Measurement
AI Visibility Tracking with Clear Definitions
Set up AI visibility tracking with defined questions, sampling notes, and separate mention, citation, referral, and inquiry metrics for clearer reviews.
TL;DR
- Fix the question, market, platform and sampling rules before comparing results.
- Count brand mentions, explicit proposed changes and owned citations separately.
- Show completion and failures beside usable-answer rates; a sample is not the whole market.
AI visibility tracking checks how a brand appears in a defined set of AI answers. The value is a repeatable record: what was asked, what appeared and which source was used. A tool score without those details is hard to act on.
Define the metric before filling the dashboard
Pew’s March 2025 browser study found traditional-link click rates of 8% with an AI summary and 15% without one; summary-source links received clicks on 1% of summary visits. The panel covered 900 US adults. These are click measures, not brand-visibility rates. Google’s AI guidance covers a different question: platform eligibility.
Worked answer-sample contract
Illustrative setup: 10 fixed questions × 2 platforms × 2 planned runs = 40 observations. Keep market, language and account state in the register. Save the answer text or screenshot and sources. Four runs fail, leaving 36 usable answers; 12 mention the brand, 4 explicitly recommend it and 6 link an owned page.
| Metric | Numerator / denominator | Result and rule |
|---|---|---|
| Completion | 36 usable / 40 planned | 90%; preserve the four failed runs and their reasons. |
| Brand mention | 12 answers naming the brand / 36 usable | 33.3%; count each answer once. |
| Recommendation | 4 answers explicitly recommending the brand / 36 usable | 11.1%; a neutral list entry does not qualify. |
| Owned citation | 6 answers linking an owned page / 36 usable | 16.7%; count once even if two owned URLs appear. |
| Unknown outcome | Failed or unreadable runs | Exclude from usable-answer rates; never count as a negative answer. |
These measures overlap. Do not add 33.3%, 11.1% and 16.7% into a combined score. Equal-run weighting is part of this example, not a claim that each question has equal demand. If one platform accounts for all failures, show its missing coverage and avoid a clean cross-platform comparison.
The input material is the saved run register. The gap is whether a change comes from answers or from missing coverage. The output is the table plus raw observations. Accept the report when another analyst can reproduce each count. For source-level checks, use AI citation tracking.
Define what is being observed
Name the market, language, questions and platforms in the sample. Save the date and any conditions that may change the answer, such as location, account state or prompt format.
Keep two groups clear. A non-branded question asks about a problem, use case or choice without naming you. A branded check asks about your firm or products. Keep a stable core set and label new test questions on their own. Adding questions changes what a rate means.
Keep each measure separate
Define the terms before you count:
- A mention means the brand appears in the answer.
- A citation means the answer names or links a source.
- An attributable referral means a site visit with enough data to trace an AI source.
- A qualified inquiry is a sales request that meets the fit rules agreed with sales.
For a rate, show what you counted and what you divided it by. Mark missing runs as missing, not as answers that excluded your brand. Use citation tracking to check the linked pages.
Review changes with context
Log what changed on the site and when it went live. Note other events too: a launch, ad campaign, press coverage or a new counting rule may affect the result.
Say what changed in the sample. One new mention does not show that a page caused the change or that all users will see it. An AI visibility audit can help set the first question set and work plan.
Connect visibility with the business journey
Where the data allows, check known AI referrals and what those visitors do. Keep unknown sources marked as unknown. A buyer can use several channels before making contact.
Use the AI lead guide to follow visits into sales requests. Keep those stages linked where there is evidence, without rolling them into one score.
Frequently asked questions
Is one prompt a reliable baseline?
It is one observation, not a representative baseline. Agree on a question set and record the sampling conditions before comparing periods.
Can visibility prove revenue impact?
Not by itself. Review attributable visits, qualified inquiries, and available sales context separately, with the limits of the data explained.
Should we rerun a failed question?
Follow a rule agreed in advance. Log the retry and keep the failed attempt; do not rerun only the answers you dislike.
Is a weighted visibility score useful?
It can be, if weights and reasons are published beside the unweighted counts. Do not label an editorial weight as search demand without evidence.
Share the question sample and reporting decision you need to make. We can scope a repeatable baseline.
