Does agentic SEO need llms.txt, Markdown or an accessibility tree?

What does agentic SEO mean here?

The label can describe improving discovery in AI answers, making a website usable by agents, or using an agent to perform SEO work. A proposal that moves among those meanings can sell an implementation without identifying its benefit.

Write the intended job first. “Make the product page eligible for search retrieval” concerns access and indexing. “Let a buyer's assistant identify delivery restrictions” concerns accurate extraction. “Let an authorised assistant select a date and reach booking confirmation” concerns an interactive journey. Success in one does not establish success in the others.

For general discovery, start with the AI-search visibility Answer. The question here is whether particular technical changes solve an observed problem.

What does llms.txt actually offer?

llms.txt is a proposed convention for a concise site summary and links to useful material. The current v2 proposal includes ways to point agents towards clean Markdown versions. It can be useful for documentation that a known agent consumes; adoption is not a general search-ranking experiment. The proposal

Google Search and Chrome’s Lighthouse assess different things. Search Central says llms.txt and special Markdown neither help nor harm Google Search visibility because Search ignores them. Chrome's Lighthouse has an agentic-browsing audit for llms.txt, but its documentation says a missing file returning 404 is optional and marked Not Applicable. A server error retrieving an offered file is a different finding. Search Central, Lighthouse's audit definition

These statements address different systems. A supplier cannot turn the existence of a Lighthouse audit into evidence that Google Search requires the file. Ask which finding occurred and which customer task the proposed change improves.

Treat llms.txt and other public copies as published information, not access controls. They cannot protect private material or grant a user permission to act. Keep those decisions in the actual access and authorisation systems.

When can Markdown help?

Markdown expresses headings, lists and links with less page furniture than many HTML documents. It can make retrieval more economical for a client designed to use it. Whether it improves a particular agent's answer still needs testing.

Cloudflare offers opt-in HTML-to-Markdown conversion when a client requests that format. Assess the delivery capability on its own merits, with separate evidence required for any citation or sales uplift. Its documentation describes separate cached formats. Where the origin supplies no content-use signal, the converted response adds a default saying AI training, search and AI-input use are allowed. Check that policy choice before enabling the feature; the signal itself cannot guarantee how recipients use the content. Cloudflare’s documentation

Generate offered formats from the same maintained, approved source, preserving substantive facts, restrictions and links. A shorter response that drops a refund condition is a worse source for a buying decision. Avoid a separately maintained Markdown price list that drifts away from the live offer. Check public and signed-in behaviour where the site has both; a format change should not redefine which information is public.

What does the accessibility tree change?

The browser derives an accessibility tree from the page: names, roles, states and relationships that assistive technology can use. It is not another text file to upload. Semantic controls, associated field labels, understandable errors and stable layouts help people and can help browser agents identify the intended action. Agents may combine this representation with HTML and screenshots. Google's agent-friendly website guidance

A booking button with a meaningful name and a date selector with an exposed selected state provide a better task interface than visually styled but ambiguous elements. This work has a user-experience purpose even if no AI-search impression changes. Measure the relevant journey rather than attaching a citation target to every accessibility improvement.

Prefer appropriate native buttons and links. Adding a button role to a visual box does not by itself implement keyboard interaction, focus or the intended action. Acceptance should include those behaviours and check that essential information remains available to assistive technology. A name exposed in an audit cannot establish that the control works.

Search access and training permission also need separate choices. OpenAI documents independent controls for OAI-SearchBot and GPTBot; allowing a training crawler is not a prerequisite inferred from an agent-friendly layout. OpenAI's crawler guidance

How would you test the proposed changes?

Use this constructed booking-page exercise. The approved source says £40 per person, cancellation allowed until 48 hours before the session, original capacity twelve and three places currently available. The task is to explain the terms, select two places for an available date and stop before an unapproved purchase. These are test inputs and expected decisions, not an agent trial we have run.

Suppose the ordinary HTML exposes all those facts, but a proposed Markdown response omits the cancellation condition. The extraction test fails even if the response is smaller. Restore the condition before assessing efficiency. A claim that “the page is easier for AI” is insufficient acceptance evidence.

Now suppose both formats are accurate, but the browser interface exposes three identical “Select” controls without the associated dates. The actionable repair is to make the options and selected state distinguishable. Adding more Markdown cannot identify which date the control will choose.

Finally, suppose the interface works and the agent correctly reports three available places. That does not prove the page is indexed, cited or generating new demand. Test retrieval eligibility separately. Record the agent/service, version when available, exact inputs, browser state, approved stopping point, errors and successful completion; repeat on held-back cases, including no availability and a changed price.

Which work deserves priority?

Repair demonstrated access, factual or task failures first. Check crawl access, intended indexing, rendered content and any relevant security rule before adding formats. For Google’s AI features, also inspect the effective Search generative AI control in Search Console where eligibility is in doubt; Include is the default, and inherited exclusions need checking. This choice is separate from AI training. Google’s inclusion-control help

Add llms.txt or a Markdown endpoint when there is a named consuming client, an affordable maintenance owner and an observable benefit worth testing. Keep ordinary public HTML useful throughout.

Assess an agentic-SEO proposal through the affected task, baseline failure, proposed change, acceptance evidence and ongoing cost, including when the terminology is fashionable. Require separate results for extraction accuracy, task completion, search exposure and commercial contribution.