llms.txt V2, published on August 10 by Answer.AI developer Jeremy Howard, adds a formal way for AI agents to locate the Markdown versions of web pages. Markdown here means a clean, plain-text version of a page stripped of navigation, ads, and JavaScript. The update is the first revision since the format launched as a proposal in 2024, and it addresses a gap that thousands of early adopters discovered: the original file told agents which pages existed but gave no standard method for finding those cleaner, agent-readable versions.
What changed in llms.txt V2?
The original spec introduced a plain-text file placed at a website’s root, listing key pages with brief descriptions. V2 keeps that foundation and adds two concrete improvements: new URL patterns and new link relations that let agents navigate directly to Markdown content without guessing or parsing raw HTML.
According to the updated spec on GitHub (Answer.AI, August 2026), V2 now supports two URL patterns for Markdown versions of pages:
- Append .md to the full filename: A page at /docs/tutorial.html becomes /docs/tutorial.html.md. This was the only pattern supported in V1.
- Replace the extension entirely: The same page can also be served as /docs/tutorial.md. V2 formally recognizes this as an equally valid option.
Alongside the URL patterns, V2 introduces two link relations, which are signals embedded in a page or its server response that point agents directly to the right files:
- rel=”alternate” type=”text/markdown”: Links from an HTML page directly to its Markdown version.
- rel=”describedby”: Points to the llms.txt file that covers a given page. A single llms.txt placed at /docs/ covers everything under that path, so every page in /docs/ can reference it with this relation.
Both link relations can be delivered as standard HTML <link> elements inside a page’s <head>, or as an HTTP Link response header (a signal the web server sends alongside the page, without touching the page’s HTML code at all). The header option works on non-HTML files too, including Markdown pages themselves, and can be added through server or CDN (content delivery network) configuration without modifying any templates.
Why did the spec need a V2?
Howard describes V2 as the result of two years of real-world adoption. According to the GitHub spec page (Answer.AI, August 2026), thousands of sites now publish an llms.txt file. Documentation platforms like Mintlify generate one automatically for every site they host. Google Chrome’s Lighthouse tool, a site-quality checker, added a check for llms.txt as part of its Agentic Browsing category, which was introduced in May. Anthropic, OpenAI, and Google’s Gemini team all publish llms.txt files for their own developer documentation.
The practical gap that emerged: agents could read the llms.txt file and find a list of relevant pages, but nothing in the spec told them where the Markdown versions of those pages lived. Agents were left to guess or fall back on parsing raw HTML, which is slower and noisier. As Howard notes in the spec, every unnecessary token (unit of text an AI processes) costs time and money for the agent running the task. V2 closes that gap with explicit, standardized signals.
Does llms.txt affect Google Search rankings?
No. Google’s Search team has said for over a year that llms.txt support is not on their agenda. Google’s official AI optimization guide (developers.google.com) states plainly that Google Search does not utilize these files, and that maintaining one “will neither harm nor help your site’s visibility or rankings.”
Google Chrome’s Lighthouse tool does check whether a site publishes an llms.txt file as part of its Agentic Browsing audit. However, according to Search Engine Journal’s reporting on V2 (August 2026), a missing file that returns a 404 is flagged as N/A rather than a failure, because the file is optional. Lighthouse does not yet test for V2’s new link relations or Markdown-page discovery signals.
The format’s primary adoption is in developer tooling. Coding agents use llms.txt to find API references and documentation faster. That use case is entirely separate from Google Search ranking, and V2 makes the spec more useful in exactly that context.
What This Means for AI-Search Visibility
For a business owner, the simplest way to understand V2 is this: before V2, an AI agent reading your site was like a visitor handed a building directory but left to figure out room layouts on their own. V2 puts up clear signage inside each room, pointing back to the directory and to a tidier version of that room’s content. The agent gets what it needs faster, with less wasted effort.
Hingewise’s read is that this matters most for businesses whose products, services, or documentation are frequently referenced in technical content, because coding agents and AI-powered IDE (integrated development environment) tools are most active in those contexts today. A SaaS API, a developer tool, a knowledge base: these are the assets that agents fetch repeatedly.
Two observations worth adding that the sources do not explicitly state:
First, the adoption signal already tells you who the real audience is. Anthropic, OpenAI, and Google’s Gemini team publish llms.txt files not to influence search rankings but because agents using their APIs benefit from faster documentation access. The format is agent infrastructure, not search crawler bait.
Second, a genuine split is forming between two distinct AI visibility tracks. One is SEO-oriented AI visibility: AI Overviews, AI Mode, and retrieval-augmented generation (RAG, the technique AI search uses to pull relevant web pages before generating an answer), all of which depend on Google’s core ranking systems. The other is agent-oriented AI visibility: coding agents, document tools, and AI assistants that fetch content directly, bypassing search entirely. The llms.txt spec lives firmly in the second category. These tracks are not competing, but they require different preparations. A business optimizing only for Google’s AI features may be invisible to agents that never touch a search index. That split is worth monitoring as agent use cases expand beyond developer tooling.
For sites already publishing llms.txt, V2 requires only a minor addition, not a rebuild. For sites that have not yet considered the format, V2 offers a clear, stable spec to judge whether their content is the kind agents actively fetch.
Quick checklist: before adding llms.txt V2 signals
- Confirm your site already publishes a valid /llms.txt file, or a path-specific one such as /docs/llms.txt.
- Check whether your platform (Mintlify, nbdev, or similar) generates llms.txt automatically, as many already do and may add V2 support in their own updates.
- Decide which URL pattern fits your server setup: appending .md to the full filename, or replacing the file extension.
- Add the rel=”alternate” type=”text/markdown” link relation to pages that have a Markdown version available.
- Add the rel=”describedby” link relation pointing to the llms.txt file that covers each page.
- Consider delivering link relations via HTTP headers if modifying page templates is impractical. This can be configured at the CDN or web server level without touching page code.
- Verify in your server logs or via a browser’s developer tools network panel that the headers are being sent correctly before declaring the update live.
The next development to watch is whether AI agent frameworks and coding tools update their behavior to actively look for V2 link relations, and whether documentation platforms beyond Mintlify and nbdev add automated V2 support. The spec is stable. Uptake on the tooling side will determine how quickly these new signals become standard practice across the agent ecosystem.
Lam Nguyen · Hingewise
Sources: Search Engine Journal, August 2026; Answer.AI / llms-txt spec, GitHub, August 2026; Google AI Optimization Guide, developers.google.com