How AI visibility tracking works
Under the hood of an AI visibility tracker: prompt scheduling, engine execution, mention detection, and why repeated runs are essential.
AI visibility trackers turn a fuzzy question — are we showing up in AI answers — into a number you can trend. This article explains the mechanics, so you can judge whether a tracker's numbers deserve your trust.
A tracker has four moving parts. Understanding each one tells you where the errors come from.
Prompt scheduling
The tracker holds a set of prompts and a schedule. Each prompt is executed against each enabled engine at a defined frequency. Two design choices here shape everything downstream.
The first is repetition. Because generated answers vary, a single daily execution of each prompt produces a noisy series. Trackers that execute each prompt several times per cycle and average the results produce visibly steadier trends. If a vendor cannot tell you the repetition count, the volatility in their charts is partly manufactured.
The second is timing consistency. Engines change behaviour over the day and between model versions. Running at a consistent time removes one source of unexplained movement.
Engine execution
The tracker sends each prompt to the engine and captures the full response. How it does this matters more than vendors usually admit.
Some engines expose official APIs. These are stable and permitted, but the API's behaviour can differ from the consumer product — different system prompts, different retrieval settings, sometimes a different model. Others have no API for their answer product, so trackers automate the consumer interface instead. That is closer to what a real user sees and considerably more fragile.
Neither approach is wrong, but they are not equivalent, and comparing a number from an API-based tracker to one from an interface-based tracker is comparing two different measurements.
Mention detection
Once a response is captured, the tracker decides whether your brand appeared. Naive implementations do an exact string match on the brand name, which fails in predictable ways:
- The brand is referred to by an abbreviation or a shortened form.
- The brand name is a common word, producing false positives on unrelated uses.
- The engine names the parent company or a product rather than the brand.
- The brand is misspelled, which models do.
- The domain is cited but the brand is never named in the prose, or vice versa.
Better implementations handle aliases explicitly, separate text mentions from domain citations, and record the position of the mention within the response. Position matters because being the first brand named carries far more weight with a reader than being fourth in a closing list.
Aggregation
Finally the tracker turns per-response outcomes into metrics: mention rate per engine, share of voice against a competitor set, citation share by domain, and the trend of each. The important property here is that competitors are measured on the identical prompt set at the same time. Comparing your score on your prompts against a competitor's score on different prompts is meaningless, and it is a surprisingly common flaw.
How to sanity-check a tracker
- Pick a prompt where you know a competitor dominates, and confirm the tracker reflects that.
- Open the raw stored response behind a metric and read it. If you cannot, treat the metric as unverifiable.
- Run the same prompt yourself in the engine and compare, accepting that variance means a single mismatch is not proof of error.
- Check whether alias handling exists by testing a shortened form of your brand name.
- Look at the volatility of the trend line. Extreme day-to-day swings usually indicate insufficient repetition rather than genuine movement.
Common questions about visibility tracking
- How often should prompts be executed? Several times per cycle rather than once. Repetition is what separates a readable trend from a chart that moves for no reason, and it is the question vendors are least likely to answer without being asked.
- Why does my tracker show a sudden jump with no explanation? Check three causes in order: someone edited the prompt set, the vendor changed access method or model version, or the engine provider shipped a model update. Genuine market movement is the least likely explanation for an overnight step change.
- Can tracking tell me how many people saw the answer? No. No engine publishes impression data for generated answers, so any figure presented as reach is modelled.
- Should I track brand mentions or domain citations? Both, separately. They fail for different reasons — a mention problem is usually reputation, a citation problem is usually retrieval or formatting.
- How do I handle brand names that are common words? Use a tracker with explicit alias and context handling, and check its false-positive rate manually before trusting the numbers. Naive string matching will badly overstate your visibility.
- Is scraping the consumer interface allowed? It sits in a grey area and can break without warning when interfaces change. API-based measurement is more stable; interface-based measurement is closer to what a user sees. Neither is wrong, but know which you are buying.
Tip: whatever tracker you use, record when you change the prompt set. Almost every unexplained step change in a visibility chart turns out to be a prompt set edit rather than anything the market did.
Still stuck? We typically reply within 1 business day.
Contact support