// guide
How do you measure AI visibility and citations?
Updated: 2026-07-31
// short answer
Measure AI visibility with four inputs: citation rate across a fixed prompt set sampled repeatedly, share of voice against competitors on those same prompts, AI-crawler hits in server logs as the earliest signal, and referral traffic from assistant domains as the commercial outcome. Single checks are unreliable because answers vary between runs.
Why rank tracking does not translate
A ranked list is stable enough that checking position once a day means something. A generated answer is produced fresh each time, varies with small changes in phrasing, and differs by user context and region. Checking once tells you what happened once.
The unit of measurement has to change from position to frequency: out of N samples of a given prompt, how often were you named.
Build the prompt set first
Everything else depends on having a fixed list of questions that matter commercially. Twenty to fifty is usually enough to start, drawn from what buyers actually ask rather than from keyword tools.
Include the three shapes separately, because they behave differently: category questions where you want to be one of the names, comparison questions where a competitor is already named, and branded questions where the assistant describes you and can get it wrong.
- Category: 'best GEO agency for B2B SaaS'
- Comparison: 'X vs Y for AI visibility'
- Branded: 'what does GET-GEO.AI do'
- Problem-led: 'why is my brand not cited by ChatGPT'
The four metrics worth tracking
Citation rate is the share of samples in which your domain is cited or your brand named. Share of voice is the same measurement expressed against the competitors appearing alongside you, which is what tells you whether a flat number is good or bad.
Crawler activity in server logs is the leading indicator: GPTBot, OAI-SearchBot, PerplexityBot and ClaudeBot hits confirm new content has been fetched, usually weeks before any citation appears. Referral traffic from assistant domains is the lagging indicator and the one that connects to revenue.
Sanity-check the referral data
Assistant referrals are systematically undercounted. Some assistants strip the referrer, some route through redirectors, and a user who reads an answer and later types your domain directly registers as direct traffic with no attribution at all.
Treat the referral number as a floor rather than a measurement, and watch its trend rather than its absolute value. Brand-name search volume rising alongside AI activity is often the clearer signal that assistants are doing their job.
Sampling discipline
Run each prompt several times rather than once, in a clean session, and record which sources were cited rather than only whether you appeared. The competitor list is where most of the useful information sits.
Keep the prompt wording frozen between rounds. Changing the phrasing changes the retrieval, and you lose the ability to compare periods.
| Metric | Source | Signal type | Reliability |
|---|---|---|---|
| Citation rate | Sampled prompt runs | Direct outcome | Good, if sampled enough |
| Share of voice | Same runs, competitor set | Competitive position | Good |
| AI crawler hits | Server logs | Leading indicator | High — logs do not lie |
| Assistant referrals | Analytics | Lagging, commercial | Undercounted; use the trend |
| Branded search volume | Search Console | Indirect demand | Noisy but useful |
Related questions
How many samples per prompt are enough?
Enough that a single unusual answer does not move the number. Three to five runs per prompt per round is a workable starting point; the right figure depends on how much the answers vary for your category.
Can I see AI citations in Google Search Console?
Not as a separate breakdown. Search Console does not currently isolate AI Overview appearances from other impressions, so overview presence has to be observed directly.
Do I need a paid AI visibility tool?
Not to begin with. A spreadsheet, a fixed prompt list and a disciplined weekly routine produce a defensible baseline. Tools mainly buy scale and scheduling once the prompt set grows.
What does a good citation rate look like?
There is no universal benchmark — it depends on category competitiveness and how established the alternatives are. The meaningful comparison is your own trend over time and your share against the specific competitors appearing in the same answers.