Evidence over vibes
AI search visibility, GEO and AEO: what the engines actually do, what the data shows, and what to do about it.
Bing was the first major engine to give publishers free citation data — total citations, cited pages, and the grounding queries behind them. Here's what it shows and what it doesn't.
Citation share is the slice of an answer's sources that are actually yours — not how often you're mentioned, and not the same thing as share of voice.
Perfect AI attribution doesn't exist. Teams that are measuring anything real rely on proxies: self-reported source, branded search lift, direct traffic, and sales-call mentions.
Pick your real rivals, ask them the same question you ask yourself, run it enough times, and write down exactly how you did it.
A good prompt set mirrors real buying questions at every funnel stage, branded and unbranded, worded several ways — not a dozen variations on your own brand name.
A score is a summary, not evidence. Before trusting any AI visibility report, check the prompts, run counts, dates, actual answers and sources behind the number.
Five numbers, one chart, a finding, a fix list, a next step. That's the whole page — anything more and the boss stops reading.
GA4 can show some visits from AI engines through a native channel and a custom channel group. It misses copy-paste visits, app traffic and most AI Overview clicks — treat it as a floor, not a full count.
There's no 40% that's good everywhere. The only honest baseline is your own competitor set, run the same way, on the same day.
Your share of voice is your slice of every brand mention across a fixed set of prompts and runs. Mix engines or prompt sets between measurements and the number stops meaning anything.
A bare score hides what it's built from. An honest score discloses its prompts, engines, run counts and date, and sits next to the evidence instead of replacing it.
Not in any sense that survives a second run. Presence across repeated runs is the honest unit to chase; position is mostly noise dressed up as a number.
Four metrics survive repeated testing and one does not — here is how to track the four and quit reporting the one.
The same prompt, asked twice, almost never returns the same brand list. Appearance frequency across many runs holds up far better than where a brand lands.