What does AI visibility actually measure?
Four metrics, each answering one question: is there a mention of you in the answer (mention rate), is the web page the answer cites yours (citation rate), what position are you in (recommendation position), and did anyone actually click through (AI referral traffic). The first three require you to sample AI yourself; the fourth comes from your analytics tool.
| Metric | Question it answers | How to calculate | Where to get it |
|---|---|---|---|
| Mention rate | Does your brand name appear in the answer | Times you are mentioned ÷ total questions asked | Sample by asking yourself, and record each one |
| Citation rate | Is your site among the web pages AI cites | Answers citing your URL ÷ total questions asked | The sources panel or citation links in the answer |
| Recommendation position | Where you rank in the recommended list | Your rank each time you are mentioned; take the average or look at the distribution | The order in the answer text |
| AI referral traffic | How many visits come from AI answers | Number of sessions from sources such as chatgpt.com | GA4 source / medium |
How do you sample mention rate and citation rate?
The same question can get a different answer from AI each time. In the SparkToro and Gumshoe study of 2026-01, 12 prompts were each asked 60–100 times per platform, and the recommended brands were not stable. So a single answer tells you little; ask many times and look at how often you appear.
1. Write 10 questions buyers would ask
Don't write “is my product good”. Write what buyers ask when they don't know you yet: “what tools are there for X”, “which is better, X or Y”, “X for scenario Z”.
2. Ask each question several times on each platform
Pick at least ChatGPT (with web search on) and Perplexity as platforms. Start a new conversation each time to avoid context influencing the answer.
3. Record one row for each answer
Record whether you are mentioned, whether your URL is cited, what position you are in, and who is recommended most. A template is below.
4. Compare only the same set of questions each time
If you change the questions, you cannot compare with last time.
5. Repeat at a fixed interval
Every time you change a page or get a new review, test again with the same set of questions.
What does a record table you can follow look like?
| Date | Question | Platform | You mentioned | Your URL cited | Your rank | Most recommended |
|---|---|---|---|---|---|---|
| (fill in) | (fill in) | ChatGPT / Perplexity | Yes / No | Yes / No | A number; write a dash (-) if none | (fill in) |
| (fill in) | (fill in) | ChatGPT / Perplexity | Yes / No | Yes / No | A number; write a dash (-) if none | (fill in) |
After each question has been asked N times, add one summary row: mention rate = number of rows where “You mentioned” is yes ÷ N; citation rate = number of rows where “Your URL cited” is yes ÷ N.
A real summary example
2026-10-26 review of GammaSlide: 21 questions asked, GammaSlide was mentioned 0 times, mention rate 0. AI read 28 cited web pages, of which 19 listed similar products and 0 mentioned GammaSlide; citation rate is also 0. This sample is only 21 runs, enough to show “not appearing at all” but not enough to compare changes over two months in fine detail.
How do you read AI referral traffic?
In GA4, look at the session source / medium in the Traffic acquisition report. OpenAI's Help Center states that ChatGPT search links carry utm_source=chatgpt.com, so the source shows up as chatgpt.com. For the setup steps, see How do you see ChatGPT traffic in GA4?.
Example: MochiStamp (the operator's own site), sessions over the seven days from 2026-10-29 to 2026-11-04, by source:
| Source / medium | Sessions |
|---|---|
| google / organic | 53 |
| chatgpt.com / ai-assistant | 44 |
| (direct) / (none) | 35 |
| openai / organic | 4 |
| bing / organic | 2 |
Of the 44 sessions from chatgpt.com, 32 were counted by GA4 as “engaged sessions”. The days are uneven: at least 2, at most 13, and no records on 10-31. This is only seven days and cannot be taken as a long-term trend.
Note: some visits from AI carry no source information and are counted under (direct) / (none), so AI referral traffic is only a lower bound. This is an inference; we have no data showing how many of the 35 direct visits came from AI.
How do you read the four metrics together?
| Combination you see | What it may mean | Next |
|---|---|---|
| Low mention rate, low citation rate | You are not among the pages AI reads | First get into the cited rankings and reviews (see the guide “What signals does AI actually look at when recommending brands?”) |
| High mention rate, low AI referral traffic | You are mentioned, but no link is given or nobody clicks | Check whether your URL appears in the answer, and whether the first screen of your official site is clear |
| AI referral traffic exists, mention rate never measured | People are using it, but you don't know for which questions | Run the sampling to find which questions bring it |
For the technical prerequisites (whether crawlers are blocked, whether the page serves its body text directly), see AI visibility audit checklist.
FAQ
Is AI visibility the same thing as SEO ranking?
No. In Ahrefs' 2025-08 study, only 12% of cited URLs ranked in Google's top 10 for the original prompt. Ranking is a useful reference, but measure mentions and citations in AI answers separately.
How many times should each question be asked?
There is no standard answer. SparkToro's study asked 60–100 times per platform. More runs are steadier; with few runs, look only at the trend and don't make fine comparisons.
How do I know whether ChatGPT is sending me traffic?
In GA4, check whether the session sources include chatgpt.com. ChatGPT search links carry utm_source=chatgpt.com, and GA4 also has an AI Assistant channel.
Why is AI referral traffic only a lower bound?
Inference: some visits from AI carry no source information and get counted as direct visits. We have no data to determine the proportion.
How often should you test?
Each time you change a page or get a new review, test again with the same set of questions, and keep the interval fixed so results are easy to compare.
Sources
- SparkToro × Gumshoe · AIs are highly inconsistent when recommending brands or products （2026-01, 12 prompts run 60–100 times on each platform）
- Ahrefs · Only 12% of AI Cited URLs Rank in Google’s Top 10 for the Original Prompt （published 2025-08-11, 15,000 prompts）
- OpenAI · Help Center · ChatGPT search （Checked 2026-11 (original would not open; excerpted from a mirror)）
- MochiStamp · MochiStamp 的 GA4 会话，2026-10-29 至 2026-11-04 （The operator’s own site, 2026-11-05）
- kagentx evaluations: the data on the case study pages; each example gives its date.