llms.txt for Contractor Sites: What It Does and Doesn't Do

The reasoning behind the proposal is sound. Web pages are built for people: navigation, ads, and JavaScript wrap the actual information, and converting HTML back into clean text is imprecise and token-expensive for a language model. The spec's answer is a curated map — a required H1 with the site's name, a blockquote summarizing what the site does, then H2 sections containing lists of links with short descriptions. The spec also proposes serving clean Markdown copies of important pages (the same URL with .md appended), so an agent that follows a link gets plain text instead of a page template.

Where llms.txt earns real use is developer documentation. The spec itself notes that AI coding agents are the heavy users, and the AI labs treat it exactly that way: OpenAI, Anthropic, and Google's Gemini team each publish an llms.txt for their own developer docs. We verified Anthropic's on August 13, 2026 — docs.anthropic.com/llms.txt resolves as plain text beginning "# Anthropic Developer Documentation." Note the precise shape of that fact: the labs publish the file for their docs. None of them has committed to fetching it from yours.

llms.txt is not a robots.txt for AI

The filename invites a confusion worth killing early: llms.txt is not an access-control file. robots.txt tells crawlers what they may fetch; llms.txt merely describes what exists. It blocks nothing, permits nothing, and no AI crawler needs it to read your site. Here is how the three root files actually divide the work:

If an AI engine cannot cite your HVAC company, the cause lives in the first row or in your page content — not the third row. That distinction is the difference between a real access audit and a decorative file.

The study is the best measurement of llms.txt that exists, so its numbers are worth walking through. Among those 137K domains, 28% publish a valid llms.txt file — and Ahrefs itself flags that as an upper bound, since its analytics users skew technical. Adoption, in other words, is wide and rising on pure speculation. Readership is not. Of the 3% of files that received any traffic, 96% of requests came from bots, and the bot breakdown dismantles the "AI visibility" story:

Read the last row twice. The retrieval bots — OAI-SearchBot, PerplexityBot, the crawlers that decide what AI search engines cite in live answers — sent 1.1% of requests. Slackbot, a chat app's link-preview fetcher, fetched llms.txt files more often than PerplexityBot did. If the file were a working AI-visibility channel, that is the row where the traffic would live.

Two more findings close the case. First, zero AI-bot requests in the study targeted llms.txt files that don't exist — AI systems never go looking for the file, so publishing one does not put you on any radar; it gets fetched only when something is pointed at it. Second, the one AI audience that genuinely fetches the file is coding agents and agentic infrastructure (10.5% of requests) — which matters if your customers are developers, and approximately not at all if they're homeowners with a broken AC. How the engines that homeowners actually use find and recommend local businesses is its own topic — our comparison of how Perplexity and ChatGPT surface local companies covers those mechanics — but the short version is: they fetch your pages, your reviews, and your mentions. Engines cite sources they can read and verify; a self-declared index file is neither.

Google's Search Advocate John Mueller had said the same thing less diplomatically a year earlier. Responding on Reddit to site owners who saw no bot ever request their file, Mueller wrote:

"AFAIK none of the AI services have said they're using LLMs.TXT (and you can tell when you look at your server logs that they don't even check for it). To me, it's comparable to the keywords meta tag – this is what a site-owner claims their site is about … (Is the site really like that? well, you can check it. At that point, why not just check the site directly?)"

The keywords-meta-tag comparison is the important part. Search engines abandoned that tag because a self-declared description is trivially gamed — and llms.txt has the same trust problem. A file the site owner writes about their own site, disconnected from the pages themselves, is exactly the input experienced engines learned to ignore. In the same discussion, a host managing roughly 20,000 domains confirmed that no AI agents were downloading the files — only niche profiling bots.

One genuine wrinkle deserves honest treatment: Google has sent a mixed signal. In May 2026, days after the mythbusting guidance, the Chrome team shipped an llms.txt check in Lighthouse's experimental agentic-browsing audits — reasoning that without the file, agents may spend more time crawling a site to understand its structure. Note what the audit actually does, though: it flags only server errors, and a missing file is marked Not Applicable, "as providing the file is optional at the moment." When Lily Ray pressed Mueller on the contradiction, Ahrefs reports, he explained llms.txt is "not done for search" — a "temporary crutch, perhaps to save some tokens" for AI coding tools parsing developer documentation. In Ahrefs' bot data, that Lighthouse audit produced 22 fetches — roughly one request in a thousand.

So the answer for an HVAC owner is unambiguous: llms.txt is not a Google ranking factor, not an AI Overviews eligibility factor, and no agency should present it as either. When we build one anyway — next section — it's for a different, smaller reason.

1. Check whether your platform already does it (2 minutes). Wix generates an llms.txt for every site it hosts, and WordPress plugins Yoast SEO and AIOSEO both offer generators. If your site is on one of these, confirm the file exists at yourdomain.com/llms.txt and skip to step 5.

2. List your 10–15 pages that answer real customer questions (5 minutes). Services, service area, pricing, financing, about, contact, and your best guides. Skip blog archives, tag pages, and anything thin.

3. Write the file (8 minutes). Per the spec: an H1 with your company name, a blockquote stating what you do and where, then H2 sections with linked lists. A contractor version in full:

   
   > HVAC installation, repair, and maintenance for the Denver metro.
   > Licensed and insured; residential and light commercial.
   
   ## Services
   - [AC repair](https://example.com/ac-repair): Same-day residential AC repair
   - [Furnace installation](https://example.com/furnace-install): Replacement and new installs with financing
   
   ## Company
   - [Service area](https://example.com/service-area): Cities and response times
   - [Pricing](https://example.com/pricing): Diagnostic fees and typical ranges
   

4. Upload it to your site root (3 minutes) so it loads at yourdomain.com/llms.txt as plain text in a browser.

5. Treat it like code from now on (2 minutes to set up). Ahrefs' study found a research crawler named prompt-injection-survey/1.0 systematically probing llms.txt files — because AI agents are built to ingest and trust them, a compromised or stale file misleads every agent that reads it. Restrict who can edit it, keep the content to plain links and descriptions with nothing instruction-shaped, and update it when your pages change.

Two things complete the honest picture. A file nothing links to will likely never be fetched — remember, AI systems don't go looking — so reference it where machine-readable pointers live if you want to raise the odds above zero. And the file's substance is only as good as the pages behind it: an llms.txt pointing at thin service pages is a well-organized map of nothing. Clean, server-rendered pages that answer real questions are the prerequisite — that's a build decision, and it's why our HVAC website builds ship content in plain HTML before any AI-facing garnish goes on.

Crawler access. An AI engine that cannot fetch your pages cannot cite you, and access fails silently far more often than owners realize — Cloudflare-level default blocks, WAF rules, and robots.txt leftovers do the damage. Running the five-minute crawler access test on your HVAC site is worth more than any number of description files, and it's where every real AI-visibility audit starts.

Content in the HTML. Major AI crawlers do not execute JavaScript, so pages that render their content client-side are invisible to them regardless of what your llms.txt claims. Server-rendered pages that answer customer questions directly — with the answer in the first paragraph, not after 300 words of throat-clearing — are the raw material engines actually quote.

Earned mentions and reviews. Engines cite sources; businesses earn citations through the reputation layer — reviews, directories, local press, real third-party mentions — not through self-declared files. Mueller's keywords-meta-tag comparison explains why: engines trust what others say about you over what you say about yourself. The full playbook is in our guide to getting an HVAC company recommended by ChatGPT.

Measurement over faith. The llms.txt episode is a small case study in why AI-visibility work needs evidence discipline: an entire tool ecosystem grew up scoring a file before anyone verified it was read. The antidote is checking logs and sampling real AI answers — we're building an HVAC AI-visibility research report on exactly that kind of first-party sampling — and it's the standard we hold our own GEO work for HVAC contractors to: measured citations, reported as evidence, never guaranteed. Nobody controls what an AI engine says, us included; anyone selling llms.txt as placement is selling the keywords meta tag with a 2026 haircut.

The honest summary: publish the file if it's free or costs you 20 minutes, skip it without guilt if you're busy, and spend the real budget where the engines actually look — your pages, your access layer, and your reputation.