I almost credited llms.txt for a Google AI Mode win. Then I read what Google actually says.

I shipped an llms.txt file for a client, their visibility in Google's AI Mode jumped over the same period, and I nearly wrote it up as a clean before/after win for the file.

Then I checked the primary source. This is the honest version — including the part where my first interpretation was wrong.

Context: I'm Christof, a software engineer and technical team lead under Sovaro Consulting in Austria. The client is California Metals, a B2B company specializing in Sustainable Metals — low-carbon alloys and finished components for aerospace, automotive and marine OEMs.

What I observed

Over one reporting cycle, in Google's AI Mode:

California Metals started appearing as the first result for relevant queries.
Two of its certifications were surfaced (a third, ISO 14001, was not — yet).
Its LinkedIn entity gained weight, to the point of showing up twice in the right-hand panel. I had just shipped an llms.txt summarizing the company, its certifications and its key pages. Tempting story: we curated a machine-readable summary, and Google's AI started reading from it. A clean before/after — the kind that's very easy to believe.

It's also almost certainly wrong.

What Google actually says

Google updated its AI optimization guide (Search Central docs, last updated 2026-06-15) with a mythbusting section that names llms.txt directly. The short version: you don't need llms.txt or any special AI markup to appear in Google Search including its generative AI features, because Google Search doesn't use those files. Maintaining one neither helps nor hurts your Google ranking.

This explicitly covers AI Overviews and AI Mode. Per Google's own AI features documentation, those surfaces run on the same index and the same ranking and quality systems as ordinary organic results. AI Mode adds a "query fan-out" — it spins off several related sub-queries and pulls from the normal Search index — but there's no separate llms.txt channel feeding it.

So the file I shipped was not what put my client at the top of AI Mode. Google grounds AI Mode in what it already crawls and trusts across the open web. Which also means the certifications shown there come from sources Google crawls — not from my llms.txt. ISO 14001 will appear in AI Mode when Google picks it up from credible pages (the site itself, certification registries, structured data), not "automatically when I update the llms.txt." That assumption was the core of my mistake.

"But doesn't ChatGPT read it?"

This is the most common pushback, so it's worth being precise. The bots that actually feed ChatGPT — OAI-SearchBot for its search answers, GPTBot for training — crawl HTML directly, and OpenAI's own crawler documentation manages them through robots.txt without mentioning llms.txt at all. Independent log studies covering hundreds of millions of AI-bot requests found hits to /llms.txt to be statistically negligible. GPTBot reportedly fetches the file now and then, but fetching isn't the same as using it to choose or cite sources.

Where llms.txt genuinely gets read is a different layer: coding and agent tooling — Cursor, Claude Code, GitHub Copilot, Windsurf — pulling a documentation site's pages with less token waste, plus emerging agent protocols like OpenAI's Agents SDK. That's real, and it's growing fast.

But notice what that use case requires: documentation worth fetching, and an audience that queries AI tools about your product.

What probably actually moved the needle

Looking at the same reporting period honestly, the likeliest drivers are textbook SEO, not the text file:

Entity/authority signals. The client's LinkedIn presence gained weight in that window — that's exactly the kind of corroborating entity signal Google's AI surfaces lean on. The double-listing is a symptom of a strengthening entity, not of a Markdown file.
The rest of the cycle's work: clearer content, internal linking, crawlability, consistent NAP/brand data across the web. None of that is glamorous, and none of it needs an AI-specific hack. That's the actual takeaway Google keeps repeating: AI search visibility is SEO.

So is llms.txt useless? No — but be precise

It's still worth shipping, for the right reason: it's a business-to-agent (B2A) surface for non-Google LLMs and agent tooling (ChatGPT and Claude agents, Cursor, Windsurf), not a Google ranking lever. Just do it correctly:

Name it llms.txt — plural — and serve it at the domain root (/llms.txt) with a text/plain content type. That's the only path agent tools look for.
Lead with a tight, factual blockquote — that paragraph is what an agent lifts as its mental model of the brand. No marketing adjectives.
Curate, don't dump. A short, high-signal link list.
Only list indexable, canonical pages. The file itself is trivial Markdown:

# California Metals

> Supplier of low-carbon, sustainably sourced metals, alloys and finished
> components for aerospace, defense, automotive, marine and luxury-goods
> manufacturers. Certifications: AS9100, ISO 14001:2015, ISO 20400.

## Core pages
- [Capabilities](https://www.californiametals.com/...): Materials and processes.
- [Sustainability](https://www.californiametals.com/...): CO2-reduction and sourcing standards.
- [Industries](https://www.californiametals.com/...): Sectors and applications.

Most company sites are on WordPress, where a static file in the web root (or a Yoast/Rank Math toggle) is enough. If you use Next.js, serve it from a route handler — same source of truth as the sitemap:

// app/llms.txt/route.ts
export const dynamic = "force-static";

export function GET() {
  const body = `# California Metals

> Supplier of low-carbon, sustainably sourced metals and alloys...
`;
  return new Response(body, {
    headers: { "Content-Type": "text/plain; charset=utf-8" },
  });
}

The takeaway

The improvement was real; the attribution wasn't. Google's AI surfaces rewarded ordinary SEO and a strengthening brand entity — the llms.txt came along for the ride rather than driving it. Ship the file as cheap B2A infrastructure if you want to be early for agent traffic, but keep it on the right side of the ledger: it isn't what gets a non-developer site to the top of AI Mode. Clear content, clean crawlability, and a coherent brand entity are.

Written by Christof Karisch — software engineer and technical team lead at Sovaro Consulting. Work done for my client California Metals, a sustainable metals supplier.

推荐订阅源

DEV Community