How often does an AI engine change its recommendation on its own?

In our July 2026 testing on a clean, logged-out ChatGPT surface, the answer changed 27.8% of the time across three measured windows with no intervention at all (25.0%, 33.3%, 25.0%). This is a small-sample, directional result, not a published statistic.

Why does engine volatility matter before buying AI-visibility work?

If an engine changes its own answer roughly a quarter to a third of the time with nothing changed, then a single before-and-after screenshot is not evidence of anything. A result has to be measured across multiple runs to be distinguishable from noise.

Can anyone guarantee an AI citation?

No. At the noise levels we measured, a guaranteed citation or ranking is not something an honest operator can promise. Get AI Cited does not sell a citation guarantee; it sells a published method and multi-run measurement.

How Much Does an AI Answer Change on Its Own?

The test

We took a fixed basket of buyer questions — the kind a real customer types when choosing a business — and asked them on a clean engine surface. Then we asked exactly the same questions again in later windows. We changed nothing in between. No new pages, no citations, no listings, no work of any kind.

If the engine were stable, every window would return the same businesses. It did not.

27.8%

of the answers changed with no intervention at all.

Three measured windows on a clean, logged-out surface: 25.0% · 33.3% · 25.0% — 10 status changes across 36 observations. Small sample, non-consecutive windows, directional result. We label it as exactly that.

Why this matters before you buy anything

Roughly a quarter to a third of the movement in an AI answer, on the evidence we have, is the engine moving on its own. That has three consequences, and they are uncomfortable for everyone selling this service — us first.

A single before-and-after screenshot is not evidence. If an answer changes a third of the time by itself, a "look, you appeared!" screenshot could just as easily be noise. So could "look, your competitor dropped."
A result has to be bigger than the noise to count. One run proves nothing. Our protocol is a minimum of three runs, averaged, against a question set fixed before the first measurement — never chosen afterwards to suit the outcome.
Nobody can honestly guarantee a citation. We do not sell one. We had a guarantee drafted and we removed it, because we did not believe our own number could support it. If a provider offers you a guaranteed AI ranking or citation, ask them what their measured noise floor is. Ask to see the runs.

What this does not say

It does not say AI visibility work is pointless — the opposite. It says the work has to be measured properly to be worth anything, and that the industry's default proof (one screenshot, one date) is worthless. It also does not say your engine behaves exactly like ours did: this is one basket, one engine, three windows. It is a directional finding from a small sample, not a law.

How it was measured

Clean surface only. Logged-out ChatGPT, temporary chat, free tier. No account, no history, no personalisation. A measurement taken while signed into an account that has been researching your own company is measuring your history, not the engine.
Fixed basket. The same questions every run, written before the first measurement.
Dated logs. Every run recorded with its date — including the runs that went against us.
Declared exclusions. Where no clean surface exists for an engine, that engine is marked suspended rather than quietly reported as a result.

You can reproduce it: pick ten questions your customers actually ask, open a logged-out temporary chat, ask them, write down who gets named. Do it again a week later, having changed nothing. Count how many answers moved.

What we do with it

This is the reason our reporting looks the way it does: multi-run averages, a published method, dated logs, and results reported either way. It is slower and less flattering than a screenshot. It is also the only version of this that survives contact with the engine's own behaviour.

Read the full method · Get the same measurement run on your business