The five signals, in order of directness
Changed answers. The bedrock proof: a question that used to name competitors, or nobody, now names you. Keep before and after screenshots per assistant; nothing survives a budget review like them.
Share of voice trend. Of your category’s buying questions and moments, what percentage of answers include you, measured monthly per assistant? Up and to the right across ChatGPT, Gemini, Claude and Perplexity is the program working; flat means your fixes are not reaching the sources that matter, the metric defined in AI share of voice.
Citation wins. When assistants cite sources, are they citing pages you created or fixed? Perplexity shows this openly, which makes it your fastest feedback loop.
AI referred traffic. Sessions arriving from assistant surfaces trend up as answers include you, visible in analytics referrers even though attribution undercounts.
Pipeline. The final word: leads and revenue from AI referred sessions, which is what makes GEO and AEO a channel instead of a hobby.
How to instrument this in one afternoon
Fix a question set: thirty to fifty real buyer questions and confided moments. Run them through all four assistants monthly, same phrasing, fresh conversations, logging mentions, positions, framing and citations. Tag AI referrers in analytics. Put it all in one sheet with a month column, and you have a working measurement system, the full KPI stack is in how to measure AI visibility.
What not to trust: single blended scores that hide which assistant moved, mention counts without buying intent behind them, and any measurement that cannot say which fix caused which change.
If the answer is “it is not working”
Flat numbers after a quarter usually mean one of three things: your content answers keywords instead of the questions people actually ask, your fixes have not touched the third party sources assistants cite, or you are measuring one assistant while losing the other three. The diagnosis order is in why your brand is not showing up in AI search. Or skip the self-diagnosis: a free AI visibility audit baselines all four assistants with screenshots and tells you exactly where the effort is leaking, on a 30 minute call.
