How to measure AI visibility without fooling yourself
Share-of-answer, citation velocity and AI-referred revenue — a measurement framework that survives scrutiny, plus the five mistakes that make GEO dashboards useless.
AI visibility is easy to measure badly. Vendor dashboards will happily report a number that goes up and to the right regardless of what happens in your business. The test of a GEO measurement system is not whether it produces a chart — it is whether it survives a skeptical CFO.
Here is the framework we use, and the mistakes we see most often.
The three numbers that matter
1. Share-of-answer
Sample a fixed panel of real buyer prompts — not keyword variants, not “top 100 AI queries” — across a consistent set of engines. For each prompt, record whether the brand is named, quoted or linked.
Share-of-answer is the percentage of prompts where you appear. The absolute number matters less than the trend, and far less than the comparison to competitors on the same panel.
Rules that keep it honest:
- Freeze the panel for at least a quarter.
- Sample the same prompts at the same cadence.
- Record competitor presence at the same time.
- Store raw answers so you can re-score later.
2. Citation velocity
Count the new, distinct sources citing you per week. This is the leading indicator: citations accumulate before share-of-answer moves, and they decay slowly when you stop earning them.
Track the source domain, not just the count. A citation from an aggregator models retrieve often is worth more than ten from a directory they ignore. Over time you build a citation graph: which sources own your category, and whether you are among them.
3. AI-referred revenue
Attribution is imperfect but not impossible. Use three layers:
- Referrer data where engines pass it (Perplexity is the most generous).
- Server-side tracking to catch redirects and app traffic that browser analytics miss.
- Self-reported attribution on the lead form: “How did you first hear about us?” with an AI-assistant option.
Match those sessions to CRM revenue. Even with 40% attribution coverage, the directional signal is enough to make budget decisions.
The five mistakes
Mistake 1: Sampling once. A single snapshot is a press release, not a measurement. Models change weekly.
Mistake 2: Using your own prompts. Your team writes prompts that include your brand. Buyers do not. Build the panel from sales calls and real search demand.
Mistake 3: Counting mentions as wins. Being mentioned in a list of “alternatives to avoid” is not visibility. Score sentiment and position, not just presence.
Mistake 4: Ignoring variance. The same prompt can produce different answers on different runs. Sample each prompt multiple times and report a confidence band.
Mistake 5: Reporting to the wrong altitude. Operators need prompt-level detail; executives need share-of-answer, citation velocity and pipeline. One dashboard cannot serve both.
A minimal, defensible setup
If you are starting from zero, this is enough:
| Component | Cadence | Owner |
|---|---|---|
| 50-prompt panel, 3 engines | Weekly | Marketing ops |
| Citation source log | Weekly | SEO / content |
| AI-referred sessions | Weekly | Analytics |
| Self-reported attribution | Continuous | Demand gen |
| Share-of-answer readout | Monthly | Growth lead |
That is a spreadsheet and two hours a week. It will outperform most paid AI-visibility tools, because you control the panel and you understand the category.
What “good” looks like
There is no universal benchmark. What matters is the trajectory relative to competitors on the same panel, and whether the movement in citations eventually shows up in pipeline.
If citation velocity rises for two quarters and revenue does not move, either your category does not buy through AI assistants yet, or you are winning the wrong prompts. Both are useful findings. Neither shows up in a vanity dashboard.