The window, pane by pane
One pane is one answer. Each row is one AI tool, each column one time we asked the same question that week. A pane lights when the answer named the product. Kettlebook’s week, for “What’s the best CRM for a small sales team?”
Kettlebook was named in 6 of 20 answers. Likely 3 to 10.
Sample. Invented products.Checked Mon 2 Nov 2026
A few weeks later
One week can bounce. So your report shows six weeks side by side, next to your top rival.
| Checked | Count and likely range, drawn from 0 to 20 | Named | Likely |
|---|---|---|---|
| 2 of 20 | 1 to 5 | ||
| 2 of 20 | 1 to 5 | ||
| 3 of 20 | 1 to 6 | ||
| 4 of 20 | 2 to 8 | ||
| 5 of 20 | 3 to 9 | ||
| 6 of 20 | 3 to 10 |
Kettlebook: 2 to 6 of 20 in six weeks. Paperwren: 14 of 20, likely 10 to 17. The band is the likely range; the dot is the count.
Sample. Invented products. Checked each Monday, 28 Sep to 2 Nov 2026.
Which questions do we ask?
The ones a buyer would ask before they’ve heard of you. We write them from your category and rivals, and you can change any of them: 25 on Starter, 75 on Growth. Three kinds:
- The best tool for a job
- “What’s the best CRM for a small sales team?”
- A way out of a rival
- “What are good alternatives to Paperwren?”
- A specific need
- “Which CRM has the best email tracking for a five-person team?”
Sample. Invented products.
We don’t put your product’s name in the questions we count.
Which AI tools, and how?
We aim to ask the way a new buyer would: logged out, with no history, from the country you sell in.
| Tool | How we aim to ask | What we record |
|---|---|---|
| ChatGPT | Logged out, web search on, no history | The answer, who it names, the pages it shows |
| Gemini | Logged out, no history | The answer and who it names |
| Perplexity | Logged out, default settings | The answer, who it names, its numbered sources |
| Google AI Overviews | A normal Google search, logged out | The AI summary when one appears, and the pages it links |
| Google search | Logged out, on a phone and a laptop | Your position, and who is above you |
Tools change without notice: Google made Gemini 3 the default model behind AI Overviews in late January 2026.
Source: DesignRush, reporting Ahrefs, April 2026. Checked 6 Nov 2026.
How many times?
Each question goes to each AI tool [N] times a week, on different days and at different times. We check your position on Google [N] times a week.
One answer is a snapshot, so the free check, which asks each tool once, says so. The samples here use one question, asked 5 times in each of 4 tools: 20 answers. To try it by hand: How to find out if ChatGPT recommends your product.
What counts as named?
When the answer recommends your product by name, by domain, or by a variant that can only mean you. When we’re not sure, the pane stays dark.
| In the answer | Counted as | Why |
|---|---|---|
| “Kettlebook is a good fit for sales teams of two to ten.” | Named | The product’s name, recommended. |
| “Small teams like kettlebook.app for its quick setup.” | Named | Your domain, or a clear variant of the name. |
| Only in the list of sources under the answer | Not named | We record it separately, as linked. |
| “Kettlebook was shut down in 2024…” | Not named | It’s wrong, so we flag it, and fixing it becomes a job. |
| “Avoid Kettlebook if you need custom objects.” | Not named | A warning isn’t a recommendation. We flag it so you see it. |
| “Kettle CRM, for field sales” | Not named | A different product. |
Sample. Invented products.
We note where in the answer you came, but never turn it into a position: AI answers have no fixed order.
The likely range, and why answers vary
In a study of about 3,000 answers to the same 12 questions, the same list of brands came back less than once in 100 tries. Some brands still showed up in 60% to 90% of answers.
Source: Search Engine Land, reporting SparkToro and Gumshoe. Checked 6 Nov 2026.
So we give a count with a range: named in 6 of 20, likely 3 to 10. We aim to use a 90% Wilson score interval: ranges worked out this way hold the true rate about 9 times in 10. More answers, narrower range.
| Answers | Range, drawn from 0% to 60% | Likely range |
|---|---|---|
| 20 | 16% to 48% | |
| 40 | 20% to 43% | |
| 80 | 22% to 39% | |
| 160 | 24% to 36% |
What we can’t measure
- Sign-ups and revenue from an answer. Ask new sign-ups how they heard about you.
- Every question every buyer asks. We ask the ones most like theirs.
- What a logged-in buyer sees. We ask the way a new buyer would.
- Why a tool picked one product. We see what it said, not its reasons.
- AI tools we don’t check yet. Such as Microsoft Copilot, Claude and Google’s AI Mode.
Known limits
- Place is approximate.
- We set the country where a tool lets us, and your report says where it doesn’t.
- Sometimes a tool doesn’t answer.
- That run isn’t counted. Your report says how many answers each count is based on.
- Not every tool shows its sources.
- So the pages an answer drew on are sometimes incomplete.
- Twenty answers is a small sample.
- That’s why every count has a range, and why we watch weeks, not days.
Changes to this method
What changed, from which week, and whether older weeks still compare.
[Launch date]
Version 1
First published method. [Final numbers: runs per week, confidence level]
[Date]
[Version]
[What changed, from which week, and whether older weeks compare]