The whole method, including the parts that are awkward.
Not keywords: questions, the way a person types them into ChatGPT. Three kinds matter, and tracking only one gives you half a picture:
| Type | What it tells you |
|---|---|
| Category | "best roofing contractor in Tampa", someone who doesn't know you exist. This is where new business comes from, and where you're least likely to appear. |
| Problem | "my roof is leaking after a storm", the symptom, described before they know what service they need. Often answered without naming anyone. |
| Brand | "is Acme Roofing any good", someone checking you out. Here the risk isn't absence, it's what gets said. |
Don't have any yet? We can generate five from your name, city and category, and you edit from there.
This is the part that makes it different from rank tracking. AI answers aren't stable: ask twice, get two answers. Checking once and calling it a ranking would be wrong most of the time.
So each question is repeated (you choose one, three or five times) on each surface you turn on. A brand appearing in four of five answers is at 80%, and that number means something. "Position 3" would not.
Not three overlapping flags. One:
| Verdict | Meaning | Score |
|---|---|---|
| Cited | Your own website was used as a source for the answer | 100 |
| Mentioned | You were named, but nothing linked to you | 50 |
| Absent | You don't appear at all | 0 |
| Flagged | You appear, described negatively or warned against | 0 |
Average those across every answer collected and you get the AI Visibility Score, 0–100. It's one number, it's arbitrary in the way all such numbers are, and its only real job is to move in the right direction over time.
Anything scored negative is sent to a second, more careful model for a check before it's recorded: that's the finding you least want to get wrong in front of a client.
Every URL an answer was built from is recorded, then split into pages you own and pages you don't, with a note on whether each one names you. A directory cited fifteen times that never mentions you is a concrete job. Knowing "yelp.com was cited" is not.
Gemini hides its sources behind Google redirect links. We follow them once per run and store the real page, so its citations count properly instead of vanishing.
Name a rival and we count them in answers we've already paid for. No extra calls, no extra cost: that's the main difference from tools that re-run every prompt per competitor. We also surface brands the AI keeps raising that you aren't tracking yet.
Nothing to us. You add a DataForSEO key and they bill you directly. The driver is the web-search fee on chatbot calls, roughly three cents each, so the number of surfaces and repeats is the lever. A typical small brand runs about $2.50 a week; the exact estimate is shown before you run anything.