Tools that track whether AI recommends your brand

Part 01 — What actually decides whether a tracker is useful.

Six things, each of which we learned by getting them wrong or by measuring them across 800+ scored answers in our published studies.

1. Which engines, and how Google AI Overviews is obtained. Engines disagree far more than people expect. In our Taiwan study, Whoscall was recommended by Google AI Overviews on 5 of 6 questions and by ChatGPT on 0 of 6 — same brand, same week, same questions. A tracker covering only chat engines will tell you a confident, wrong story. Note that Google AI Overviews has no public API: it has to come from location-targeted search data, so ask any vendor how they get it and for which country.

2. Multi-run consensus, not a single pass. Across three separate studies we measured roughly 15% of answers flipping between identical runs. If a tool asks each question once, it cannot tell you whether a change is your work or noise — and neither can you.

3. Citation tracking. This is the difference between a score and a to-do list. Our retrieval study found answers built from cited sources named the local brand 68% of the time versus 25% for answers written from memory — the sources are the mechanism. And they are often not yours: in the travel category, the most-cited source was klook.com, 15 times, against 2 for the client's own domain.

4. Per-question granularity. Averages hide the question you lose. McDonald's Taiwan ranks first in 15 of the 18 answers that name it — and is absent from all four engines on "which Taipei fast food has the best value", because the engines read that phrasing as bento shops. An aggregate score would have shown a healthy number and hidden the entire problem.

5. History, and what changed. A number without a baseline is not a scoreboard. What you need weekly is the delta: which question × engine flipped, which source started naming you, which one stopped.

6. Language and market. If your buyers ask in Traditional Chinese, the questions must be asked in Traditional Chinese — translated questions measure a different market. Same for the search location behind Google AI Overviews.

Part 02 — The tools buyers actually ask about.

These are the products that came up when we ran the category's own buyer questions. We list what we could verify from each vendor's public site on Sep 8, 2026 — engines claimed, whether pricing is public, and who it is aimed at. We do not rank them, and we have not run their products; where a vendor does not publish pricing, that is what we say.

Two honest notes. Classic SEO suites — Semrush, SE Ranking, Ahrefs — now ship AI-visibility features too, and in our measurement AI recommends them for these questions more often than it recommends any specialist tool. And this table describes claims, not audited behaviour: engine counts come from vendors' own marketing, including ours.

Part 03 — What the category looks like when you measure it.

In August we ran 16 category buyer questions through four engines — 186 scored answers — and published the result, including the row that embarrasses us.

We re-ran our own scan on Sep 8 — six category questions, four engines, 24 checks. Ainswer: still 0. Ten days, four published studies, and the number has not moved. We are publishing that because a tracker that hides its own number is not worth buying, and because it is the clearest possible demonstration of the thing we measure: being good at something does not put you in the answer; being in the sources does.

The sources feeding these answers, in our latest scan, were competitors' own blog posts — pages titled "best ChatGPT tracking tools", "best Google AI Overviews trackers" — plus Reddit, LinkedIn and YouTube. That is exactly the pattern our category studies keep finding, applied to us.

Part 04 — How to test any vendor in ten minutes.

You do not need to trust a comparison table, including this one. Ask three questions:

"Run my five buyer questions and show me the raw answers." Not a score — the actual text each engine returned, with the citations. If a tool cannot show you the answer behind the number, the number cannot be checked.

"How many times do you ask each question, and what do you do when runs disagree?" The honest answers are a number greater than one and a consensus rule. "Once" means the tool cannot separate your progress from drift.

"Which sources fed each answer, and which of them never mention me?" That list is the work. A tool that gives you a percentage but not the door list has handed you a problem without a next step.