Rank reports answer a question that is becoming less important: where does this URL sit in a list of links. They cannot tell you whether an assistant recommends you. That needs a different measurement system.
The five metrics
- Citation rate. Across a fixed prompt set, in what share of answers is your brand named or linked? This is the headline number.
- Share of voice. Of all brands named across those answers, what proportion are you? Reveals whether you are gaining or the whole category is.
- Description accuracy. Does the assistant describe your services, pricing model and positioning correctly? Errors are actionable — they point at a source that needs fixing.
- Source composition. Which domains are cited in your category? Usually a mix of review platforms, directories, publications and community threads. This is your corroboration roadmap.
- Sentiment and framing. Are you presented as the safe choice, the cheap option, the specialist, or a hedge? Framing predicts whether citations convert.
Build the prompt set properly
Write 40 to 120 prompts in customer language across the buying chain: category questions, comparisons, cost questions, risk questions, recommendation requests, and local variants if geography matters. Include prompts you expect to lose — a set you always win measures nothing.
Then freeze the wording. Changing prompts between runs makes the trend meaningless.
Run it the same way every time
- Same assistants each cycle: ChatGPT, Gemini, Perplexity, Copilot, plus AI Overviews.
- Clean sessions, no personalisation or memory carried over.
- Monthly cadence, and immediately after any major site change.
- Record raw answers, not just a score, so you can audit conclusions later.
Accept the noise
These systems are non-deterministic. Two identical prompts can return different brands. Run each prompt more than once, report rates rather than single results, and read direction over three cycles instead of reacting to one bad month.
Connect it to revenue signals
AI visibility does not appear cleanly in analytics. Referral data from assistants is partial at best. Supplement it with:
- Branded search volume trend.
- Direct traffic to deep pages.
- A "how did you hear about us" field on your enquiry form.
- Sales notes recording when a prospect says an assistant recommended you.
Together those give you a defensible picture. Any single one of them alone does not.
Reporting that drives decisions
A useful monthly report says: here is the citation rate and its direction, here are the questions we lost and to whom, here is what those winning sources have that we do not, and here are the three fixes we are making this month. Screenshots of a good answer are not a report.
That is the discipline behind our AI Citation Monitoring service, and it is how we report on every AI Search Growth engagement.