llms.txt for Local Service Businesses: What to Put In It and What It Won't Do

llms.txt is a proposed plain-text file, placed at the root of a domain (yoursite.com/llms.txt), containing a short Markdown-formatted summary of the site: what the business does, links to key pages with one-line descriptions, and sometimes an expanded version at /llms-full.txt with more detail. The idea, proposed by developer Jeremy Howard in September 2024, is to give a large language model a compact, curated map of a site instead of forcing it to crawl and parse the whole thing.

It is a convention, not a standard ratified by any search engine or AI lab, and no major AI company — not OpenAI, not Anthropic, not Google, not Perplexity — has publicly committed to systematically fetching and using it as part of how they answer questions. Some AI tools built for developers (certain coding assistants and documentation-focused crawlers) have adopted it more directly, particularly for technical documentation sites, but that use case is different from a local roofing or HVAC company hoping a consumer-facing assistant will read it.

robots.txt is a directive file: it tells crawlers what they are and are not allowed to access, and well-behaved crawlers (including the major AI crawlers) respect it. sitemap.xml is a complete inventory: it lists every URL you want indexed, with no editorializing about which pages matter most. Both have been standard for two decades and are honored by essentially every crawler that matters.

llms.txt is neither a directive nor an inventory. It is a curated, human-written pitch — closer in spirit to an executive summary than to a technical protocol — and it is entirely voluntary on both ends: voluntary for you to publish, and voluntary (currently mostly unconfirmed) for any AI system to read. That distinction is the single most important thing to understand before investing meaningful time in one.

Below is a realistic llms.txt for a hypothetical HVAC company, with the reasoning behind each section. Adapt the specifics to your trade and market.

Do not use llms.txt as a place to stuff keywords, marketing copy, or claims you would not put on the visible site — there is no evidence it carries special ranking weight, and inflated claims here carry the same reputational risk as anywhere else if a system does surface them. Do not duplicate your entire site's content into it; that defeats the purpose of a concise summary and risks becoming stale immediately.

Do not include private or sensitive information — internal pricing formulas, unpublished phone lines, staff personal details, or anything not already intended for public consumption. And do not treat it as a place to make claims that contradict your schema or your visible pages; any inconsistency across these surfaces undermines the entity corroboration you are trying to build, which matters far more than the file itself.

Host it as a plain UTF-8 text file at the domain root, same level as robots.txt. It requires no server configuration beyond that — any static file host or CMS that allows a custom file upload can serve it. Keep it under roughly one to two printed pages; the value proposition collapses if it becomes long enough that no summarization benefit remains.

The maintenance burden is the real cost. Every time you add a service, change a service area, update pricing, or retire a page, the file needs a matching edit or it becomes actively misleading rather than merely unused. Assign it to whoever already owns your schema and NAP consistency, and review it on the same cadence — quarterly is reasonable for most local service businesses.

No AI assistant guarantees it reads llms.txt, and none has published documentation stating it factors into how businesses get named or ranked in consumer-facing answers. Publishing one will not move your share of AI-assistant mentions in any measurable, attributable way for most local service categories, and you should not budget meaningful time against an expectation that it will.

The honest case for building one is small but real: it costs little to create once your content is already organized, it does no harm if kept accurate and in sync, and if adoption among consumer AI assistants does expand over the next few years, having a well-maintained file already in place costs you nothing to have started early. Treat it as a low-priority, low-cost addition — not a project that competes for time against schema, content, or review generation.

Unlike llms.txt, the crawler directives in your robots.txt file have a confirmed, immediate effect: they determine whether specific AI companies' bots can access your content at all. The major named user-agents as of 2026 include GPTBot (OpenAI), ChatGPT-User (OpenAI, used for live browsing during a chat), PerplexityBot (Perplexity), ClaudeBot (Anthropic), and Google-Extended (which controls use of your content for Google's AI features, separate from standard Googlebot indexing).

Blocking these bots is a real, binary decision with tradeoffs, not a style choice. Allowing them makes it possible for your content to be cited in AI answers; blocking them removes that possibility entirely for that assistant, full stop, regardless of how good your schema or content is. Some publishers block AI crawlers over content-licensing concerns, which is a legitimate business decision for content-heavy media companies — but for a local service business whose entire objective with AI search is to be named and recommended, blocking these bots directly works against your own goal. Check your current robots.txt for accidental blanket disallows (sometimes left over from a staging site configuration) before doing anything else in this space.

If you think of AI visibility as a stack, llms.txt sits at the very top as an optional, lightweight summary layer. Structured data (schema markup) sits below it as a confirmed, machine-readable layer that both traditional search and AI systems demonstrably use. Below that is the foundation: plainly written, answer-shaped content on your actual pages — content that states your service area, your pricing model, and direct answers to real customer questions in ordinary sentences a system can lift and quote.

In practice, an llms.txt file that points to well-structured, schema-backed, answer-shaped pages is coherent and low-risk to add. An llms.txt file built as a substitute for that underlying work is a wasted afternoon dressed up as an AI strategy. Spend the bulk of your effort on the foundation; treat llms.txt as the last five percent, not the first fifty.