LLM Visibility Stack®

Measurement

Why every question is asked 12 times

Ask ChatGPT the same question twice and you can get two different lists of names. A report built on one check tells you which answer came up that time, and nothing about how often it comes up.

This is how Tier 5, Impact, is measured every month.

What one check misses

Two numbers set the method

The test behind these numbers is published as One AI check against twelve.

21.7%

How often a single ChatGPT check gave the less common answer about whether a brand was named.[1]

4.7%

Of the question months where ChatGPT named a brand at least once, the share where it named that brand in all 12 runs.[1]

Put the two together and one check makes a poor monthly number. About one time in five it says the opposite of what most of the 12 runs would have said, and even a brand that ChatGPT does name gets named in every run in fewer than one question month in twenty.

Try the simulator: set a rate and a number of asks, and watch how far one report can sit from the truth.

Figures dated 2026-10-25, from answers collected 12 to 26 September and 1 to 5 October 2026. ChatGPT here is gpt-5-mini with web search on.

The method

How a month is measured

Fixed questions
Written with you before the baseline, in each language your customers use. They stay the same from month to month, so this month can be compared with the last.
Twelve runs
Every question, on each engine, in each language, every month.
Engines kept apart
ChatGPT and Perplexity behave differently. In the same study a single Perplexity check gave the less common answer 5.7% of the time, against 21.7% for ChatGPT.[1] One blended number would hide which engine moved.
A margin of error
Every rate comes with a 95% interval: the range the true rate most likely sits in.
Real or noise
Every change from last month gets a verdict. A move the margin of error can explain is called noise, however good it looks.
Dated and labeled
Every number carries a date, an engine and a model, because the models change underneath all of us.

Limits

Where this measurement stops

  • Two engines. ChatGPT (gpt-5-mini) and Perplexity (sonar), asked through their APIs. Those answers can differ from what a signed-in person sees in the apps. Gemini and Claude are not yet measured, and Google's AI features stay with your SEO agency.
  • One question shows only big swings. At 12 runs, a single question's rate carries a margin of about plus or minus 17 to 25 points, so a month to month change on one question has to be 33 to 42 points before it counts as real.[1] Pooled across the full question set, the margin is narrower.
  • Model updates look like results. When a provider changes its model, rates can move with no work done. The date and model on every number are there so a model change is not mistaken for progress.
  • Zero is not absence. A brand named in about one run in ten is missed by all 12 runs 28% of the time.[1] Not named in 12 asks means rarely, not never.

The same measurement runs every month on 11 websites.[2]

Sources

  1. The Best SEO Podcast, One AI check against twelve. Figures dated 2026-10-25, from answers collected 12 to 26 September and 1 to 5 October 2026.
  2. EWR Digital's monthly AI visibility tracking, October 2026 run: 11 websites measured on ChatGPT and Perplexity. Internal, not published.

Cite this page

"How the LLM Visibility Stack® is measured." llmvisibilitystack.com, updated October 28, 2026. https://llmvisibilitystack.com/measurement

To cite the method itself, cite the book: Bertram, M. (2026). LLM Visibility: A Decision-Grade System For Winning AI-Mediated Discovery.