according to a declared scope. A file that downloads successfully can still contain stale titles, broken destinations or contradictory summaries.

When the index changes automatically, compare the generated file with the source of truth. Preserve original publication dates on linked articles and update modification dates only for real changes. If a resource moves, update the index and check the destination. A successful build is valuable evidence about generation, but public checks establish what visitors can actually retrieve.

For Dupple's public audience data, that means making methods and limits easy to find beside the figures. The first-party data definition explains why owning a measurement does not make it representative or independently audited. An index should point to those qualifications instead of compressing them into an unqualified headline.

Observed file requests, observed answer citations and attributable business inquiries are separate measurements. A request log does not reveal an assistant's final answer. A brand mention does not establish a visit. Our AI-referral traffic guide and attribution-window definition explain the later measurement boundaries. Keep the advertising glossary available for readers comparing these terms.

FAQ