You ask ChatGPT for the best accountant in Norwich on Monday and you're named second. A colleague asks on Tuesday and you're not there at all. Nothing about your business changed overnight. The answer did, because AI answers vary with the wording, the place, the assistant and the day.
That variation is why a single check proves nothing, and why a short, repeatable routine proves a great deal. This lesson builds the routine.
Why one check proves nothing
Four things move between one question and the next. The model generates text probabilistically, so the same question asked twice can come back with a different order or a different third name. Retrieval is fresh each time, so a new review or an updated "best in town" list can reshuffle the sources overnight. Location is a guess: when the question doesn't name a place, each assistant decides where you are from the device, the account or the connection, and two people in one office can get different answers. And your own account carries history: an assistant with memory has seen you ask about your business before, which makes your logged-in session the worst place to test from.
So treat a check like a survey rather than a screenshot. You are sampling a distribution of answers, and the useful numbers are how often you appear and what is said when you do, measured the same way each time.
Build a fixed question set
Start from the grid you built in service-plus-location keyword mapping: your main services against the towns you serve. For each pair that matters, write three phrasings.
- The plain recommendation: "best emergency plumber in Bristol".
- The need with a condition: "plumber in Bristol who can come out on a Sunday".
- The fit question: "which plumber in Bristol is good for a landlord with several flats".
Always name the town. "Near me" hands location to the assistant's guess, and you can't reproduce a guess next month. Keep the set to about a dozen questions; a routine you repeat beats a thorough one you abandon. Then freeze the wording, because once a question changes you can't tell whether the answer moved or the question did.
Run it the same way every month
The routine is dull on purpose. Same questions, same assistants, same session conditions, same order, once a month. Ask in ChatGPT, Gemini and Perplexity, and search the plain questions on Google to see whether an AI Overview or AI Mode answer appears and what it shows.
Record one row per question per assistant with five columns: whether you were named, your position if you were, the reason given in the assistant's own words, which sources it cited or linked, and which competitors were named. The reason column is the one people skip and the one that teaches most, because it tells you which of your sources did the work. "Reviewers mention same-day appointments" points at your reviews; "their site says they cover the whole county" points at a page.
Add a notes column for what changed on your side that month: new reviews, a rewritten page, a mention in a local guide. You are not proving cause; you are giving yourself a fair chance of pairing a change with an effect a month later, the same discipline reading rank movements honestly asks of rank tracking.
Monthly is the right cadence for a local business. Weekly checks mostly measure the randomness described above, and quarterly checks miss the window in which a fix could be confirmed.
Read the results honestly
After two or three rounds the sheet answers a few questions.
If you are never named and competitors are, the problem is upstream: assistants aren't finding enough about you in the sources, and the fixes are the four jobs in the previous lesson, starting with the Profile. If you are named sometimes, you are in the candidate pool, and the specifics in your reviews and pages are what move you from sometimes to usually. If you are named with the wrong facts, an old address or hours you no longer keep, find the source the assistant cited and correct it there; the assistant is repeating a listing, not inventing.
Count your share as well as your position: across all questions and assistants, the proportion of answers that name you compared with each rival is a steadier measure than whether you were first on any given day. Position moves with noise; share moves with substance.
One reading is easy to get wrong. When Google shows no AI Overview and no AI Mode answer for a local question, that isn't a loss; many local queries still have no AI surface at all, and the map pack and web results carry them. Record it as "no surface" and keep watching, because it can change.
What to take away
- AI answers vary with wording, location, assistant and day, so a single check is one sample and a monthly routine is the measurement.
- Build a fixed set of about a dozen questions that name the town, freeze the wording, and ask them the same way each month from a clean session.
- Record named, position, the reason given, sources cited and competitors named; the reason column shows which of your sources is working.
- Judge yourself on share of answers across the set rather than position in any one of them, and treat "no AI surface" on Google as a fact, not a failure.
Next
With the routine running, the next chapter turns to the mentions that feed it: Local link building shows where local links and "best in town" mentions really come from.