GEO Analytics Professional

AI visibility is a chain of four links: AI crawlers read your pages, assistants mention you or cite your page, people click through, and some visits turn into bookings. Each link leaves its trace in a different place: server logs, sampled answers, web analytics, the booking system. Measure every link where it lives, follow the chain to the link that breaks, and judge it by what comes out the end: visits beat mentions.

A measurement prompt set is a sample of the questions real buyers ask, not the questions you hope they ask. Build it from calls, support tickets, site search, reviews and sales notes; stratify it by intent (discover, compare, decide) and by segment (city, service, audience); weight each prompt by expected demand; freeze a panel you ask the same way every period, and keep a rotating set for discovery. Version it, so a trend always compares like with like.

Teams argue about numbers when the definitions live in heads. A metric dictionary writes each metric down once: name, definition, formula, source, grain, period, owner and caveats. The core AI metrics are mention rate, citation rate, share of voice, average position, accuracy, AI referral sessions and AI-assisted key events. Every dashboard, report and experiment uses these definitions and no others.

The same prompt gives different answers on different runs, even with settings that sound deterministic, so a single answer is an anecdote. Sample: ask each prompt several times (5 for a quick look, 10 to confirm, 15 to be rigorous), in clean sessions, spread over days, in each assistant you care about. Report every rate with a confidence interval, and call a change real only when the intervals separate. Log every run.

A crawler read is a line in your server or CDN logs: IP address, time, request, status, bytes, referrer and user agent. Identify AI crawlers by user agent, then verify them against the published IP ranges, because user agents are easy to fake. Separate three kinds: training crawlers, search indexers and user-triggered fetchers, which mean someone asked about you right now. Then analyse: which pages are fetched, with which status codes, how fresh each key page is, and how volume moves over time.

AI referral visits land in web analytics, but under several names: referrers such as chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and claude.ai; a tag ChatGPT adds to its links; and, as of 2026, an AI Assistant channel in GA4 that excludes Google AI Overviews and AI Mode, which count as organic search. Build one channel that catches them all, watch landing pages, mark key events, and accept that visits from apps that strip the referrer show up as direct.

The three data sets meet at the page. Build a page ledger: for each page, crawler fetches and freshness from the logs, citations from the sampling log, AI sessions and key events from analytics. Then find the break per page: fetched but never cited, cited but never visited, visited but never converting; each has its own fix. For attribution, last click undercounts AI, because an answer often leads to a search or a direct visit later, so triangulate with assisted conversions, branded-search lift and what customers tell you.

A rate on its own has no scale. Benchmark it against named competitors in the same prompt panel: share of voice, average position, and which domains the answers cite for each brand, including directories and review sites. Watch who is mentioned together with you. Benchmarks need the same intervals as your own numbers, and the useful output is the gap and its cause, not the rank.

A change followed by a better number proves little: platforms shift, seasons move, and answers vary. Run experiments that hold up: change one thing on a set of pages, keep a control group of similar pages unchanged, measure both before and after (difference-in-differences), use enough runs and weeks to beat the noise, and write the metric, the threshold and the decision date before you start. Judge wins by visits and conversions, not mentions alone.

A dashboard is for decisions, not decoration. Use the chain as its spine: one row per link with the number, its interval, the change, and the source it came from. Add trend lines with annotations for the changes you made and the changes platforms made. Set a few threshold alerts: key pages no longer fetched, citation share below a floor, AI sessions down for two weeks. Every tile answers "so what, do what".

Rank prompts by what their pages earned, not by mentions. For each prompt: the pages that serve it, their crawler reads and indexing, the AI and search visits they brought, and the conversions. A prompt can win without being cited; a cited prompt can bring no one. Double down on winners with more prompts like them and more pages; send cited-but-unvisited prompts back to the page, and uncited prompts to the content team as gaps.

Leadership needs one page: the four links with deltas and intervals, the money (conversions times value, as a range), the decisions taken and the decisions needed, what changed on the platforms and when, and the next experiments. Report ranges, not false precision, and say what you cannot see: visits that arrive as direct, impressions without clicks, citations without visits. Lead with the decision.