Agent readiness

What is llms.txt, and do you actually need one in 2026?

The honest answer is nuanced: llms.txt is a ten-minute, zero-risk addition with no evidence behind it, sitting inside an agent-readiness stack where other artifacts demonstrably matter more. Here is what the file is, what the data shows, and what to prioritize.

Updated July 13, 2026 · 9 min read

Ask ten SEO consultants whether your site needs an llms.txt file and you will start a fight. Proponents call it the robots.txt of the AI era. Skeptics point out that no major AI company has ever confirmed using it for anything. Both camps are partly right, and the honest answer fits in one sentence: llms.txt is a cheap, harmless addition that belongs near the bottom of your agent-readiness list, not at the top.

The stakes behind the question are real. Salesforce reported that AI influenced $262 billion of 2025 holiday sales, and assistants now summarize, recommend, and increasingly book and buy on behalf of their users. Business owners are right to ask what makes a website readable to these agents. The problem is that llms.txt got marketed as the whole answer while the artifacts agents demonstrably consume went ignored.

This guide covers what the file actually is, what the evidence says (with sources, because this topic is drowning in vendor claims), how to create one in ten minutes, and the priority order for everything that should come before it.

What llms.txt actually is

llms.txt is a plain markdown file served at the root of your domain, so it resolves at yoursite.com/llms.txt. It was proposed in September 2024 by Jeremy Howard, co-founder of Answer.AI, and the spec lives at llmstxt.org. The pitch: language models have limited context windows, and a typical web page is mostly navigation, scripts, and boilerplate. A hand-curated markdown map lets a model skip the noise and load only what matters.

The format is deliberately simple. An H1 with the site name, a blockquote summary, then H2 sections containing annotated links to your canonical pages. For a physiotherapy clinic it might look like this:

# Riverside Physio
> Sports physiotherapy clinic in Austin, TX. Initial assessments,
> follow-up treatments, and dry needling. Online booking available.

## Services
- [Initial Assessment](https://riversidephysio.com/assessment): 60 min, $140
- [Follow-up Treatment](https://riversidephysio.com/follow-up): 30 min, $85

## Policies
- [Cancellations](https://riversidephysio.com/policies): 24-hour notice required

There is also a heavier sibling, llms-full.txt, which inlines the complete text of every page into one giant file instead of linking out. Documentation sites adopted the idea fastest: Mintlify generates both files automatically for the docs it hosts, and Anthropic publishes one for its own developer documentation. That docs-world adoption is the strongest real-world signal the format has.

The case proponents make

The arguments for llms.txt are plausible on their face, which is why the file spread so fast:

• Token efficiency. An agent that fetches one curated markdown file spends far fewer tokens than one that crawls and strips twenty HTML pages.
• Narrative control. You decide what an AI reads first: your best pages, your current prices, your actual service area, instead of whatever a crawler happens to land on.
• Cheap insurance. It costs ten minutes and nothing can break. If agents start honoring it later, early adopters are already in place.
• Precedent. robots.txt and sitemap.xml both started as informal conventions before becoming universal. Maybe llms.txt follows the same path.

Every one of those claims describes what the file could do. The open question is whether anything on the consuming side actually honors it. That is where the evidence gets uncomfortable for the maximalists.

What the evidence actually shows

Start with the strongest data point. Ahrefs studied sites with and without llms.txt and found no correlation with AI visibility or rankings. Sites that added the file did not get cited more often in AI answers, and their traffic from AI sources looked no different from sites without it.

Google has been unusually direct. Its guidance says llms.txt is not required and that Google’s systems do not use it; Gemini and AI Overviews work from ordinary Google crawling and indexing. Google’s John Mueller went further and publicly compared llms.txt to the keywords meta tag, the canonical example of a signal search engines learned to ignore because site owners could write anything they wanted into it.

OpenAI and Anthropic tell a similar story by omission. Both document that their crawlers (GPTBot, ClaudeBot) obey robots.txt, but neither lists llms.txt as an input to search, browsing, or shopping features. When ChatGPT surfaces products to buy, OpenAI’s own merchant documentation points to structured product feeds, not markdown manifests. There is a real irony in Anthropic publishing an llms.txt for its docs while never committing to read anyone else’s.

None of this makes the file useless. It costs ten minutes, it cannot hurt you, and some smaller agents and custom crawlers do fetch it opportunistically. The problem is opportunity cost: every hour spent perfecting llms.txt is an hour not spent on the artifacts with actual evidence behind them.

What AI agents demonstrably consume

While llms.txt sits mostly unread, real agent traffic flows through a different set of artifacts. These are the ones with documented, verifiable consumption today:

Clean, crawlable HTML is the baseline. Every browsing agent and AI crawler parses your rendered pages, and content that only appears after JavaScript executes is invisible to many of them. Structured data comes next: JSON-LD using schema.org vocabulary is documented by Google, parsed by Bing, and read by the answer engines, because it turns prose into typed facts (prices, hours, service areas, review scores) a machine can trust.

Then come the commerce-specific surfaces. ChatGPT’s shopping experience runs on merchant product feeds, and OpenAI revamped Instant Checkout in March 2026 toward discovery-first, feed-driven merchant surfacing. Google’s UCP works the same way on its side. And for agents that act rather than just read, machine-readable action surfaces (OpenAPI specs, agent.json files, MCP servers) let an agent check availability or complete a booking instead of just describing your business. If those acronyms are new, the UCP vs ACP vs MCP explainer untangles which protocol does what.

Notice the pattern. Everything with strong evidence is either structured (feeds, JSON-LD) or executable (APIs, MCP). llms.txt is neither; it is a suggestion, and suggestions only work when someone commits to reading them.

How to create an llms.txt anyway

Since the cost is ten minutes and the downside is zero, there is no reason not to have one once the higher-priority work is done. The manual route:

1. Open a text editor. Start with an H1 (# Your Business Name) and a two-sentence blockquote describing what you do, where, and for whom.
2. Add one H2 section per topic (Services, Pricing, Policies, Contact) containing markdown links to your canonical pages, each with a one-line note.
3. Keep it short. Curation is the entire point; pasting your sitemap into it defeats the purpose.
4. Save it as llms.txt and upload it to your web root so it loads at yourdomain.com/llms.txt as plain text. No registration or submission step exists; being at that URL is the whole deployment.
5. Put a review reminder on your calendar. A file quoting last year’s prices is worse than no file at all.

If you would rather not hand-write it, free generators exist from several SEO and docs vendors. Nexez’s Agent-Ready Kit takes a different angle: it generates llms.txt as one artifact in a bundle that also includes JSON-LD, an agent.json link, and redirect recipes for your own site, and the WordPress plugin injects the bundle automatically. The point is not the markdown file; it is that the file ships alongside the artifacts agents actually use.

The priority-ordered agent-readiness checklist

If you get one afternoon a month for this work, spend it in this order. Each step builds on the one before it.

1. Make sure agents can fetch your pages at all. Server-render your content, keep critical information out of JavaScript-only rendering, and check that robots.txt is not blocking GPTBot, ClaudeBot, or Google-Extended unless you mean to.
2. Add JSON-LD structured data for your business, services, and offers. This is the highest-leverage hour on the list: it is cheap, and the consumers are documented.
3. Publish a structured catalog or feed if you sell anything or take bookings. Feeds are the enrollment ticket for agentic shopping surfaces; the ACP enrollment guide walks through the OpenAI side.
4. Expose actions, not just facts. An OpenAPI spec or MCP server lets an agent move from "here is a physio clinic" to "I booked your Tuesday 3pm." For service businesses this is where AI booking becomes real revenue rather than a demo.
5. Then add llms.txt. At this point it costs nothing, and there is a nonzero chance some agent someday reads it.

Steps one through four are exactly the stack Nexez publishes for a business automatically: structured listings served as clean HTML plus JSON-LD, agent.json, llms.txt, an OpenAPI spec, and a per-merchant MCP server, with ACP and UCP feeds for the shopping surfaces. Discovery artifacts are included on every plan.

So, do you actually need one in 2026?

Need? No. Nothing measurable happens when you add llms.txt today, and both the biggest study and the biggest search company say so plainly. Anyone telling you otherwise is selling something.

Should you have one? Sure, once the real work is done. It is the cheapest line item in the agent-readiness stack, and the format could still win adoption the way sitemap.xml eventually did. Just be honest with yourself about which artifact earns the next hour of your time. In 2026, that hour belongs to structured data, feeds, and callable endpoints, in that order, with llms.txt riding along for free.