What exactly counts as a citation?
Definitions first, because sloppy definitions produce fake precision. We count a citation when an engine’s answer links to or explicitly names the client’s domain as a source: a Perplexity citation chip, a ChatGPT source link, a Google AI Overview link card. A brand mention without a source link is tracked separately — it feeds the Brand Perception and Visibility dimensions, not Citation Rate. Ambiguous cases are resolved against the written rubric before scoring, never during, so the scorer cannot move the goalposts.
Why a fixed question set?
Comparability. We build 50-100 questions per client from real buyer prompts — sales calls, support tickets, community threads, fan-out mapping — then freeze the set for at least a quarter. Changing questions between cycles makes scores incomparable; a frozen set makes month-over-month movement meaningful. The set covers the whole fan-out tree deliberately: definitional, comparison, pricing and how-to prompts, per query fan-out.
How does repeated sampling work?
- Each question is run multiple times per cycle per engine, in fresh sessions, because answers vary run to run.
- Citation Rate per question = share of runs where the citation appears. Citation Rate per engine = average across the question set, expressed 0-100.
- The same runs feed all four dimensions — Visibility, Citation Rate, Semantic Coverage, Brand Perception — so one sampling pass produces the full dashboard.
- Every run is archived with screenshots and timestamps. If a stakeholder disputes a number, we open the archive.
How do we keep the numbers honest?
- Control questions. A subset of prompts about topics the client does not address detects engine-wide shifts masquerading as client progress.
- Engine annotations. Model updates and interface changes (like the I/O 2026 AI Overviews/AI Mode merger) are logged, so deltas can be read in context.
- No cherry-picking. Questions are never removed mid-quarter for performing badly — a rule we write into client agreements.
- Threshold discipline. Movements under roughly five points are reported as noise, not trend.
How does this plug into the 90-day POC?
Week 1 produces the baseline: full question set, all engines, plus a crawler-log snapshot. Weeks 2-10 are the intervention window. Week 12 re-runs the identical set, and the deltas — test topics versus controls, this cycle versus baseline — are the POC result. It is deliberately boring: the value is that clients can re-run every question themselves. The full standard is public at the citation-rate standard and the measurement standard; the experimental design around it is in the GEO testing framework.
What do we refuse to report?
Position in answers (answers have no stable positions), single-run screenshots as proof, and traffic guarantees derived from citation gains. Citation Rate measures earned presence inside AI answers; what that presence is worth commercially is a separate, honest conversation about branded demand and referral behavior — see zero-click 2026.
One implementation detail that saves real pain: separate data collection from scoring. Runs are collected and archived raw; scoring against the rubric happens afterward, in a second pass, ideally by someone who did not run the collection. When a client questions a score months later, you can re-score from the archive under the original rubric — an audit trail that turns a methodology argument into a ten-minute lookup.