What the number actually is
There is no visibility score inside ChatGPT to look up. What you can build is a measurement, and it has three parts.
Appearance rate. Of your fixed prompt set, in what share of answers does your brand appear at all? This is the headline figure and the one that moves first.
Share of voice. Within the answers where any supplier is named, what proportion name you, against named competitors? Appearance is binary and easy to feel good about; share of voice is what tells you whether it means anything.
Description accuracy. When you are named, is the description correct and is it the one you want? A brand named with the wrong specialism has a visibility number and a positioning problem.
Why the prompt set has to be frozen
Answers vary between runs. Aral, Li and Zuo (2026) measured markedly lower response variety in AI search than in classic results — repeated questions converge more than ten blue links do — but lower is not zero. A single run is a sample, not a reading.
Two rules follow. Use enough questions that one flip does not swing the number; twenty to forty covering your real buying questions is a workable range. And never edit the list to make the trend look better, because the trend is the entire value and a changed list resets it to nothing.
Write the questions the way a buyer types them. Not "best industrial flooring supplier Netherlands" — that is a keyword. A buyer asks "who can install food-grade flooring in a working production hall without shutting it down for a week".
What moves the number, in order of measured effect
Kargaev (2026) ranked generative-engine signals and the hierarchy is steep enough to be an operating plan.
| Signal | Strength (NIS) |
|---|---|
| Brand entity mentions | 0.918 |
| Statistics present in content | 0.747 |
| Citations within content | 0.671 |
| Content length | 0.043 |
| Page speed | 0.000 |
The top of that list is not a content task, it is an identity task: one name, one description, corroborated outside your own domain. The middle is a writing task. The bottom two are the levers most teams pull first, and on their own they do not move AI visibility at all — which is not a reason to have a slow site, only a reason to stop expecting speed work to change this number.
Iyappan (2026) adds the shape dimension: long-form contextual content cites at 92%, entity-rich at 89%, structured-data-heavy at 85%, FAQ-formatted at 67%, keyword-focused thin content at 41%. If your library is mostly the last category, rewriting a handful of existing pages will move the number faster than publishing new ones.
A baseline you can build in an afternoon
- Write 20–40 buyer questions, in the languages your buyers use. Freeze the list.
- Run them with web search enabled. Record, per question: were you named, how were you described, who else was named, which URLs were cited.
- Compute appearance rate and share of voice. Both will be low. That is the point of a baseline.
- Repeat monthly, same list, same conditions.
- Compare the cited URLs against your own pages. What gets cited on a competitor's site tells you what shape of page wins in your category.
Tools such as Otterly.ai or Peec AI automate exactly this and are worth it once there is a trend to watch. Before then, a spreadsheet is enough and teaches you more.
Interpreting what you see
Appearance rising, share of voice flat. More answers include suppliers, and you are riding the category rather than gaining on it. Look at positioning.
Share of voice rising on narrow questions, flat on broad ones. This is the normal and healthy shape. Concentration in AI answers favours large sites on broad questions (Aral, Li, and Zuo, 2026); narrow questions are where a specific supplier wins.
Named but described wrongly. A corroboration problem. The model assembled you from sources that disagree.
No movement in three months after real work. Check eligibility before concluding the work failed — a crawler block or client-side-rendered content will hold every number at zero regardless of quality. The diagnosis is in how to appear in ChatGPT.
Frequently asked questions
How do I increase my visibility in ChatGPT?
Baseline it with a frozen prompt set, then work the signals in measured order: entity clarity first, evidence-bearing content second, external corroboration throughout. Re-measure monthly against the same list.
Is there a ChatGPT visibility score?
Not one published by OpenAI. Any score you see comes from a third-party tool running its own prompt set — useful, but it is that tool's measurement, not a figure the model exposes.
How often should I measure?
Monthly. More often mostly captures run-to-run variance; less often and you cannot tell which change caused what.
How many prompts do I need?
Twenty to forty covering the real buying questions, in every language you sell in. Fewer and one flipped answer swings the percentage; many more and the discipline of running it every month tends to collapse.
Does improving Google rankings improve ChatGPT visibility?
Partly, and indirectly. Classic search work tends to fix the eligibility layer and builds the authority that corroboration rests on. It does not by itself make you the named answer, which is why sites can rank well and stay invisible in AI answers.
Why does my visibility drop without any change on my side?
The retrieved set changes as other sites publish, and response variety means some movement is noise. This is exactly why the trend matters more than any single month.
Should I buy an AI visibility tool?
Once you have something to track. Monitoring a brand that appears in none of its prompts produces a flat line and a subscription; build the baseline by hand first.
Does ChatGPT visibility convert?
Well, on the evidence: AI-referred traffic has been measured converting at 14.2%, far above typical organic rates. Volumes are lower, so measuring appearance rather than sessions is the honest approach.
If you would rather not build the prompt set by hand, the free scan produces the first measurement for you, and the result is yours whatever you do with it.
