Guides
How to track AI visibility (without a fake score)
AI visibility is measurable, but not as one number. This guide sets up five measurements you can defend — access, AI Overview presence, citation rate, AI referrals and branded demand — and explains what each one cannot tell you.
By the Visdenda research team · Reviewed by the Visdenda editorial team
Published 2026-11-12 · Updated 2026-11-13 · 8 min read
Define AI visibility as measurements, not a score
A composite AI visibility score is easy to sell and hard to defend: when it moves, nobody can say which input moved. Replace it with five measurements that have separate sources and separate windows:
- Crawler access — can an AI agent fetch the page at all?
- AI Overview presence — did Google show an AI answer for the tracked query?
- Citation rate — did that answer reference your domain?
- AI referrals — did a visit arrive from an AI surface, and to which page?
- Branded demand — are people searching for you more often?
The first three live on the AI Overview tracker; referrals and branded demand are separate reports. None of them needs the others to be valid.
What is an AI visibility score — and when is it fake?
A visibility score is a composite: a vendor picks signals — mentions, citations, sentiment, position — and collapses them into one number. That number is not measured by an engine; it is a weighting decision made by the vendor. It can be a useful first read, but it cannot tell you which input moved.
It becomes fake when the inputs are not disclosed: a simulated answer counted as a ChatGPT answer, a sentiment classifier presented as the engine's opinion, a mention total with no sample size, or a share of voice with no denominator. A number nobody can reproduce from raw answers is a marketing asset, not a measurement.
A defensible score states the prompt set, run count, date window, locale and model version, and links each input to a raw answer. Even then, compare it only with itself over time. For decisions, use the five separate measurements above instead of the composite.
Step 1: check the access layer
Before measuring anything, confirm that AI crawlers can fetch the pages you want cited. Run the free AEO checker: it tests robots.txt, llms.txt, content signals, bot identity and agent discovery. A blocked path makes every later metric ambiguous — you cannot tell whether an answer omitted you or never had the page.
Re-run the check after every deploy. Most access regressions are configuration changes, not decisions by the assistant.
Step 2: baseline AI Overviews
Pick the queries where an AI answer would plausibly cite you: category terms, comparisons and alternatives. Then sample them. A sample records two facts — whether an AI Overview appeared, and whether it referenced your domain — plus the cited URLs as evidence.
Keep the sample size and the locale fixed, or the trend is noise. The console samples the US/English locale; paid sites rotate five tracked terms a day, ranked by recent Search Console impressions, and manual checks run on demand.
How to track AI Overviews rankings over time
AI Overviews do not expose a rank grid. What you can track is a fixed sample: the same queries, locale and device checked on a schedule, with presence and citation recorded per run. If the order of cited sources matters to you, keep it, but treat that order as presentation rather than a stable position.
The comparable unit is a rate: runs where an AI Overview appeared, runs where it cited your domain, and the share of cited sources that were yours. Store the raw answer or the cited URL list as evidence. A day-over-day change is usually sampling noise; compare full weeks or months with the same query set.
When the sample changes, version it. Add queries by appending to the set and record the date; never swap queries silently, or the trend line compares two different measurements. The AI Overview tracker keeps presence and citation together, so a missing answer stays distinguishable from a missing citation.
Step 3: record AI referrals
Citations are the supply side; referrals are the demand side actually reaching your site. Install the probe or page beacon so visits from AI surfaces land in AI traffic analytics with their landing page.
Read the number as a floor. Referrers get stripped by apps and privacy browsers, and GA4 can threshold small segments. First-party records are the closest thing to a reliable count, and even they cannot see a visit with no referrer.
Step 4: read branded demand as context
Branded queries in Search Console show that people are looking for you. They do not show why. Use them as a lagging signal next to the citation and referral data, and resist the temptation to credit a single answer for a lift in brand searches.
If the branded demand moves without any citation change, the cause is probably elsewhere: a launch, a mention in traditional media, or seasonality.
Step 5: review monthly with fixed windows
Compare the same window length, the same query set and the same locale. When a number moves, write down what changed in the sample before writing down a cause. Then look at the crawler report: a new agent or a blocked path often explains a change that looks like an answer-engine decision.
The AI crawler report gives that side of the story: hits by agent, status, machine-file reads and 404s.
What to ignore
- Share of Model or share-of-voice percentages without a stated denominator.
- Per-brand sentiment scores sold as if an engine published them.
- Assistant rank trackers that report a position inside ChatGPT or Perplexity; those products do not expose one.
- A single screenshot of an answer presented as a trend.
A one-page checklist
- Access check passes for the crawlers you care about.
- 10–30 tracked queries with a fixed locale and sample size.
- Presence and citation recorded, with cited URLs kept.
- Probe or beacon installed; AI referrals read per landing page.
- Branded queries reviewed as context, never as attribution.
- Monthly review with the same windows and a written change log.
Frequently asked questions
How to track AI Overviews rankings over time?
Run the same query set, locale and device on a schedule and record three facts per run: whether an AI Overview appeared, whether it cited your domain, and which URLs it cited. Compare rates over full weeks or months, not day to day. If you add queries, append them and record the date so the trend stays comparable.
What is an AI visibility score — and when is it fake?
A score is a vendor's weighting of signals such as mentions, citations and sentiment, not a number an engine publishes. It is fake when the prompt set, run count or denominator is hidden, or when a simulated answer is presented as a real one. A defensible score links each input back to a raw answer.
How often should I check AI visibility?
Weekly is enough for access checks after deploys; AI Overview samples and referrals are better reviewed monthly with fixed windows. Daily sampling adds noise faster than signal unless you are watching a launch. Re-baseline when the query set or locale changes.
Can I track AI visibility without a paid tool?
You can cover the access layer for free with a crawler or AEO checker, and Search Console gives you branded demand. Citation sampling needs a SERP source, and referral recording needs a first-party beacon. The free route works for a first baseline, not for a fixed ongoing sample.
Is AI visibility the same as AI traffic?
No. Visibility is whether an assistant can reach, cite or mention you; traffic is whether someone clicked through. A citation can produce no visit, and a visit can arrive without a recorded citation. Keep them as separate columns with separate limits.