One Sample Lies: Why We Run Every Prompt 3 Times
AI answers are unstable, so a visibility score built on one sample is mostly noise. Kevin Indig's research found that the same prompt, run three times on the same platform, produced consistent citations only 2.2% of the time. Our standard follows from that number: a fixed set of buyer questions, every prompt sampled three times per platform, rates reported as appearance rates across runs — and every report carries a methodology note stating exactly what was sampled, where, and when.
Prompt tracking is polling, not ranking
Kevin Indig, who writes the Growth Memo and runs some of the largest prompt-tracking datasets in the industry, has made the point precisely: prompt tracking is polling, not rank tracking. A search ranking is a stored position you can look up. An AI answer is generated fresh each time, from a system that samples as it writes. Ask the identical question an hour later and the wording changes, the brands rotate, and the sources shift. His consistency finding is the hard version of that point — run the same prompt three times and the cited sources line up only 2.2% of the time.
Anyone who has sold or bought a “ChatGPT visibility score” built on single runs has traded in coin flips. A brand recorded as “mentioned” on Tuesday's single sample may simply have won that generation's lottery; a brand recorded as invisible may have lost it. Decisions made on those numbers — budgets, fix lists, celebrations — are decisions made on noise.
Our standard: three samples, fixed questions
This is our measurement standard, adopted because of the instability above. Full scans run every buyer question three times per platform. A brand's mention rate for a question is not a yes or a no; it is an appearance rate across the three runs — 0 of 3, 1 of 3, 2 of 3, 3 of 3. Consistent presence (3 of 3) and occasional presence (1 of 3) are different business situations, and the report shows both instead of collapsing them into a single percentage that pretends to know more than it does.
The question set is fixed and small on purpose: five to ten real buyer questions per measurement cell, drawn from prompt research and checked against how actual customers phrase their needs. Quality of questions beats quantity of questions. A hundred invented prompts sampled once tell you less than eight real prompts sampled three times — and a fixed set is what makes month-over-month comparison honest, because the instrument does not change between readings. Single-question spot checks still exist in our workspace for quick lookups; they are labeled as spot checks, never as measurements.
Why three runs and not thirty? Because the metric is used for direction and trend, not for decimal places. Three samples separate “never”, “sometimes”, and “always” — the three states a fix list actually acts on — at a cost that lets a small brand afford monthly tracking. It is a deliberate trade, and we state it in the report rather than hiding it.
We measure the products buyers use, not the models
The second half of the standard: we test the six platforms people actually ask — ChatGPT, Perplexity, Copilot, Gemini, Grok, and Claude — as products, in the form a buyer meets them. A raw model endpoint is not what a customer sees. The product wraps the model with browsing, source selection, ranking, and interface behavior, and those layers change which brands surface and which pages get cited. Measuring the model underneath would be measuring a shadow of the thing that decides your sale.
Results are reported per platform, never averaged into one score. A brand can be an always-answer on Perplexity and a never-answer on ChatGPT; the average of those two facts is a fiction that helps no one decide anything.
The methodology note is part of the product
Every report we issue states its sample size, its question set, its platforms, and its dates, next to the numbers. This is the same discipline Indig argues for with polling: a poll without a sample size and a margin of error is an advertisement. If your current GEO report does not say how many times each prompt was run, ask. The answer will tell you how much of the report to believe.
Sources & method. The 2.2% three-run citation consistency figure and the “prompt tracking is polling” framing are from published research by Kevin Indig (Growth Memo). The sampling standard — fixed buyer-question sets, three runs per prompt per platform, appearance rates, methodology notes in every report, six product platforms — is our own, live in our tracking workspace since its v5 release.
© 2026 Jun Liu · JUN & GEO. This article is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0) — https://creativecommons.org/licenses/by-sa/4.0/. You may share and adapt it with attribution and under the same license.