Run your own AI citation study — the honest methodology
“Studies” claiming ChatGPT cites X% of the time usually recycle each other. If you need a number you can defend — for a report, a pitch, or your own strategy — measure it. Here’s a methodology that holds up.
The design decisions that matter
Prompt selection. Random prompts tell you nothing about your market. Build a stratified basket: your category’s top intents, weighted by how often buyers actually ask them. 100+ prompts is the floor for percentages; 500+ for confident cross-market comparisons.
Repetition. Answers vary run to run — the same prompt cited you yesterday and a competitor today. Run each prompt at least 3× per cell (engine × market) and report the mean with the variance, not a single draw.
Markets. US numbers don’t transfer. Citation and trigger rates differ per country — measure the markets you care about with country (and usState for the US), not a VPN assumption.
Time window. Publish the measurement dates. A study from March describes March — engines change monthly.
The measurements
trigger_rate = prompts where a citation block appeared / total
citation_share = your domains in sources[] / all sources[]
mention_rate = prompts naming the brand in entities[] / total
source_mix = domain-type distribution across sources[] (forums, docs, news, reviews)Queue the whole basket as an async batch — 500 tasks per call, webhook delivery. The math then runs on stored responses; keep the raw JSON so any number is auditable.
What makes it citable
State the sample, method and date up front, publish the prompt basket, and let others re-run it. That transparency is exactly what most published AI-search statistics lack — and why a careful first-party study earns links the borrowed ones can’t.
The share-of-voice framework defines the metrics; monitoring cadence covers sampling depth.