How do I measure if my GEO or AI visibility efforts are working? Our client method
How do I measure if my GEO or AI visibility efforts are working? The four-layer method we use on client programs, built around Promptwatch reporting.
Clients rarely ask "what is our Visibility Score?" They ask, three months in, "is this working?" Our answer is never one number. We report four layers, and we set that expectation in the first call.
We run these programs on Promptwatch, so the examples below use its reports. The method itself would work with any tool that covers the same ground.
Layer one: appearance
We start with a fixed prompt panel. Typically that is a few dozen to a couple of hundred questions, organized by topic and by buyer persona, with competitor-comparison and brand prompts kept in their own groups. The panel is frozen for the baseline period. New prompts go into a separate group, so the trend line stays honest.
Promptwatch runs the panel daily per model, country and language. From it we report Visibility Score, Share of Voice against agreed competitors, Position in Answer and Sentiment. The A/B period comparison is the view we use most in client meetings, because it shows before and after for the same prompts.
One setup step people skip: the Brand Book and brand aliases. They decide what counts as a mention. If the rules are wrong, every downstream number is wrong, so we write them with the client and review them after the first week.
Layer two: access
Before we change a single page, we check whether the AI crawlers can reach the site. Promptwatch's Agent Analytics gives us IP-verified crawler logs, which means spoofed user agents are removed. We look for GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot and PerplexityBot, and we look for errors, blocked paths and important pages that never get fetched.
This step regularly changes the plan. Promptwatch's TreeHouse case study, for the Pararius and Huurwoningen.nl sites, describes AI search becoming about 25% of organic conversions, 99% of it from ChatGPT, after crawler access was fixed and weekly reporting was automated. That is TreeHouse's result, and we do not promise anyone the same. We do take the lesson: if the crawler cannot see the page, nothing else matters.
Crawler logs start on Promptwatch's Professional plan, and they need a log source such as Cloudflare, CloudFront, Vercel or the WordPress plugin. We scope that with the client's developer early.
Layer three: citation
Next we look at citations: which of the client's pages are cited, which domains the models trust for their topics, and where the gaps are. Citation Trends is useful here, because it shows when a page was first cited, how quickly it ramped, where it peaked and how fast it decayed. When decay arrives, that page is on the refresh list.
We add the client's key URLs and a handful of competitor URLs to Page Tracker. We also read the offsite mentions report, which flags third-party pages that name the brand without linking to it.
Layer four: visits and conversions
The last layer is the one the finance person cares about. Promptwatch's visitor analytics counts AI-referred visits and conversions from ChatGPT, Perplexity, Gemini, Claude, Copilot and others, through a script, a GTM template or the WordPress plugin. Search Insights then overlays Google Search Console and Bing Webmaster data, so we can show classic search and AI side by side.
We are careful here. AI-referred traffic is often small at first, and some AI use never produces a click at all. We tell clients that visibility can run ahead of visits by weeks. If we only promised traffic, we would lose the argument in month two.
What we put in the monthly report
- The same prompt panel, compared with the previous period.
- Share of Voice against the named competitors, with the movement explained in a sentence.
- Crawler errors fixed and pages newly reached.
- New and lost citations, with the pages behind them.
- AI-referred visits and conversions, with a plain note on how small or large the sample is.
- Three actions for next month.
Promptwatch's custom reports, scheduled PDF exports and the Looker Studio connector let us automate most of this, and the Action Board gives us a prioritized to-do list with evidence behind each item.
What measurement cannot tell you
It cannot tell you that a model will recommend the client next week. Answers vary from run to run, and a single good day proves nothing. It also cannot isolate cause perfectly: if you refresh a page and cite it more often, other factors may have helped. We say this out loud and look for patterns that repeat across several topics before we claim a win.
Start small
If you are doing this yourself, build the prompt panel first, record the baseline, check crawler access, then change one topic at a time and note the date. Promptwatch has a free Explore plan with 10 prompts on ChatGPT, which is enough to learn how the reports look before you commit to a paid tier.