How do I measure GEO performance? What we track for client programs

Clients rarely ask us "how do we measure GEO performance" in those words. They ask "is any of this working?" The answer needs more than one chart, so here is the measurement routine we run on client programs and the reasoning behind it.

Start with a prompt set, not a score

Before we touch a page, we agree on a list of buyer prompts with the client. Category questions ("best X for Y"), comparison questions, and a few brand questions. We tag them by topic and intent and run them daily on a schedule, per engine and per country. The set stays frozen for the first month. If you keep changing the prompts, you cannot tell whether the brand moved or the questions did.

We run this in Promptwatch, the platform our client programs sit on. Its monitors run prompt panels daily per model, country and language, so the baseline is a trend line rather than a one-off screenshot of a chat window. Answers vary between runs, and one lucky answer proves nothing.

From the prompt runs we report four numbers:

• Visibility Score and mention rate across the set.
• Share of Voice against the competitors we agreed to benchmark.
• Position in Answer: first named brand or fifth.
• Sentiment, which Promptwatch scores per response with a weekly summary.

A caution we repeat to every client: what counts as a "mention" is a definition, not a fact. We write it into the Brand Book (aliases, product names, what to ignore) so the count is consistent across months.

Citations explain the visibility number

A mention rate tells you the result. Citations tell you the cause. We pull citation analytics at page and domain level: which of the client's URLs get cited, which competitor pages get cited for the same prompts, and which third-party sources keep appearing. Reddit and YouTube citation reports are worth their own look, because brands are often discussed there with no link to the brand at all. Offsite mentions (the brand named on a third-party page without a link) show up as a separate list.

For individual pages we use the Page Tracker. When we ship a new article, we track its URL and watch whether it is cited. Citation Trends classifies each citation by content type and source type and shows first cited date, ramp-up days, peak rate and decay half-life per page. That is the closest thing to a "did this article work" report we have found.

Crawler logs: the check people skip

Before blaming content, we check access. If GPTBot, OAI-SearchBot, ClaudeBot or PerplexityBot cannot fetch a page, it will never be cited, however good it is. Agent Analytics in Promptwatch shows IP-verified AI crawler logs, so spoofed user agents are filtered out, with errors and failure explanations per request. It also shows a citation rate per page, which links crawling to citing.

One practical note: crawler logs start on the Professional plan, not Essential. When a client is on a tighter budget and needs this, we say so up front rather than let them find out in month two.

Several times, the first "GEO win" on a program has simply been fixing a blocked path or a page returning errors to a bot. It is not glamorous, and it is usually the cheapest fix on the list.

Traffic and conversions

Then we ask whether any of it sends visitors. In GA4 we set up a segment for AI referrers, but some assistants do not pass a referrer, so that number is a floor. Promptwatch's visitor analytics (lightweight script, GTM template or WordPress plugin) recognizes AI referrers such as ChatGPT, Perplexity, Gemini, Claude and Copilot and ties them to conversion events. Search Insights adds Google Search Console and Bing Webmaster data on top, so the client sees AI crawls, citations and classic search in one view. For Google's own AI features, the clicks sit inside regular web search data in Search Console, so we never claim an exact split there.

We do not promise clients a traffic figure. We report what the data shows, and we report the caveats next to it.

Our monthly reporting rhythm

1. Week one: frozen prompt set, baseline Visibility Score, Share of Voice and Position in Answer, crawler access check, analytics tagging.
2. Weekly: a glance at the Action Board, where Promptwatch's Discovery agent suggests evidence-backed actions from the data. We edit and prioritize them rather than accept the list blindly.
3. Monthly: a client report covering prompts that moved, pages that gained or lost citations, competitor changes and AI-referred visits. Promptwatch generates PDF reports and Looker Studio connectors, which saves time on formatting.
4. Quarterly: review the prompt set itself. Add prompts when the client launches products or enters markets.

What we tell clients it cannot do

Measurement does not buy placement. No tracker can guarantee a model will recommend a brand, and answers change when a model updates. We frame results as direction and evidence, never as a promise. If a vendor tells you otherwise, ask for the method.

We also say plainly what we cannot see: clicks that never happen because the answer was good enough, and referrers that are stripped. Those gaps are why we triangulate across prompts, citations, crawler logs and analytics instead of trusting any one source.

Where to begin

If you are doing this yourself, build the prompt list this week, set up GA4 tagging, and check your robots.txt and server errors for AI bots before you write anything new. Add a platform when the manual process stops scaling. For us, the deciding factor was never the prompt tracking alone. It was having the crawler logs and traffic attribution in the same place, which most trackers do not offer.