AI visibility tracking asks a specific question: when people ask about a topic, does an AI answer mention your brand or cite your website? A useful study answers that question for a defined set of prompts and collection conditions. It does not claim to describe every answer that every user might see.
ProtoScrape supplies the observations for this kind of research through its Google AI Mode API. Your application or analysis notebook defines the question set, stores the history and calculates the metrics. Start with a clear method before collecting a large volume of data.
Define a question set around customer decisions
Group questions by the decisions your audience makes. A project management product might study category discovery, comparisons and specific requirements such as permissions or reporting. Include questions that do not name your brand, as well as a separate group of branded questions.
Keep those groups distinct. A question that explicitly asks about your company is much more likely to produce a brand mention than an open category question. Mixing them into one visibility percentage makes the result hard to interpret.
Give each question a stable ID and preserve its exact text. If you change a question, record a new version rather than silently comparing it with the old one.
Hold collection conditions steady
Record the engine, country, language, device and time for every observation. The current ProtoScrape scope is Google AI Mode with US, English and desktop settings. Use a consistent schedule that fits your research question and available credits.
Repeated observations help reveal variation. They do not turn a small, selected question set into a representative sample of all search activity. Label results as findings from your study, and avoid presenting the percentage as your share of all AI search impressions.
Choose metrics you can explain
- Mention rate: successful answers containing a defined brand name or alias, divided by successful answers in the relevant question group.
- Citation rate: successful answers citing your chosen domain, divided by successful answers in that group.
- Collection coverage: successful observations compared with the observations you scheduled.
- Source distribution: which domains appear in citations, counting each domain at most once per answer.
Define aliases carefully and manually review ambiguous names. A common word that is also a brand can produce false matches. Keep mention rules consistent between reporting periods, and preserve the answer text so someone can check the classification.
Show the counts behind the chart
Imagine a study with 30 successful answers, of which 12 mention a fictional brand and six cite its domain. You would report a 40% mention rate and a 20% citation rate, alongside the counts 12/30 and 6/30. These numbers are an illustrative calculation, not ProtoScrape performance data.
Report failed observations separately. If several questions fail in one period, the successful sample may represent a different mix of topics. Compare matched questions where possible and explain any missing coverage before interpreting a movement in the percentage.
Turn an observation into a review task
Read the answer and its cited pages together. Check whether a product fact is missing, outdated or unclear in your own documentation. Then make a useful editorial change and continue measuring with the same method. A later change in mentions can guide further investigation, but it does not establish that the edit caused it.
Google says that its existing SEO fundamentals remain relevant to AI features, with no special optimization requirement for AI Mode or AI Overviews. See Google’s guidance for AI features and websites. Use visibility research to understand information gaps, rather than treating it as a guaranteed route to inclusion.
The citation guide explains source normalization, and the agency reporting guide shows how to present the evidence clearly.