The abstract argues that knowing which brands a large language model names is only useful if those mentions persist long enough to act on. To test that, the authors ran a pre-registered replication.
They repeated the same 162 requests — 54 questions in English, Estonian, and Finnish — to Gemini Answers, then compared the results 20 days later. The goal was to see whether historical brand-mention counts could forecast future ones.
The pre-registration is notable because it commits to the analysis plan in advance, reducing the risk of cherry-picking. The available text of the abstract does not include the results, so the outcome of the stability test is not yet known from this summary.