How to measure whether AI recommends your company
Last updated
“Does ChatGPT recommend us?” sounds like a yes-or-no question. Measured carelessly, it gives whatever answer you hoped for. This is the method we use, and you can run it by hand with a spreadsheet.
Decide what you are counting
Four different things get called “visibility”. Keep them apart:
| Measure | What it means |
|---|---|
| Named | The answer mentions your company, and means your company. |
| Recommended | It presents you as its first pick or as one of the options. An answer that only compares you, or advises against you, names you without recommending you. |
| First pick | It recommends you above the others. |
| Cited | It used a page of your website as a source. An answer can cite you without naming you, and name you without citing you. |
1. Write the questions your buyers ask, without your name
Ask an assistant about a company by name and it will talk about that company. That tells you what it believes about you, which is worth knowing, and nothing about whether a buyer who has never heard of you would be pointed to you. So write questions the way a buyer types them: the problem (“our invoices keep getting paid late”), the search for an option (“best payroll software for a restaurant”), the doubt before buying (“is it worth paying for”).
Keep one question that does name you, on its own line of the sheet. Never add it to the count.
Questions an AI wrote for you are guesses. The better ones come from sales calls, support messages and the searches in your Search Console.
2. Write down exactly what you asked, and of what
The assistant, the model, whether web search was on, the country, and the date. Two results are comparable only when all of these match. Answers through an API and answers in the app are different surfaces: the app adds the person’s history, settings and location. Label which one you measured, and don’t mix them in one number.
3. Keep every answer, whole
A count you can’t trace back to the words behind it is not evidence. Keep the full answer, the sites it cited, and whether it searched the web or answered from memory.
Here Wellread does less than this guide asks. It reads and keeps the first 3,000 characters of each answer, and marks an answer that was cut. By hand, with a few dozen answers, you can keep all of each one.
4. Read each answer for who it means
A name match is not enough. When our own audit asked what Wellread is, ChatGPT described a different company with a similar name. Matching the word alone would have counted that as us. For each answer, decide whether it names your company, how it presents you, and which other companies it names.
A company the question itself names doesn’t count for that answer. Ask “A versus B” and the answer has to mention both.
5. Count with honest denominators
- An answer that names no one doesn’t count against you. Many how-to answers recommend no company at all. Being absent from those says nothing about you.
- An answer from memory can’t cite anyone. Count “used your site as a source” only over the answers that searched the web.
- A lost answer is not a “no”. If a request failed, say so and leave it out. Don’t fold it into the misses.
- Say the denominator every time. “Named in 3 of 24 answers where buyers asked for options”, not “12% visibility”.
In our own audit of 30 questions, 12 asked for options, which gave 24 answers from two assistants. Wellread was named in 0 of them. Of the answers to the other questions, 27 searched the web and 9 came from memory. Read our own audit.
6. Expect the answers to move by themselves
Ask the same question twice and you can get two different answers. A 2026 paper on measuring visibility in AI search puts it plainly: “Answers can vary across runs, prompts, and time, making one-off observations unreliable” (Don’t Measure Once). So:
- Ask many different questions once, before you ask a few questions many times. The spread across questions tells you more.
- Treat a change of one or two answers between two rounds as nothing.
- Compare two rounds only when the questions, the assistants and the settings are the same.
- When a number changes after you changed something, you have seen that the answers changed. You have not proved what caused it.
A sheet you can copy
One row per answer. These are the main columns of Wellread’s own spreadsheet export:
| Column | What goes in it |
|---|---|
| AI | ChatGPT, Gemini |
| Question | Word for word, as asked |
| Buyer stage | Has a problem, looking for a solution, comparing, about to decide |
| Searched the web | Yes or no |
| Counted in your report | Yes, or why not: named no one, answered from memory |
| Named you | Yes or no |
| How it presented you | First pick, option, compared, warned against |
| Linked your site | Yes or no |
| Companies named | Everyone else the answer put forward |
| Sites it drew on | The sources it cited |
| Answer | The answer’s text (Wellread’s export has up to its first 3,000 characters) |
What this doesn’t tell you
- How much traffic or revenue AI sends you. That is in your analytics.
- Why the assistant chose who it chose. Asking it “why” gets you a plausible story, not a record of what happened.
- What every buyer sees. It tells you what was said to these questions, on this day, on this surface.
Wellread does this for 30 questions on ChatGPT and Gemini in a few minutes. How we measure gives our exact rules.