Someone arrives in a new town and, instead of opening Google Maps, asks ChatGPT: "Which body shop can I trust in Montecatini?". They get three or four names, each with a couple of lines of explanation. If your business is not on that list, for that customer it does not exist. And you never find out, because nobody sends you a notification.
On Google you can at least search for yourself and see where you rank. With AI assistants it is harder: the answer changes from one day to the next, from one assistant to another, and with the way the question is phrased. That is why in LocalProAI I built a check that runs on its own every Monday morning. Here is how it works, because the choices behind it are useful even if you never use it.
The right questions for each kind of business
The first mistake would be asking everyone the same question. People looking for a dentist do not ask what people looking for a pizzeria ask. So the questions are written by hand, per type of business, in the words a customer would use:
- pizzeria: "What is the best pizzeria in Pistoia?", but also "Pizzeria open on Sunday evening in Pistoia?" and "Best pizzeria with home delivery in Pistoia?";
- hair salon: "How much does a blow-dry cost in Lucca?", "Where can I get a good hair colour in Lucca?";
- garage and body shop: "Who do I call for roadside assistance in Prato?", "Where can I get hail damage on my car repaired in Prato?";
- dentist: "Emergency dentist in Livorno: who do I call?".
Today there are seven categories: pizzeria, restaurant, hair salon, gym, dentist, medical practice and garage. Each one has between four and eight questions. The questions include the town, the neighbourhood if the business has set one, the dish or service it is known for and, for medical practices, the specialty. If a piece of information is missing, the question that needs it is skipped instead of being asked halfway.
For doctors and dentists there is a note I also wrote into the code: these are questions patients ask, not advertising for the practice. The rules on healthcare advertising apply to what the practice publishes, and in fact the texts the system prepares later (I will get there) follow different rules.
Three assistants, not one
Every question goes to three assistants with web search turned on: Perplexity, OpenAI (the company behind ChatGPT) and Google's Gemini. One alone would give a partial picture: it is common for a business to show up on one and not on the others.
One thing needs saying clearly: this is not the ChatGPT app on your phone. It is the same providers queried automatically, with web search on. The answers can differ a little from what a customer sees, since they may have a chat history or phrase the question differently. The check is there to show the trend week after week, not to capture the exact answer every single person will get.
What gets recorded in each answer
For every answer the system notes four things:
- whether the business is mentioned, looking for the name and its variants: "Pizzeria Da Mario" must be recognised even when the answer just says "Da Mario";
- in which position: first on the list or fourth makes a real difference;
- in what tone: "renowned", "recommended" or "avoid", "overrated";
- who else is named, meaning the competitors the assistant suggests instead of you, and the sources it says it read.
I do not use another AI model to read the answers. These assistants almost always reply with numbered lists and names in bold, and a few handwritten rules are enough to pick them out. Having a second AI re-read every answer would triple the cost of the check with no real benefit. It is the same reasoning I described in when you don't need an LLM: the big model goes where it is needed, not everywhere.
From numbers to action
A chart going down without telling you what to do is not much use. So two and a half hours after the check, a second step looks at the last two weeks and searches for three situations:
- a question where the business never appears, on any assistant;
- a question where it appears, but on average fourth or lower;
- the same competitor named in its place on at least three different questions.
For the most serious cases the system drafts an article for the business's website and a few FAQs, written to answer exactly that question, with the business name and a concrete fact at the start of each section. That is the kind of sentence assistants like to pick up.
There are three brakes, on purpose:
- at most two actions a week per business: a few good pieces beat a website full of similar texts;
- if a question was already worked on in the last 60 days, the same article is not rewritten every Monday;
- the article stays in draft until the owner reads it and approves it from the dashboard. It only goes out on its own if the owner chose automatic publishing.
The owner gets a Telegram message explaining, in plain words, which question they were not being mentioned for and what has been prepared.
What the owner sees
The dashboard shows a summary of the last twelve weeks: how often the business was mentioned out of all questions, the average position, the week by week trend, the questions where it does best and worst, and the ten competitors named most often. For the most recent answers you can also see the sources the assistant used.
The sources are often the most useful part. If assistants keep citing a local guide, a review site or a category directory where you are missing, you know where to work. And that work needs no technology: an up-to-date listing, consistent information everywhere, real reviews.
If you want to try it by hand
Even without tools you can run a smaller version of this check:
- write five questions the way a customer would, with your town and your main service;
- ask them to ChatGPT, Perplexity and Gemini, in a private window or logged out;
- note whether you appear, in which position and who comes before you;
- look at the sources the assistant cites: those are the places where it pays to be present and accurate;
- repeat in a month with the same questions, worded the same way.
The last step is the one that matters. A single answer says little; the same question repeated over time tells you whether you are heading in the right direction.
If you run a local business and want to know what ChatGPT, Perplexity and Gemini say when your customers ask about your kind of service, we can look at it together on a free 30-minute call.
Originally published at gabrielepieretti.dev. I write about software for small businesses, Laravel, FastAPI and privacy-by-design. More here.
Top comments (0)