How do I measure my brand's visibility across generative AI engines? The way we do it for clients
Our client method for measuring brand visibility across generative AI engines: prompt panels, share of voice, citations, crawler logs and traffic, in plain steps.
Measuring visibility in generative AI engines is not hard, but it is easy to do badly. The common mistake is to check a few prompts by hand and report whatever came back. We set up measurement the same way for every client, in a way that survives a skeptical finance director. Here is the setup.
Define what counts as visible
Before collecting data, we agree on definitions with the client. Does a mention count if the brand is named last? Does a mention in a negative context count? Do name variants and old brand names count? In Promptwatch the Brand Book handles part of this, because its analysis instructions decide what counts as a mention, and brand aliases catch variants. Settling it early avoids arguments later.
Build a panel that looks like buyers
We write 25 to 40 prompts with the client, drawing on sales calls, support tickets and Search Console queries. Prompt Explorer helps by suggesting prompts from Search Console, ranking keywords, People Also Ask or any page. We group them as category, comparison, alternatives and brand questions, and tag them by topic, intent and persona. Then we set the countries and languages that match the client's markets.
We keep the panel stable. When we need to add prompts, we add them as a new cohort so trends stay clean.
The numbers we report
We run client programs on Promptwatch and report five things:
- Visibility Score per engine, topic and country.
- Share of Voice against three to five named competitors.
- Position in Answer, because order matters.
- Sentiment, with the weekly AI summary as a starting point for qualitative reading.
- Citations, at page and domain level, including Reddit, YouTube and LinkedIn reports.
Prompt panels run daily per model, so we see movement and not single snapshots. Promptwatch monitors ChatGPT Search, Google AI Overviews, Google AI Mode, Copilot, Gemini, Perplexity, Alexa and Meta AI through their live interfaces, and tracks other models through APIs. We tell clients which is which, because the distinction affects how literally to read a result.
Two measures most dashboards skip
Crawler access. Agent Analytics shows IP-verified crawler logs for bots such as GPTBot, OAI-SearchBot, ClaudeBot and PerplexityBot, with errors and a citation rate per page. If an engine cannot read a page, no content change will help. Crawler logs start on the Professional plan, and we flag that to clients choosing a tier.
Traffic and conversions. Visitor analytics counts AI-referred visits and conversions through a lightweight script, a GTM template or the WordPress plugin. Without it, visibility is an abstract number.
How we read the data
We look at trends across weeks. We separate engines, because the same brand can behave differently in each. We compare against competitors, so a rise is read in context. And we use A/B period comparison to judge whether a change we shipped moved anything.
We also read a sample of raw responses every week. Numbers can hide odd descriptions, wrong facts or competitor comparisons that are unfair. The raw text is where we find reputational issues the score does not capture.
Reporting to the client
A monthly or weekly summary with one headline number, three or four findings and a prioritized action list from the Action Board. Scheduled PDF reports save us from rebuilding the same view every time. We keep the format steady so the client can see progress without relearning the report.
What we say about limits
The data tells you where you stand. It does not tell you what an engine will do next month, and it cannot guarantee a recommendation. Responses vary by run, and models change. We say so in the first meeting, because clients who expect a guaranteed ranking are disappointed, and clients who expect a trustworthy baseline are not.
If you are starting from nothing, the shortest path is: write the panel, add competitors, run it for two weeks, check crawler access and set up traffic attribution. After that, you can measure honestly and decide where to work.