The reaction is nearly always the same, and it is nearly always the wrong diagnosis. An owner sees a competitor named in an AI answer and concludes one of two things: that the competitor is paying for the position, or that the assistant has judged their work to be better.
Neither is what happened. Understanding what actually happened is the difference between fixing this and spending a year on the wrong problem.
There was no comparison
Assume, for a moment, that an assistant were genuinely evaluating businesses. It would need to know both of you exist, hold comparable information about each, and apply some consistent standard of quality. None of those conditions hold. It has no roster of businesses in your city. It has no way to assess workmanship, outcomes, or how you treat people who call at five o'clock on a Friday.
What it has is text. When a customer asks which firm to use, the assistant assembles the most plausible answer from the material available to it — its training data, and whatever pages its search back-end retrieves at that moment. If your competitor's name appears in that material attached to the category and the city, and yours does not, the answer writes itself. No comparison took place. There was only one candidate in the room.
This is the part worth sitting with, because it inverts the usual instinct. Your competitor did not beat you. They were the only one present.
What "present" actually means
Presence is not the same as having a website, and it is not the same as being findable. Three things make a business legible to an assistant:
It is named alongside the category and the place. "Cedar & Stone, a standing-seam metal roofing contractor in West Yorkshire" is a complete unit of information. An assistant asked about roofers in West Yorkshire can retrieve it. A homepage that says "Quality Roofing Since 1998" over a photograph of a roof supplies almost nothing retrievable, because it never states what kind of roofing, for whom, or where.
It appears on pages that compare things. Assistants lean disproportionately on sources whose whole purpose is to enumerate options: directories, trade association member lists, "best X in Y" roundups, review platforms, forum threads. These pages are structurally convenient. They already contain a list of names in a category, in a place, with reasons attached. An assistant composing a recommendation is doing the same job those pages did, so it borrows their work.
It is described the same way in more than one place. One mention is an anecdote. The same description on your site, a directory, an association listing, and a local news piece is a pattern — and patterns are what a model retains. Consistency is not a branding nicety here. It is the mechanism by which scattered mentions accumulate into a single recallable fact.
Your competitor probably did not set out to do any of this deliberately. Most who win on AI assistants got there by being in a directory since 2019, or by being the example someone reached for in a blog post about the local market.
Where the advantage compounds
There is a second-order effect that makes this urgent rather than merely interesting.
Once a business is regularly named in AI answers, people write about it more. Journalists researching a piece use an assistant. Someone compiling a "best in the region" roundup uses an assistant. A prospective customer, having heard the name from ChatGPT, searches it, visits the site, and occasionally mentions it somewhere public. Each of those becomes new text, which becomes new material for the next answer.
The gap does not stay the same size. The business that got there first accumulates the mentions that keep it there, and the cost of catching up rises every quarter. This is not a race that rewards moving at the same speed as your competitors; it rewards noticing earlier.
What to do with the finding
The useful response to seeing a competitor recommended is not to look at their website. It is to look at why the assistant named them — which usually means reading the citations, where the assistant shows them.
That list of sources is the actual scoreboard. It tells you which four or five pages control the answer in your category and your city. Those pages are a finite, addressable list. Some you can appear on by asking. Some require a correction to a listing you already have but never checked. Some are review platforms where you have a profile with three reviews and a competitor has sixty.
None of that is guesswork once you have the citation list in front of you. The work is unglamorous and specific, which is the good news: unglamorous and specific problems are the kind that actually get solved.
The bad news is that you cannot do any of it until you know what the assistants are currently saying, across all four of them, for the questions your customers genuinely ask. Guessing at that is worse than useless, because the answers vary by assistant, by phrasing, and between runs — and the version you happen to see once is not the version your customers are reading.