A Gemini study tests whether brand-mention measurements hold for 20 days
The preprint reports useful mention-history forecasts in one setup. It does not establish lasting web citations, traffic or AI-search rankings.
An AI brand-mention measurement is useful only if it remains informative when a team acts on it. A new Gemini preprint examines that question over roughly 20 days. Its reported result supports a bounded forecasting claim about one set of brand mentions, rather than a general rule for AI search rankings.
The study repeats a declared request set
Dmitrij Żatuchin of Rankfor.AI submitted Twenty Days Later on October 14. The paper describes 54 questions in English, Estonian, and Finnish; 12 candidate software brands; and 162 requests per wave. The follow-up ran on October 13, 20.3 days after the third earlier wave. Under its preregistered rule, earlier question-level mention history remained informative.
The author’s Rankfor.AI affiliation is relevant commercial context. This is an original preprint, and SearchEngineAnswer has not independently replicated it. The result covers one model, one interval, and the sampled software categories. The paper estimates no general decay rate.
Search was enabled, but the reported answers did not search
The paper reports no search query in the collected answers despite search being enabled. Its outcome is a lexical brand-mention label. That distinction prevents a common transfer error: treating stable mentions in this setup as evidence that web citations, referrals, or live search results will remain stable for 20 days.
A brand can be named without a supporting link. A linked source can appear without the answer naming its publisher. A later visit requires another event. The AI visibility measurement crosswalk explains why those observations belong in separate fields.
What a repeat measurement should preserve?
| Field | Retain | Interpretation boundary |
|---|---|---|
| Task and request | Question wording, language, model identifier and request settings | A revised question begins a different comparison. |
| Time | Each collection window and the interval between waves | One interval does not establish a decay curve. |
| Mention label | Brand spelling rule, answer text and ambiguous cases | A lexical match is different from recommendation or factual support. |
| Search and sources | Observed search calls and linked destinations, where exposed | Permission to search does not establish invocation. |
Imagine a fictional team finding the same brand in three later answers. If the team changed the model between waves, it cannot assign that agreement to time stability alone. If the answers contain no web search, the team should report mention recurrence under those conditions rather than call the observation a search-ranking result.
Use the finding to improve the next comparison
For a follow-up study, freeze the questions and labeling rules before collection, retain incomplete runs and compare the historical forecast with a simple baseline. Keep the account, client, language, and observation window with the result. Report disagreements as well as stable answers.
Those are proposed safeguards, not a new SearchEngineAnswer experiment. The prompt-family measurement guide addresses a related design problem on another platform; its results should not be pooled with this Gemini sample. The useful next question is whether the same pattern holds under a separately declared model, task, and interval.
Keep learning
Continue this topic
Measure referrals
Set up the GA4 investigation
Check definitions
Keep each visibility event separate
Test the score
Compare visibility with actual visits
Research
Ask a question or join the discussion