The four-layer report
Layer 1: one number and its trend
Lead with a single visibility score — the share of AI answers to the client's customer questions that name them, averaged across ChatGPT, Gemini, Perplexity, and Claude over a trailing window. One number, 0–100, with a delta since last report. Executives don't remember methodology; they remember "we went from 31 to 44." The methodology lives one layer down for whoever asks — and someone good will ask.
Layer 2: the per-engine split
Four engines, four bars. This layer does two jobs: it shows you're measuring the market rather than cherry-picking the friendliest engine, and it explains your strategy. "Gemini's fine because your SEO is strong; Claude has never heard of you, which is why we're doing footprint work" is a sentence that earns renewals.
Layer 3: question-level wins and losses
Pick the five to ten questions that matter most commercially and show the actual movement: named versus absent, position in the answer, facts right or wrong. This is where "alternatives to [market leader]" going from zero mentions to named-with-accurate-pricing becomes a story the client retells internally. Losses go here too, with a diagnosis attached — clients forgive an honest "not moved yet, here's why" far more readily than they forgive discovering it themselves.
Layer 4: receipts
Before-and-after answer text for two or three highlights. Nothing you say about AI visibility is as persuasive as the engine's own words naming the client where it previously named only competitors.
Cadence and honest expectations
Monthly is right for most engagements — engines shift meaningfully month to month, and weekly reporting amplifies noise into anxiety. Set the timeline expectation in the first report: retrieval-driven improvements (cited sources fixed, new pages picked up) show in weeks; memory-driven improvements (footprint, entity work) show over quarters and model updates. Write both down early. The alternative is explaining in month two why the memorized answer hasn't budged, which is a worse meeting.
What to refuse to report
- Single-conversation screenshots as evidence. Engines are stochastic; one great answer proves nothing. Averages over repeated runs or it didn't happen.
- Guaranteed placement. There's no bribable ranking. Anyone promising "we'll get you into ChatGPT's answers" by Friday is selling something adjacent to astrology.
- Metrics without a question set. "AI mentions are up" — asked what? By whom? A fixed, versioned question panel is what makes the trend line mean something.