Gemini changes models. Does the comparison change?

On 8 October, Google changed the route for requests to Gemini 3.7 Flash. Its official changelog says they are automatically routed to Gemini 3.8 Flash, while 3.7 is no longer selectable in AI Studio. A second transition routes Gemini 3.5 Flash requests to Gemini 3.6 Flash. The deprecations page clarifies that these endpoints are not shutting down: they continue to respond through the successor model. [gemini-changelog-routing][gemini-deprecations]

The requested name is only part of the record

For a team tracking how AI describes its brand, this creates a concrete measurement issue. A collection can keep the same requested model name in its code while crossing a change in the model actually serving it. If a later response cites another page or describes a product differently, a change to the company website is not the only explanation to examine. Routing belongs in the comparison.

What the announcement establishes

Google documents API behaviour, not its effect on a particular brand’s citations. We have no measurement establishing that this transition increased or decreased brand visibility. The announcement does not establish a change to Google Search, AI Mode or the consumer Gemini app either. Those are different surfaces. Keeping that boundary clear avoids turning a technical update into a supposed universal GEO rule.

Keep the measurement in context

A useful observation record includes the requested model, date, any returned version, tools used and exact prompt. When documentation announces a routing change, mark it in the time series. Then repeat a fixed set of questions, separating branded prompts from questions describing a customer need. If the effective model identity is not returned, record that limitation instead of guessing it.

The answer itself contains different signals. A mention names the brand; a citation points to a source; a recommendation proposes a choice under particular conditions. Google’s grounding documentation describes references to web content. It does not make every answer a recommendation, nor does a citation establish that someone visited the linked page. Combining these signals into one count removes the distinction the analysis needs. [gemini-grounding]

The first decision after this update should concern the measurement design, before rewriting the website. A change that coincides with a model transition should be annotated rather than credited automatically to a new page. An incorrect claim repeated across observations and supported by an outdated source points to a different, concrete editorial problem. These diagnoses call for different work.