Running an AI-operated company
ChatGPT named you once. Ask it another 60 times
One AI answer naming your business proves little: AI lists change on almost every run. How to measure AI visibility so a change means something.
You change your About page on Monday. On Tuesday you ask ChatGPT for the best firms like yours, and there you are, third on the list. Job done.
On Wednesday you ask again and you are gone. Nothing changed. You were never really third.
That is how AI recommendation lists work, and it catches nearly everyone who starts checking their AI visibility by hand. We have planned a weekly Presso Visibility scan of our own sites, and the plan carries one rule: model answers vary from run to run, so score over several runs and never claim a move from one.
How much do AI answers really change?
More than most people would guess. In January 2026 SparkToro published a study run with Gumshoe, a company that tracks AI answers. 600 volunteers ran 12 prompts through ChatGPT, Claude and Google’s AI a combined 2,961 times, and recorded every list of brands they got back (SparkToro, published 28/01/2026, read 12/11/2026 (opens in a new tab)).
What they found:
| What you might check | What the study found |
|---|---|
| Will the same question give the same list twice? | Less than a 1 in 100 chance, for ChatGPT or Google’s AI |
| Will the list come back in the same order? | Closer to 1 in 1,000 runs |
| Will the list be the same length? | No. Sometimes 2 or 3 names, often 10 or more |
| Do people ask the same question the same way? | No. Across 142 prompts that people wrote for the same need, the average similarity between two of them was 0.081 |
So a single answer is one draw from a lottery. If you do not like it, ask again.
So is checking AI visibility pointless?
No, and this is the useful part. The order is noise, but how often a name appears is not.
In the same study, ChatGPT named one West Coast cancer hospital in 69 of 71 answers. It was the first name in only 25 of them. Position jumped around. Presence did not.
That is the number worth tracking: the share of answers that name you, across many runs of many questions. The study’s lead author started out sceptical and came round to that view. He called visibility percentage “a reasonable metric”, and said any tool that gives you a “ranking position in AI” is “full of baloney”.
How do I measure it properly?
- Write the questions your customers ask. Not your brand name. “Best accountant for a one-person business in Leeds”, not “Is Smith & Co any good”.
- Ask each one many times. The study suggests at least 60 to 100 runs if you want to know a model’s real set of answers.
- Count appearances. Named in 40 of 60 answers is a fact. Third place on Tuesday is a coin toss.
- Change one thing, then repeat the same questions. Same wording, same number of runs. Otherwise you cannot tell your fix from the lottery.
- Never celebrate one run, or panic over one.
Whatever tool you use, ours included, ask how many times each question was run before you trust the score.
Where do I start?
The free AI visibility scan gives you your score in ChatGPT, Claude and Gemini, and a fix guide to keep. No card, no call. Send us your website address at hello@pressonetwork.com.
Then do the boring bit. Ask the same questions again next week, the same way, and count.