In this piece: what llms.txt actually does, why Google and Chrome gave founders contradictory signals in the same two weeks, the decision framework for whether you need one, how to write one that isn't a waste of an afternoon, and the one move to make this week.
No, not for Google. Yes, for almost everything else. That's the honest answer to whether your B2B SaaS needs an llms.txt file in 2026, and it's more useful than the "add this now or get left behind" posts flooding LinkedIn.
Google's Search team published guidance on May 15, 2026 stating flatly that llms.txt does nothing for AI Overviews or AI Mode citation, part of a "mythbusting" section aimed at exactly this kind of speculation, as Passionfruit's decision-framework breakdown lays out in detail. Days earlier, Chrome shipped a Lighthouse audit that checks whether your site has one. Two teams at the same company sent founders opposite signals in the same two weeks, and most SaaS blogs picked a side instead of explaining why both are right.
What llms.txt actually is
llms.txt is a markdown file hosted at your domain root (yourdomain.com/llms.txt) that lists your most important pages with a one-line description of each, so an AI system doesn't have to guess what matters on your site. It was proposed in late 2024 and has no formal governing body behind it, just a specification page and a growing pile of opinions.
An Ahrefs study of 137,000 domains found 28% now publish a valid llms.txt file, roughly 38,000 sites. That adoption rate is misleadingly high, since the sample skews toward technically sophisticated, AI-aware site owners who were already likely to add one. Adoption among websites generally is almost certainly lower.
Here is the number that should reset your expectations: of those 38,000 domains with a valid file, 97% received zero requests for it in May 2026. No bots, no crawlers, no humans. The file existed and almost nothing read it. Of the small sliver of traffic that does arrive, most of it isn't even the AI tools founders assume: SEO audit crawlers and unidentified scrapers outnumber every named AI bot category combined.
The mistake founders keep making
The mistake isn't skipping llms.txt. It's treating it as one single decision instead of four different ones, because "does it help AI visibility" depends entirely on which AI surface you mean.
Google Search's May guide is explicit: AI Overviews and AI Mode pull from the same index classic Google ranking uses, and llms.txt changes none of it. John Mueller has compared the file to the old keywords meta tag, a claim a site makes about itself that nobody is obligated to check, let alone believe. Gary Illyes and Amir Taboul confirmed at Search Central Live Deep Dive Asia Pacific that Google isn't pursuing it as a ranking input.
But Anthropic explicitly recommends llms.txt in its Writing for Agents guidance, and OpenAI maintains one for its Agents SDK and the Agentic Commerce Protocol. Chrome's Lighthouse added an Agentic Browsing audit in early May 2026 that checks for the file and notes that its absence means "agents may spend more time crawling the site to understand its high-level structure," according to Ahrefs' breakdown of the standard. The same Ahrefs bot-traffic analysis found that among the small share of llms.txt files that do get fetched, AI agents and agentic infrastructure, not search or retrieval bots, are the single largest reader, which lines up with Mueller's framing that the file mostly serves coding agents rather than search visibility.
Both things are true at once: worthless for the surface most of your traffic comes from today, and genuinely used by the agent layer that's growing underneath it.
The decision: when to build one, when to skip it
Build one if any of these are true for you:
- You run developer docs, an API reference, or an MCP server that agents need to parse.
- Claude or OpenAI agents already show up with meaningful frequency in your traffic logs.
- Your site is under 1,000 pages and marketing owns the CMS, so upkeep is cheap.
Skip it, or stop maintaining it, if any of these are true instead:
- Your only goal is Google Search visibility, including AI Overviews and AI Mode. Spend that afternoon on direct-answer content instead.
- Your site has 10,000-plus pages. The maintenance cost of keeping the file accurate outgrows the agent-readiness upside.
- Your team is already stretched thin on higher-impact SEO work. Do the third-party citation work first.
For a seed-stage B2B SaaS with a marketing site, a docs page, and a handful of integration pages, this resolves to a straightforward call: build it. The cost is roughly one afternoon and the downside is close to zero, aside from making it marginally easier for a competitor to scrape your positioning in one pass instead of ten.
How to write one that's actually useful
Most of the llms.txt files that get ignored deserve it. They're auto-generated sitemaps with the file extension changed, not curated documents. Four things separate a useful one from a hollow one.
- Treat it as an edit, not an export. List 10 to 30 pages that matter, not your full URL inventory. A curated list an agent can actually use beats a 400-line dump every time.
- Point at markdown where you can. Agents parse a .md mirror of a page faster and more reliably than the equivalent HTML. If your CMS doesn't output markdown natively, a lightweight server-side render of your top 20 pages as .md alongside the HTML closes most of the gap.
- Write literal, not marketing. Each line should describe the page the way a buyer or an agent would search for it, not the way your landing page copy sells it. "Pricing tiers and what's included at each" beats "Simple, transparent pricing that scales with you."
- Review it quarterly. A stale file feeding an agent last year's pricing or a deprecated integration is worse than no file at all. Put a recurring calendar reminder on it the same day you publish it, or it won't happen.
What actually moves AI citation, if llms.txt mostly doesn't
The comparison worth internalizing: llms.txt affects browser-agent readiness and a subset of Claude and OpenAI agent workflows. It does not affect Google Search, and its effect on ChatGPT or Perplexity citation is indirect at best. The levers that move citation across every AI surface at once are the same ones that were already true before llms.txt existed: content where the first sentence under a heading stands alone as a complete answer, consistent entity naming across your own site and third-party platforms, and getting cited on the sources these models actually pull from, not just your own domain.
If you've already built the weekly routine that gets a bootstrapped SaaS mentioned by AI assistants, an llms.txt file is a small addition on top of that work, not a substitute for it. The same is true if you're already tracking which of your visits actually originate from AI referrals instead of watching them disappear into "Direct" in your analytics.
Your first move this week
Don't start by writing the file. Start by checking your server logs for GPTBot, ClaudeBot, and PerplexityBot activity over the last 30 days. If any of them show up with meaningful frequency, you already have your answer and your file just needs to exist. If none show up, spend the afternoon on a direct-answer rewrite of your three highest-traffic pages instead. That work pays off on every AI surface, including the ones llms.txt can't touch.
Frequently asked questions
Does llms.txt help my SaaS show up in Google's AI Overviews?
No. Google's May 2026 AI optimization guide states this directly. AI Overviews and AI Mode use the same index as standard Search ranking, and llms.txt has no effect on either.
Do ChatGPT and Perplexity actually read llms.txt files?
OpenAI uses it for its Agents SDK and Agentic Commerce Protocol, and Anthropic recommends it for agent workflows. Perplexity has been observed pulling from it independently of standard retrieval. None of the three treat it as a primary ranking or citation signal in their consumer chat interface.
How long does it take to build one?
For a site under 1,000 pages, a single afternoon: pick 10 to 30 pages that matter, write a one-line literal description for each, and host the markdown file at your domain root.
Is there any real downside to adding one?
The main risk is making it marginally easier for a competitor to scrape a clean summary of your site and positioning in one request instead of many. For most early-stage SaaS companies that risk is small next to the near-zero cost of building the file.
Should I delete my llms.txt file if nobody's requesting it?
Only if you can't commit to reviewing it quarterly. A stale file that feeds an agent outdated pricing or a killed integration is worse than none. If you can maintain it, keep it. If you can't, deleting it is cleaner than leaving it to drift.
What should I prioritize instead if I only have a few hours a month for this?
Direct-answer content structure and third-party citations on the platforms AI models actually cite, like comparison threads, review sites, and Reddit. Those move citation across Google, ChatGPT, and Perplexity simultaneously. llms.txt moves one narrower lever.
Most founders will read this, nod, and never check their server logs. The ones who spend fifteen minutes pulling that log this week are the ones who'll know, three months from now, whether this was ever worth their time instead of still guessing.