Gemini brand mentions prove stable over 20 days, study finds

Twenty days is a long time in model-land. Google shipped Gemini 2.0, 2.5 and three Nano Banana image releases inside a single year; the idea that a brand's standing in Gemini's answers might simply hold still for three weeks sounds like a claim nobody should make without a notarised witness. The researchers behind this paper brought one anyway. They pre-registered their prediction, hashed it, filed it on OSF before making a single call to the model, and then went and checked.

The design is unglamorous and that is the point. The team sent 162 identical requests, 54 questions about three software categories in English, Estonian and Finnish, each asking the model to pick among four of twelve candidate brands, three times each. They fired this battery at gemini-3.1-pro-preview in three waves across ten hours on 22 and 23 September 2026, then fired the exact same requests again on 13 October, 20.3 days later. No prompt tweaks, no new brands, no changing of the guard. Just the same question, asked again, after three weeks of silence.

The forecast that won was the boring one

A forecast built from each brand-and-question cell's own mention rate over the nine earlier answers scored a Brier score of 0.0281 and a log loss of 0.126. For readers who do not keep scoring rules in their back pocket: lower is better, 0 is omniscience, 0.25 is the score you get from a coin flip calibrated to a 50% base rate. This cell-level history forecast beat a model using brand-only history (0.104), a model using the overall mention rate across everything (0.234), and the laziest possible benchmark, a constant guess of 0.5. It also came close to a registered ceiling of 0.0559, the theoretical best score given how noisy Gemini's answers are even within the same ten-hour window. In plain terms: knowing how often a brand got named for a specific question last month predicts, with real precision, how often it will get named for that same question three weeks later.

That is not a small finding dressed up in statistics. It is a direct rebuttal to the working assumption inside most marketing and comms teams tracking "share of model," namely that LLM outputs are too volatile to budget against. Volatility exists, mostly within a single session, from one of the three repeated calls to another. But the exercise of running a prompt a dozen times today to establish a brand's baseline mention rate is not wasted effort that expires by next Tuesday. Per the researchers, that baseline survives three weeks largely intact, at least for Gemini's current preview build answering stable, repeatable, category-level questions.

Why this is inconvenient for the "nothing in AI search is measurable yet" crowd

Financial-services comms teams and policy institutions have spent two years treating LLM citation tracking as something between astrology and guesswork, reasonably enough, given how often model weights change underneath them. This paper does not say LLM visibility is stable forever. It says something narrower and more useful: for a fixed model version and a fixed question set, mention rates are sticky enough to forecast, and sticky enough that last month's audit is not obsolete data. For a multilateral institution tracking how often its name surfaces against peer organisations on a specific policy topic, or an industrial group checking whether Gemini still lists it among the default options in its category, that stickiness is the difference between a measurement programme and a one-off snapshot.