Why does ChatGPT recommend a different business every time I ask?
Because the answers genuinely change week to week. There is published measurement of how much, and it decides how you should be buying this.
You ask ChatGPT which med spa in Buckhead does the best filler work. You ask again on Thursday and get a different list. Nothing about the businesses changed. This is the single most misunderstood thing about AI search, and it has a direct consequence for what you should be paying anyone to do about it.
How much it actually moves
Orbit Media ran a controlled version of exactly this. They took 72 identical prompts and re-ran them two days apart — same wording, same models, no variables changed. Then they measured how many of the previously-cited URLs survived:
| Engine | Kept its earlier citations |
|---|---|
| Perplexity | 65% |
| Gemini | 38% |
| Claude | 30% |
So on the least stable of the three, seven out of ten sources changed in forty-eight hours.
The engines are not even answering the same question
The same study looked across 1,792 query-and-domain combinations and asked how often all four major engines cited the same domain for the same query. The answer was 1.7% of the time.
That is close enough to zero to be worth restating plainly: there is no such thing as "how you rank in AI." There are four different answers, they disagree with each other almost always, and each one disagrees with itself over time.
What follows from that
A single check tells you nothing. If you ask an engine about your business once and it names you, you have observed one draw from a distribution. Ask again Thursday and you might not be there. The reverse is also true, which is the more expensive mistake: people conclude they have a problem, buy a fix, ask again after the answer happened to move, and credit the fix.
A one-time "AI visibility score" is not a measurement. It is a snapshot of a process with 30–65% week-over-week churn. Anyone selling you a number with a date on it and no tracking behind it is selling you a coin flip in a PDF. You can check this yourself in five minutes: ask the same question three days running and write down what you get.
You need a baseline, not a reading. This is why we run three weeks minimum before we will tell you anything. Three weeks of four engines is enough draws that the average means something and the movement is signal rather than noise. One week is an anecdote.
And you need to keep measuring after the fix, because the thing you are trying to move keeps moving on its own. That is not a sales argument for a retainer, it is the reason a retainer is the only honest shape for this work. If the underlying number were stable, we would sell you a one-time audit and go away.
What to ask a vendor
Three questions, and they are quick:
- How many times will you check, and over what period? If the answer is "once," you are buying a snapshot of a moving target.
- Which engines? All four disagree. A report on one is a report on one.
- Will you show me the raw transcripts? The specific queries, the specific answers, the dates. If the methodology is not something you can inspect, the number is not something you can trust.
Source: Orbit Media, "AI Citation Sources" — 13,184 citations analysed, 72 prompts re-run at a two-day interval, 1,792 query/domain combinations.