The question a buyer actually types
Not a keyword. A full sentence, with a place in it and a job in it.
Search used to mean typing a few words and reading ten blue links. Asking an AI assistant is different. A buyer opens ChatGPT, or a similar assistant, and writes a real sentence, the way they would ask a person: who does this kind of work, in this town, and can they do it soon. The assistant does not hand back ten links. It reads what it can find and writes one answer, and that answer names, at most, a handful of businesses.
That changes what a business is competing for. It is not competing for a spot on a results page anymore. It is competing to be one of the names the assistant decides to put in its answer, for that exact question, asked that exact way. A page built to rank for a keyword and a page built to answer a buyer's question are not the same page, and it is usually only the second one that gets quoted.
The search behind the answer
An assistant does not remember a business. It goes and reads for it, right then.
When an assistant answers a buyer's question about a local business, it is not pulling the answer from memory. It runs a live web search first, the same moment the question is asked, and builds its answer from whatever it can read. ChatGPT's search, for example, draws on Bing's index and on OpenAI's own crawler, OAI-SearchBot. Other assistants read the web with their own crawlers: names like GPTBot, PerplexityBot, ClaudeBot, Google-Extended and Bingbot show up in server logs doing the same kind of reading.
Two files decide whether any of that reading ever happens. robots.txt can allow or block each of those crawlers by name, and a site that blocks them is invisible to the search behind the answer, no matter how good the page underneath is. llms.txt is the newer file, a plain-language page some sites add that tells an assistant what the business does and where, in the assistant's own terms, instead of making it guess from the rest of the site.
Why one competitor gets named and another does not
Two real businesses ran the same six-question check. The gap between them was not small.
The gap is not theoretical. These are two real first runs of the same check, on two real businesses. Neither is named here, on purpose.
Named in 0 of 6
First real run, an HVAC contractor
An HVAC contractor in San Antonio, in the market twenty years, was named in zero of the six buyer questions asked about the work it does every week. In every one of the six answers, a competitor was named instead.
Named in 5 of 6
Second real run, an exterior remodeling contractor
An exterior remodeling contractor in the Chicago suburbs was named in five of the six answers. It already had an llms.txt file in place, the plain-language file described above, before the check ever ran.
Both are established local businesses doing real work their buyers actually need. The difference between zero and five was not the work. It was what an assistant could find, read and quote when it went looking, which is exactly what the rest of this page is about.
What an assistant looks for on a page
An assistant is not persuaded. It is matched against a question, one part at a time.
When an assistant reads a page while trying to answer a buyer's question, it is checking that page against the question almost line by line. These are the same five things it checks for, in order.
- 01The question in the heading. Not a slogan, not just a service name. The heading should read close to what the buyer actually typed. A fictional example: a page titled "Who repairs a broken garage door spring in Naperville, IL" is answering a question an assistant is holding. A page titled "Garage Door Services" is not.
- 02The answer in the first paragraph. An assistant reads top to bottom and often stops at the first paragraph that answers the question. Bury the answer under three paragraphs of company history, and the assistant may already be reading the next source.
- 03The area. The buyer's question named a place, so the page has to name it back, in words on the page, not only inside a map embed the assistant cannot read.
- 04The specifics. What exactly the business does, for that exact question, stated plainly enough that it could not be true of every competitor's page as well.
- 05The proof. Business details in a form an assistant can read as fact rather than prose: LocalBusiness or Service schema for hours, address and service area, plus whatever licenses or certifications back up the claim.
Why the site, the Business Profile and the directories have to agree
An assistant does not call to ask which listing is right.
A buyer question and an assistant's search do not stop at a business's own website. The assistant also reads what it already trusts: the business's Google Business Profile, and the directories and review sites that come up for that kind of question. If the phone number on the website, the hours on the Business Profile and the address on a directory do not all say the same thing, the assistant has no way to know which one is current.
It does not pause to sort that out. It either follows the version it already trusts more, which is often not the business's own site, or it leaves the business out of the answer rather than state something that might be wrong. Getting the site, the Business Profile and the directories to agree is not a formality. It is part of what makes a business readable enough to trust.