Measure AI visibility by repeating a fixed set of buyer questions and counting the answers that name your brand or link to your website. Start with at least 10 runs per key prompt. Keep mentions, citations and verified crawler visits separate. Each answers a different question.
The worksheet below gives you a baseline. For help connecting that evidence to a content plan, see Toimi's AI visibility services.
What should count as a mention or citation?
A mention names the brand. A citation links to its website. Record both.
A recommendation can name your company while citing a directory. That is a brand mention and a third-party citation; it is not a citation to your site. Save the directory URL too. It may explain where the answer obtained its information.
For each prompt, calculate two rates:
- Mention rate = answers naming the brand / valid answers recorded.
- Website citation rate = answers linking to your domain / valid answers recorded.
Count an answer once per metric, even if it contains several links. Keep URL-level counts in another column.
Which prompts belong in the panel?
Choose real purchase questions. Ask sales and support what buyers need to decide.
A useful panel includes comparisons, situations and fact checks. For example: “Should a small distributor buy a CRM or build one?” A branded prompt, such as “What does Company X offer?”, checks factual accuracy. Report it separately from discovery prompts.
Freeze the wording. Store a panel version. Changing a question starts a new series for that question.
How many runs should you record?
Use at least 10 repetitions for each key prompt as an initial working minimum. Spread them across your defined collection window. Larger samples can support firmer decisions.
Record the product interface your customers use. Keep API experiments in a separate dataset. Product mode, search access and account context can differ, so mixing them leaves you comparing different conditions.
Use fresh conversations. Record language, region, plan, search mode and the model label when visible. An unavailable label should remain “not shown.” Avoid guessing.
What goes in the recording sheet?
Copy these columns into a spreadsheet:
| Field | What to save |
| Panel version and prompt ID | Stable identifiers |
| Collection time | Date, time and time zone |
| Product and mode | Platform, plan, search setting |
| Context | Language, region, fresh conversation |
| Raw answer | Full text or export |
| Brand named | Yes/no and exact wording |
| Website cited | Yes/no and destination URLs |
| Other sources | All linked URLs |
| Run status | Valid answer, error or refusal |
Log failed attempts. Report their number. Predefine how you treat refusals; selectively dropping inconvenient answers inflates the rate.
How do you show uncertainty?
Write “6 of 10 answers” beside “60%.” Those are hypothetical results. A 95% Wilson interval for 6 of 10 is approximately 31%–83%, showing how uncertain a small sample remains.
This interval assumes independent trials under stable conditions. Repeated AI answers may violate those assumptions. Treat the interval as a sampling summary, not a guarantee about future answers.
A move from 6 of 10 to 7 of 10 is weak evidence by itself. Collect more runs before assigning credit to a content change. Non-overlapping intervals offer a conservative screening signal; a formal comparison also needs a statistical test suited to the collection design.
Compare each platform separately. Keep the panel conditions constant. A blended score can hide a gain on one product and a loss on another.
How do you verify AI bot visits?
A crawler visit shows access. It does not prove a citation.
Match the logged user agent and originating client IP against the provider's current published ranges. Behind a CDN, confirm that the log records the trusted client IP rather than the proxy address. Keep the list retrieval date with the result.
OpenAI documents separate roles for OAI-SearchBot, GPTBot and ChatGPT-User. Perplexity publishes crawler and user-fetcher guidance, including IP sources. Read the current lists rather than copying old addresses.
Classify unverifiable requests as unverified. Never call them genuine solely because the header says “bot.”
What should the monthly report say?
Show the prompt results, search impressions, branded search demand and qualified inquiries together. Add referral visits as supporting evidence. Use the platform reports available in your account and disclose gaps.
Keep the raw answers. Repeat the same panel after a meaningful change. The resulting record should explain what changed, where it changed and how much confidence the sample supports.