Businesses invest in AEO and GEO and then have no idea whether it worked, because the channel does not report itself the way search does.
There is no console showing citation share. Most AI answers generate no click at all. The value is real and the measurement is genuinely partial. Anyone selling you a clean number for this is selling you something.
Here is what can actually be measured.
1. Direct Prompt Testing, Run Like a Method
This is the primary instrument, and it works if it is run with discipline instead of curiosity.
Build a fixed list of 20 to 30 prompts phrased the way buyers actually ask. Not keywords. Questions. "Who does immersive web design in Phoenix." "What should a cinematic brand video cost." "Best drone videographer in Arizona for a resort."
Run them monthly across ChatGPT, Perplexity, Claude, Gemini, and Google AI Overviews. Use a signed-out or fresh session every time, because personalization and memory will show you a flattering result that no prospect will ever see.
Record four things per prompt: whether the brand appeared, its position in the answer, how it was described, and which source pages were cited.
The fourth column is the most useful and the most ignored. It tells you which of your pages the engine considers authoritative, which is a direct instruction about where to invest next.
2. Referral Traffic, Understood as a Floor
Segment analytics traffic from AI hosts. Perplexity and ChatGPT browsing sessions do produce clicks, and that traffic is real.
It is also a severe undercount. A large share of AI answers are consumed without a click. Someone who reads a recommendation, remembers the name, and searches for the brand three days later shows up as branded organic or direct. The citation caused the visit and the data will never say so.
Treat this number as a floor that proves the channel exists, not as its measurement.
3. Branded Search Volume
Rising branded search with flat paid spend and flat unbranded impressions is one of the better available proxies for AI recommendation. People are hearing the name somewhere and going to look it up. Search Console makes this visible over a long enough window.
4. Accuracy Auditing
Appearing is not the goal. Appearing correctly is.
When a model describes the business, check whether the description is right. Wrong price, outdated service list, wrong city, a service you no longer offer. Every one of those is a lead cost, and every one is fixable at the source, because the model got it from somewhere. Usually a stale directory listing or a contradicted fact. The consistency audit is the repair.
5. The Question That Actually Matters
Ask new leads how they found you, and record it. Not a dropdown with six preset options, an open field or a direct question on the call.
"I asked ChatGPT and it mentioned you" is now a real answer people give, and it is the only unambiguous evidence in the entire stack. Twenty of those in a quarter tells you more than any dashboard.
What The Timeline Honestly Looks Like
Retrieval-based engines read the live web, so fixing rendering, schema, robots access, and content structure can change answers within weeks. Training-data influence is slower and compounds as models update. Entity confidence accumulates rather than switching on.
Which is the argument for starting now rather than measuring first. The first cited brand in a category tends to stay cited.
UM Media builds the structure that makes citation possible and is direct about what is and is not guaranteed. Structure only, no ranking or citation guarantees. See the SEO / AEO / GEO Foundation, or the Cinematic Brand System.