How to track whether AI mentions your company (a 20-minute setup)
· by boneLau
Short answer: you cannot track this by asking ChatGPT once. The same prompt can return a different answer on the same day — in our own runs, mention rate swings roughly ±10 percentage points between identical runs. What works is fixing three things (the questions, the engine list, the interval), recording the same five fields every time, and comparing against a baseline instead of against last week. The whole setup below takes about 20 minutes.
1. Why a single check tells you nothing
Searching "does ChatGPT mention my brand?" once gives you an anecdote. Three problems with it:
- It changes. We run the same five questions against the same assistants every week. On one brand, two runs on the same day produced mention rates of 63% and 79%. Nothing had changed on the brand's side.
- You will not remember it. By next week the screenshot is buried, and you cannot tell whether the answer changed or you are remembering a different question.
- It has no denominator. "ChatGPT didn't mention me" on one question tells you nothing about whether you are missing from 10% or 90% of the answers that matter.
Anecdotes are not useless — they are how you find a question worth tracking. They are just not a measurement.
2. Fix these three things, or the numbers mean nothing
a) The questions
Pick five questions your buyers actually type, not your taglines. A workable split:
| Type | Example | Why it matters |
|---|---|---|
| Discovery (×3) | "best {category} for {use case}" | where new buyers start |
| Comparison (×1) | "{your brand} vs {competitor}" | where deals are decided |
| Objection (×1) | "is {your brand} any good" | where doubt gets resolved |
Five is deliberate: fewer and you cannot tell a trend, more and you will stop maintaining it.
b) The engine list
Pick at least four: ChatGPT, Perplexity, Gemini, Google AI Overviews.
Not because four is magic, but because they retrieve from different evidence. In our data, Perplexity cites a brand's own site heavily while ChatGPT leans on third-party pages. If you average them into one number, you lose the only part that tells you what to do.
c) The interval
Weekly, on the same day. Not daily: retrieval indexes do not move that fast, and daily runs just multiply noise. Not monthly: you will miss the four-to-eight week window where a fix either worked or didn't.
3. The 20-minute setup
One table, five columns:
| Question | Engine | Mentioned? | Position | Sources cited |
|---|---|---|---|---|
| best AI visibility tools for SaaS | ChatGPT | no | — | g2.com, reddit.com, alternativeto.net |
| best AI visibility tools for SaaS | Perplexity | yes | 4 | (own site) |
Steps:
- Create a table — Notion, a spreadsheet, anything with rows.
- Run each question against each engine. Five questions × four engines = 20 rows.
- For each row record: mentioned yes/no, position if mentioned, and every source URL the answer cites.
- Paste the raw answer into a notes column. This is the step almost everyone skips, and it is the one you will need in three weeks when you are trying to tell a real change from a remembered one.
That is the entire setup. You now have a baseline.
4. How to tell a real change from noise
Three rules, in order of usefulness:
Rule 1 — anything inside ±10pp is not a story. That is the natural run-to-run spread we measured on mention rate. Week 1 at 60% and week 2 at 55% is not a decline.
Rule 2 — require two rounds pointing the same way. One run moving up is a data point. Two consecutive runs moving the same direction is a signal. This single rule has saved us from shipping "fixes" for noise more than once.
Rule 3 — watch the cited sources before the mention rate. Sources are upstream of mentions: if the set of domains the engine cites starts including a site that describes you, the mention tends to follow. If the cited domains are unchanged, your mention rate usually will be too.
5. When a spreadsheet stops being enough
Be honest about the ceiling:
| Brands you track | What works |
|---|---|
| 1 | A spreadsheet is fine — 20 rows a week |
| 2–3 | Starts to drag; you will skip weeks |
| 4+ | You will stop doing it, and skipped weeks are worse than no weeks |
The arithmetic is what kills it: five questions × four engines × four brands is 80 rows every week, plus reading each answer to judge position and sources. Most people manage two or three weeks before the log goes stale.
That is the point at which automating it stops being a luxury — and, in full disclosure, it is why we built BrandScope. Not because a spreadsheet is wrong, but because a spreadsheet you stop updating is.
FAQ
How long before a fix shows up? Usually 4–8 weeks. A new third-party source has to be crawled, evaluated and then retrieved for the right question. Anyone promising days is measuring something else.
Does incognito mode change the answers? Not meaningfully. Search results are personalised, but the assistants' answers to category questions are not usually tied to your cookies. Use a fresh session for consistency, not because it unlocks a different truth.
Do I need to pay for anything to do this? No. The five-question, four-engine setup above runs on the free tiers. Paid plans matter when you want the same questions run automatically every week, or more than a couple of brands.
Should I track the Chinese assistants (Qwen, GLM, Kimi)? Only if your buyers use them. For most Western B2B products they return nothing — in our runs, one brand had zero mentions across two of the three. Track them if your market is there; otherwise they are extra rows of nothing.
Can I get the spreadsheet template? The five columns above are the template. We deliberately do not gate a download behind an email — the structure is the whole value, and copying it into your own tool takes a minute.
BrandScope runs this exact setup automatically every week across the four core answer surfaces, verifies each citation against your own domain, and reports what changed. You can [run one brand free](https://www.brandscope.dev/) and see the current state before paying for anything.