Where answer engines get local recommendations

The question we get asked most this year, usually near the end of a meeting, is some version of “how do I get into the AI answer”.

Nobody submits a business to an answer. There is no queue and no placement to buy. When someone asks AI Mode, Gemini or ChatGPT for a good dentist near them, the machine reads what already exists about the businesses in that area and writes a paragraph out of it.

So the only honest starting point is to go and read what it currently says about you. That is an hour with a notepad, it costs nothing, and almost nobody does it. Here is how we run it, before the autumn schedule fills up.

Write the prompts a customer would actually type

Fifteen is the right number. Enough to show a pattern, few enough to finish in one sitting.

Not “best dentist”. The sentences people actually write, with a constraint in them. You want some of each of these four kinds:

1. The need with a constraint. “Who can do a same day crown in [your town] and takes evening appointments.” “Which HVAC company in [your town] does emergency weekend calls.” The constraint is the part that decides whether you are named, because it is the part most businesses have never published an answer to.
2. The comparison. “Best rated family dentist near [the neighbourhood people say] for someone nervous about the dentist.” These lean on review language rather than on listing fields, and they tend to name a different set of businesses.
3. The direct question about you. “Is [your business] taking new patients.” “Does [your business] take [plan name].” Being asked about by name is the most common way you appear at all, and the most common way a wrong fact reaches somebody.
4. The question the front desk gets all day. Asked the way a stranger would ask it, not the way your industry writes it.

Write all fifteen down before you run any of them. Once answers start coming back the temptation is to keep rephrasing until you like what you see, and that is not a sample, it is a search for reassurance.

Run the same set through two engines

Engines read different things, so one of them is not a measurement. Run the identical fifteen through two, on the same day, in a signed-out window if you can manage it, on a phone rather than a desktop because that is where the question gets asked.

Do not help. No follow-up prompts, no corrections, no “actually we do offer that”. You are recording what a stranger gets, and a stranger gets one go.

Log two columns, then a third for what is wrong

For each prompt write down who was named, in order, and what the answer leaned on: the links it offered, the sources it credited, the phrases it clearly lifted. Then a third column for every fact about you that came back wrong.

The citation column is the one that changes minds. In the checks we have run this year across local categories, what gets quoted is rarely the business’s own website:

• Apple Maps, which keeps its own record of you through Apple Business Connect and feeds assistants that are not Google’s.
• Insurance and plan directories, licensing boards, trade association member lists and supplier locators. Tiny traffic, high trust, and often years out of date.
• Neighbourhood forums and Reddit threads where somebody asked for a recommendation and got six replies. Google has had a data partnership with Reddit since early 2024 and those threads have been visible in results ever since.
• A “best of” list from a regional publication, treated as an independent source even when it was mostly advertising.

Most of that list is not yours, which is the useful and unwelcome discovery.

Read the variance before you change anything

Answers move between runs. Run the same fifteen again a week later and compare, because a single run will tell you something confident and half of it will not repeat.

What repeats is worth acting on:

• A competitor named every time on a prompt you assumed was yours. That is a description problem, not a ranking one.
• The same source quoted across several prompts. That record is doing the work, whether or not you have ever logged into it.
• A wrong fact that survives the rerun. One appearance is noise. Twice is the answer of record.

What varies is not worth a meeting. Resist charting any of it. This is a dated sample you repeat, not a number to watch weekly, and anything sold to you as a score is somebody else charting their own sample.

Where it goes wrong

The failures we find in the third column are almost never on the business’s own site. They sit in records somebody else maintains:

• A procedure the practice stopped offering two years ago, still on an insurance directory, still being recommended to people who arrive disappointed.
• An association or supplier page carrying the address from a move nobody told them about.
• A phone number from the previous owner, or a tracking number that appears on one listing and nowhere else, so the business looks like two businesses.
• An aggregator record nobody has ever claimed, feeding the same wrong suite number to everything downstream of it.
• Apple Maps showing hours you changed last year, because the listing was never claimed.

Each one is a ten minute fix, and each one is being repeated to customers in a confident sentence right now.

What to do this month

• Write your fifteen prompts, run them through two engines, and log who is named and what is cited. Keep the page; you will repeat this in the new year.
• Rerun the same set a week later and act only on what appears both times.
• Claim or check your Apple Maps listing if you have not looked since you opened.
• Pick the two third-party records you have never logged into, claim them, and fix the hours and the service list.
• Read the last thirty reviews and pull out the phrases customers repeat. If the words that matter to your business are missing, change what you ask for.
• Give the service that makes you the most money a proper page this month: what it costs and what changes the cost, the process, the timescale, and photographs of your own work.

The uncomfortable part of this exercise is how little of what gets quoted is yours to edit. The business profile is the one record you can correct this afternoon. The rest are other people’s copies of your facts, kept by directories and aggregators that have never met you, and the work is finding out who is holding an old one. Put the same fifteen prompts in the diary for January. Most of what you fix in between will be ten minutes of local search housekeeping, and the reviews you collect in the meantime supply the words the next answer quotes.