AI Mode now runs on two different models at once

Model releases from Google usually get covered as capability news: faster, cheaper, better at code. The more consequential fact about Gemini 3.7 Flash's arrival inside AI Mode isn't what it can do, it's that it arrived quietly, as an option sitting alongside whatever model was already answering AI Mode queries, with no public documentation of which one wins by default.

What actually happened

Gemini 3.7 Flash launched generally on August 13, 2026, positioned by Google as a low-cost model built for coding and agentic workflows, with a 1,048,576-token context window and introductory pricing of $0.75 per million input tokens through the end of the year. Google added it to AI Mode the very next day, August 14, as a manually selectable option available to Google AI Pro and Ultra subscribers, accessible through the "+" icon in AI Mode's model picker, rolling out in English globally. Free-tier AI Mode users don't get the option. The rollout was announced not through Google's official Search Central channels but through a personal X post from Robby Stein, Google's VP of Product for Search, the kind of announcement that reaches AI-search practitioners scanning social feeds far faster than it reaches a typical SEO team's change-monitoring workflow.

It's worth being precise about what "selectable" means in practice, because it's easy to round it up to "optional" and move on. A subscriber has to actively open the model picker and choose Gemini 3.7 Flash for that specific session; nothing about a default AI Mode query changes unless a user takes that extra step. That sounds like it should limit the blast radius to a small, technically-inclined slice of paying users. But AI Pro and Ultra subscribers are disproportionately the same power users who run repeated, deliberate queries, including the exact kind of manual verification checks a practitioner might run to confirm a citation result, which means the population most likely to hand-check AI Mode's output is also the population most likely to be the one triggering the alternate model without realizing it changes what they're checking.

Google has framed Gemini 3.7 Flash's advantage inside AI Mode as "better instruction-following and a better read on intent," language that describes how the model synthesizes retrieved material into an answer, not a change to how material gets retrieved or ranked in the first place. That's the detail worth sitting with: this isn't a ranking change. A page's position in the underlying retrieval set can stay identical while which sources actually get cited in the generated answer shifts, purely because a different model is doing the summarizing.

The pattern this fits

This is not an isolated event. Google made AI answers the global default across its search surfaces on July 10, 2026, a change powered by Gemini 3.5 Flash, and that shift, covered here at the time, turned what had been a two-tier opt-out feature into the primary lever publishers had left. Gemini 3.7 Flash's arrival inside AI Mode a little over a month later is the third meaningful model-layer change to a Google AI search surface inside roughly nine months, following a now-familiar shape: introduce quietly, make selectable rather than mandatory, decline to specify a default or a timeline, and let practitioner-side monitoring tools catch up after the fact.

Why snapshot tracking breaks

Most AI citation tracking tools work by running a fixed set of prompts on a schedule and logging what comes back: which sources got cited, in what order, how often. That approach implicitly assumes the thing being measured, the AI Mode retrieval-and-synthesis system, is a stable target between snapshots. Selectable models break that assumption cleanly. A tool running the same prompt twice in the same week could now be sampling two structurally different synthesis processes depending purely on which subscribers' sessions happen to trigger the model picker's default versus manual selection, and there's no way to distinguish the two in the output.

The practical failure mode is a false trend. Citation presence can shift week over week for a tracked query, and a team reading that shift as a content-quality signal, our structure improved, so citations improved, or our structure regressed, so citations dropped, may actually be reading model-selection noise. Without knowing which model answered which snapshot, there's no way to separate a genuine content signal from an artifact of Google quietly running two systems in parallel. That's a problem our own coverage of Search Console's generative AI report already flagged from a different angle: the available first-party measurement counts frequency only, with no model, query, or placement attribution attached, which means there's currently no first-party way to check whether a citation shift traces back to a model change at all.

The two failure modes compound each other rather than existing independently. A team with no model-attribution data and a snapshot-based tracking cadence is flying blind in two directions at once: it can't tell whether a shift happened, and if it did notice one, it can't tell why. That's a worse position than either problem alone would create, because it removes the two most obvious ways a team might otherwise catch itself making the wrong call, cross-checking against a first-party source or re-running a query to confirm a result actually reproduces.

What to do about it

None of this means AI Mode's citation behavior has become unmeasurable, it means it's become measurable only with a different method than the one most teams already have running. The fix isn't more frequent checks on the same fixed prompt set, it's checks designed to detect a change in the underlying system, not just a change in the output. That's a genuinely different kind of monitoring than most reporting and analytics stacks were built for when AI Mode first launched as a single, presumably stable system.

It also strengthens the case, already made in our review of the H1 2026 market-share data, for treating each engine's AI surface as a genuinely separate, moving target rather than a stable channel you optimize once and monitor passively. A B2B SaaS team running a quarterly GEO reporting cycle should build a standing question into that cycle now: has Google changed anything about the model layer since the last report, and if so, does last quarter's trend line still mean what it looked like it meant. Answering that honestly, quarter over quarter, is a smaller lift than it sounds, and it's the difference between reacting to real content performance and reacting to a model Google swapped out without telling anyone.

The nine-month pattern, three changes to the model or measurement layer inside Google's AI search surfaces, argues this isn't the last quiet swap either. Building the time-series habit now, before the next one lands, is cheaper than rebuilding a team's trust in its own dashboard after a fourth unexplained shift shows up with no changelog attached.

There's a broader lesson here that extends past AI Mode specifically. Every engine worth tracking now carries some version of this same instability, a live-versus-cached retrieval split, a model that gets swapped without notice, a measurement surface that only shows part of the picture. Treating any single engine's citation behavior as a fixed target to be measured once and monitored passively was already a fragile assumption before this specific model addition; Gemini 3.7 Flash's arrival inside AI Mode is just the clearest, most recent example of why that assumption keeps failing in practice. A GEO program built around continuous, model-aware measurement is more expensive to run than one built around a quarterly snapshot, but it's the only version that survives contact with how these systems actually change.

None of this is a reason to distrust AI citation data wholesale, or to stop investing in the content and structural work that earns citations in the first place. It's a reason to be more careful about what a single data point is allowed to mean. A citation-rate chart with a sudden dip looks like a problem to fix. A citation-rate chart with a sudden dip, annotated against a known model change two days earlier, looks like exactly what it is: noise from an infrastructure shift, not a signal about the content underneath it. The annotation is the whole difference, and it only exists if someone was tracking model-layer changes in the first place. Building that habit now, while the current pattern is still just three data points instead of ten, is the cheap version of a fix a team will otherwise be forced to build under pressure later, after a real content problem gets misdiagnosed as one more unexplained model swap.

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.