How we measure, and how you check us.
Share of answer is a number you can reproduce. This page explains exactly what we ask, how we score it, what we store, and the five ways to catch us if we're wrong. It does not list our prompt sets or our vendors; that part is the work you pay for.
What we ask
For each client we build a set of questions (typically 60 to 120) that real customers type into an AI assistant: discovery ("best X near me"), first-time buyers, comparisons, price and deals, and logistics like hours or delivery. Each question is tied to one location. The set is documented and handed to the client on day one, and every question in it has a button that opens it in the real apps.
Who we ask
ChatGPT and Gemini, through their official developer interfaces, with web search or search grounding switched on and the customer's city passed in, so the engine answers the way it would for someone standing in that neighborhood. Perplexity and Google AI Overviews are added per client. The model that answered, whether it used live search, and the exact time are stamped on every stored answer.
Without it, an AI model answers from memory and can invent businesses that don't exist. We found this on our own first run and turned it off for good. If an answer was ever produced without live search, the dashboard says so on that answer.
How often
Every prompt is asked three separate times per engine, every week. A single AI answer contains randomness: ask twice and the third name can change. Three samples turn one lucky or unlucky answer into a rate. The dashboard shows "named 2 of 3" rather than a yes or no, and the weekly trend is built from those rates.
How we score
Scoring is a plain text match. No model decides whether you were named; a program checks for your name and your domain. That's deliberately dumb so it's auditable.
What we store
The full text of every answer, every source URL the engine cited, the model name, whether search was used, and the timestamp. Nothing is edited or summarized. Clients can read every answer in the dashboard and download all of them as a spreadsheet at any time.
Five ways to verify us
1. Open it in the real app. Every prompt has one-tap buttons for ChatGPT, Claude, Perplexity and Google AI Mode, and a copy button for Gemini. Compare the app's answer to ours. Order and wording will differ between runs; whether you're named should agree most of the time.
2. Click the sources. Each answer lists the pages the engine read. If it says you're named and cites a review site, open the review site.
3. Re-run it live. From the dashboard, any prompt can be re-asked on the spot while you watch, with the answer shown next to the stored one.
4. Log a spot check. When you check a prompt in the real app, record what you saw (named first, named, or not named) with one tap. The dashboard keeps every check with who logged it and when, and shows the running agreement rate between the real apps and our data. We do ten of these ourselves every month for every client.
5. Watch the traffic. Our free tracker counts visits arriving from AI assistants on your own website. If we say you're being named and cited, those referrals should appear over time. They're your data, not ours.
What this can't tell you
Developer interfaces are close to the consumer apps but not identical: the apps add personal history, precise location and whatever model version is live that day. Memory is the big one. If ChatGPT knows you already run ATMs, it answers you as an ATM operator, not as a store owner shopping for one. That is why our verify buttons open ChatGPT as a temporary chat, and why comparing against a logged-in account with memory on is not a fair test of what your customers see. Google AI Overviews has no developer interface; measuring it needs a separate data provider and is added per client. And AI-referred website visits understate the real effect, because many people read the answer and simply drive to the store.
See it on your own business.
Three real questions, two engines, twenty seconds, no call.