We do not estimate what a model would say about you. We send your customers’ questions, read the answers and count how often your name appears.
Get the free scanIn Williamsburg the best rated are Sal’s Bakery, Bedford Bread Co. and Grand St Deli. The first has the most recent reviews.
In a search engine a customer gets ten links and picks one. In a conversation with an assistant they get three names, sometimes one. There are no positions and no links, only names, so either yours is among them or somebody else’s is.
The second difference matters more: the answer changes every time it is asked. Same model, same question, two minutes later, three different names. That is why one screenshot means nothing and the only honest measure is a share across many answers. A shop owner will not do that by hand, because it means asking sixty questions and counting.
The third thing surprises people most. Your own assistant knows you. It remembers your conversations, knows your location, knows you typed your own name into it. So it names your business to you more often than to anyone else. What matters is a stranger’s assistant, and that is what we measure.
Nothing runs until you confirm the request with one click in the email.
The model looks your business up on the web and describes it in one sentence. You get that sentence in the report. If it is wrong, or if the assistant cannot write it at all, that is already the finding, because an assistant only recommends businesses it has something to read about.
Questions come from what the model established about you, not from a category typed into a form. A roofer in a small town gets questions about roof repair in that town, not about “services”. Every question must contain the town name and must never contain your business name, because such a question would guarantee a hit.
Through official APIs, three runs each, with search enabled so the model reads current sources. On the plan they go to four engines at once: ChatGPT, Claude, Perplexity and Gemini. We store the full text of every answer, the engine, the run number and the time to the second.
The stored answers go back to the model with one question: is this business named here, and where. The model understands a short form, an inflected form and a small typo, so “DachTech” counts the same as the full listing name, while the town or trade name alone does not count at all.
Whether a Google profile exists and whether it exists twice. Whether robots.txt blocks assistant crawlers. Whether your home page carries LocalBusiness structured data. Those are the three most common reasons a model has nothing to read about you.
Most tools crawl your site and estimate what an assistant would probably say about it. That is guessing, expressed as a number.
It reads your public pages, runs them through its own model and returns one figure. No source, no quote, no timestamp. A week later there is no way to check where the number came from or compare it to anything.
A question in the customer’s own words went to a real engine, and the answer was kept in full, with a run number and a time. If the report says two of sixty, you can count the lines yourself.
Because you pay for the repair, not for the measurement. To see whether the repair worked you need the same measure before and after. A simulated score moves with the tool’s own algorithm. A recorded answer never moves.
Because it is the only way to check us. The report carries full conversations with timestamps and the names your customer heard instead of yours. You can ask the same questions yourself and compare.
It is not a copy of your customer’s screen. The ChatGPT app has its own memory, personalisation and maps; the programming interface has none of them. Our number is a reliable indicator of what a stranger hears, but it is not a forecast of what you will see on your own phone.
It is not a promise either. We measure the share of answers, we do not guarantee it. We repair what can be repaired: structured data, listings, pages that answer specific questions, replies to reviews. The rest depends on the model and on what your competitors do.
And it does not work the same for every business. A manufacturer or a wholesaler has customers searching nationally rather than in a town, so questions with a town name tell you less. If that is how it turns out, we say so in the report rather than showing a zero and pretending we measured something.
Three questions, real answers, a report by email. No card, no phone call, no strings.