Why does an AI never give the same answer twice?
Published on
Because each time, the AI reformulates the question differently, launches new searches, and selects its sources all over again. On September 15, 2026, we asked questions to ChatGPT, Claude, and Gemini, then asked them again identically about an hour later: none gave the same list of sources twice, and almost no search was reformulated identically. In median, only 21% to 25% of the pages cited the first time were cited again. A screenshot of an answer therefore tells you almost nothing about a company's visibility.
For what type of question? It all depends on the question. For a question that names a brand or a specific source, the answers vary less, because the AI will look for the same source. For an open-ended question, they vary a lot.
What we measured between two answers
We asked ChatGPT, Claude, and Gemini our questions about visibility in AIs, then we asked them again identically about an hour later. None of the three AIs returned the same list of sources twice, for any question. Almost no search was written the same way twice: 3% for Claude and Gemini, less than 1% for ChatGPT. The question is reformulated from scratch for each answer.
The share of pages cited the first time and cited again the second time is low: 21% median for Claude, 21% for Gemini, 25% for ChatGPT. At the website level, it is slightly more stable: 26% for Claude, 21% for Gemini, and 41% for ChatGPT, which returns more readily to the same publishers. In repeated monitoring of other questions collected by another tool, we measured greater stability for Gemini and Claude, with 50% to 67% of cited pages found again: variability depends heavily on the subject.
Why the answer changes
Three mechanisms overlap. First, a language model produces its text by choosing each word from several possible ones: two answers to the same question are never written identically. Second, before answering, the AI writes its own searches, and as these searches change, the pages found change. Finally, the web itself moves: pages are updated, others become slow or inaccessible.
This last point can be measured. We measured the speed of the same pages twice, twelve minutes apart: nearly one page out of two had changed its tier, from fast to medium or from medium to slow. Yet with ChatGPT, the slowest pages are about half as likely to be cited. A page's speed is a state of the moment, and with it, its chances of being cited.
What other studies measure
SparkToro had volunteers ask ChatGPT, Claude, and Google's AI answers the same recommendation requests in November and December 2025, using their usual settings. Conclusion: there is less than a one in a hundred chance that ChatGPT or Google's AI will give the same list of brands twice, and about a one in a thousand chance of getting two lists in the same order.
Rand Fishkin, who led the study, draws a consequence for measurement from this: an "AI rank" makes no sense, whereas a percentage of presence over many runs of many questions is a reasonable measurement. This is the same conclusion as our own measurements.
What remains stable despite everything
Not everything varies at the same rate. The major choice mechanisms, however, are found from one measurement to another: for ChatGPT, the weight of the rank and of the named sources; for Claude, the weight of the consensus and of pages found by multiple searches. Pages found by search are also slightly more stable than pages cited.
This is what allows us to work despite variability: you cannot predict the next answer, but you can increase the probability of being chosen, by building pages that correspond to these stable mechanisms.
What this changes for a company
Beware of any promise of a "position" in an AI and of any proof based on a single answer. Visibility in AIs is not measured like a ranking: it is measured like a frequency, by asking many questions, multiple times, to several AIs, and counting the share of answers where the company appears.
This is also why FaireDuBruit does not promise citations: what is guaranteed is the presence of pages built to be chosen, and the monitoring of their reading by AI bots. Being cited remains a probability, which can be worked on and measured over time.
What we do not know
Our comparison is based on two responses generated approximately one hour apart via the models' API. Over a longer period, the variability may be even greater. The stability measured in the tracking of another tool relates to other questions and, for OpenAI, to another model.
Sources
- SparkToro, Rand Fishkin — inconsistency of brand recommendations by AIs, .
- FaireDuBruit measurements, from September 4 to 16, 2026 — see "How we measured".