Prompt-set design rules
- Source prompts from real buyer language — sales calls, support tickets, community threads, PAA boxes. Invented corporate phrasing measures nothing buyers do.
- Grade by intent depth — ~20% head ("best X"), ~40% mid-tail (with constraints: size, region, budget), ~40% long-tail situational ("what should a 20-person logistics firm use to…").
- Include the uncomfortable prompts — "[brand] problems", "[brand] vs [competitor]", "is [brand] legit". You need Brand Cognition measured where it hurts.
- Version the set. Prompts evolve with the market; change at most 10% per cycle or trends break.
- Localize. Outbound brands test in the buyer's language and, where relevant, the buyer's locale setting.
Sampling a probabilistic system
AI answers are stochastic. Our standard protocol: each prompt runs 3 times per engine (fresh sessions), and mention/citation/sentiment are scored by majority. Temperature and personalization mean single runs are anecdotes; three-run majority turns them into data. For regulated or reputation-critical categories, use 5 runs.
Scoring and aggregation
- Per-engine first. Never average across engines before examining — the average hides exactly the engine-specific problem you need to fix.
- Then composite. GEO Score = 0.35·Visibility + 0.25·Citation Rate + 0.25·Semantic Coverage + 0.15·Brand Cognition. Weights are published so you can recompute with your own priorities.
- Trend over level. A single score is a photograph; the monthly series is the movie. Budget decisions belong to the series.
Cadence and reporting
| Layer | Cadence | Content |
|---|---|---|
| Alert layer | Continuous / daily | Negative-sentiment flips, new wrong facts, lost citations on money prompts |
| Operating layer | Monthly | Full four-dimension run, per-engine detail, fix-list progress |
| Strategy layer | Quarterly | Category benchmark vs named competitors, prompt-set revision, budget allocation |
The GEO Checker implements a lightweight version of this methodology (30-prompt graded set, 5 engines, four-dimension output) free; the full standard is deliberately reproducible without us.
Frequently asked questions
Why not just use an AI visibility tracking tool's default score?
Default scores hide prompt selection — and prompt selection is the measurement. A tool that tests 'best [category]' ten ways measures one cluster. Insist on seeing and controlling the prompt set; the methodology matters more than the dashboard.
How stable are AI answers week to week?
In our monitoring, within-engine Visibility for established brands moves a few points week to week, with step changes around model updates and index refreshes. New or thin-source brands swing much more. Stability is itself a maturity signal.