What model powers Gemini?

Google’s Gemini app and assistant (gemini.google.com).

As of July 2026, the free Gemini app answers with Gemini 3.5 Flash — the model Google made its global default for the app at Google I/O 2026 (May). Discoverably also reads the model directly from Gemini’s own answers, so this reflects the live free experience, not only the announcement. Paid tiers add a daily allotment of the heavier Gemini 3.x Pro.

Current model
Gemini 3.5 Flash
Evidence
Observed
As of
May 2026
What we capture
gemini.google.com, the free default

Discoverably reads this model from the engine’s own answers, so it reflects what the free experience actually returns — not only what the vendor last announced.

How Gemini picks its model

The free Gemini app runs a default model that Google auto-selects; you don’t choose it. Gemini 3 Flash was the app default from December 2025, and at Google I/O 2026 (May) Google made Gemini 3.5 Flash the app’s global default. Discoverably also reads the model from Gemini’s own answers, so this reflects the live default and updates as Google changes it.

For many questions Gemini also grounds its answer in Google Search: it decides, per prompt, whether live results would help, then composes the reply from what Search returns. Google can change the default model at any time; Discoverably follows it rather than pinning a version.

Model history

Every family-level change to the model behind Gemini, newest first, each linked to its primary source.

Gemini 3 Flash
Model changeVendor-stated

Gemini 3 Flash becomes the stated default model in the Gemini app.

Gemini release notes ↗ (opens in a new tab)
Gemini 3.5 Flash
Model changeObserved

Gemini 3.5 Flash becomes the Gemini app’s default model globally (Google I/O 2026).

Gemini 3.5 Flash model card ↗ (opens in a new tab)
  1. Gemini 3.5 Flash

    Gemini 3.5 Flash becomes the Gemini app’s default model globally (Google I/O 2026).

    Gemini 3.5 Flash model card ↗ (opens in a new tab)
  2. Gemini 3 Flash

    Gemini 3 Flash becomes the stated default model in the Gemini app.

    Gemini release notes ↗ (opens in a new tab)

What decides whether Gemini names your brand

A model-first assistant: most answers come from training (parametric), with Google Search grounding invoked selectively — so a brand is earned in the corpus months before it’s cited.

How answers form & what drives brand mentions

Mixed path. VENDOR-DOCUMENTED Google says some Gemini Apps responses are grounded on Google Search, and Gemini can use public Search information even when signed out; it does not publish a deterministic consumer grounding trigger. The API documents the mechanism — the model decides whether a search would help, generates one or more queries, retrieves, and returns inline citations. OUR INFERENCE: the consumer app uses the same decide-then-search mechanism; current-events, pricing, and volatile-fact queries ground far more than evergreen ones.

Sources are conditional. VENDOR-DOCUMENTED Gemini “sometimes” shows sources; a related link “is not necessarily what the Gemini app used,” and “Double-check” runs a post-hoc Search comparison — not the original retrieval trace.

Two brand-selection layers. OUR INFERENCE Parametric: whether the model already associates the brand with its category (pre-cutoff web/Wikipedia/third-party coverage). Grounded: when Search fires, visibility inherits Google’s ranking systems, so Google organic strength and entity presence carry over. Parametric answers show no sources, so they are earned months earlier in the corpus, not at answer time.

Training/grounding control. VENDOR-DOCUMENTED Google-Extended (robots token) controls use of a site’s content for Gemini training/grounding in Google’s non-Search AI; it does not affect Search AI features (AI Overviews/AI Mode).

Evidence is thinnest for Gemini. OUR INFERENCE Most 2025–2026 citation studies focused on the other engines, and Gemini-app-specific citation data is sparse and often conflated with AI Overviews. We flag Gemini as the engine where our factor weights rest most on inference.

Factor weights

Weight and confidence are our editorial estimate of each factor’s influence, and of the evidence for its direction — not a vendor’s ranking weight. No engine publishes a brand-selection formula.

FactorWeightConfidenceBasis
Google organic ranking strength (inherited when grounding fires)High (grounded only)Med-HighVENDOR-DOCUMENTED mechanism; weight OUR INFERENCE
Training-corpus prominence (pre-cutoff third-party/Wikipedia coverage)HighMedOUR INFERENCE (mechanism follows model-first design)
Query type as a regime switch (volatile→grounded, evergreen→parametric)High (regime)MedVENDOR-DOCUMENTED mechanism; thresholds undisclosed
Allowing Google-Extended (eligibility for Gemini training + grounding corpus)Med-High (gate-like)HighVENDOR-DOCUMENTED
Accurate entity/Maps/Business facts across Google surfacesHigh (local/travel)MedOUR INFERENCE
Branded web + YouTube mentionsMedMedINDEPENDENT STUDY (Ahrefs 2025-12: YouTube ~0.737 top correlate across Google surfaces; Gemini app not separately broken out)
Freshness of site contentLow-Med (grounded only)MedOUR INFERENCE
Schema/structured data as a direct Gemini-app factorLowLowOUR INFERENCE; no vendor doc of schema use in app answer selection

What you can do (ranked)

  1. Win the grounded layer through Google organic strengthcross-engine (shared with AI Mode/AIO)VENDOR-DOCUMENTED mechanism

    The indexation/snippet/ranking work that serves AI Mode and AIO serves grounded Gemini answers.

  2. Keep Google-Extended allowed on public marketing contentengine-specificVENDOR-DOCUMENTED

    Blocking it removes content from Gemini training/grounding while doing nothing for Search AI features; allowing it on marketing pages is near-costless.

  3. Invest in durable third-party coverage for the parametric layercross-engineINDEPENDENT STUDY

    Long-lived reference/trade/YouTube coverage is what future snapshots absorb; Ahrefs found off-site brand signals (YouTube ~0.737, branded mentions ~0.66–0.71) correlate with AI visibility far more than on-site metrics — correlation, not causation.

  4. Establish clean entity datacross-engineOUR INFERENCE

    One canonical name, consistent org facts, accurate Business Profile; reduces entity confusion in retrieval and generation.

  5. Target the query classes that groundengine-specificOUR INFERENCE

    Comparison/pricing/“current best” questions trigger retrieval, the only layer influenceable quickly.

  6. Separate generation, grounding, and Double-check in measurementengine-specificVENDOR-DOCUMENTED

    A related link may not be a generation source; do not score Double-check as an original citation.

Grounding, citation & locality

VENDOR-DOCUMENTED Not every response includes sources; a related link may not be what Gemini used; long verbatim quotes always get a link; Double-check marks corroborated vs uncorroborated statements. Locality: at minimum a general area from IP or Home/Work; precise device location and prior activity are opt-in. OUR INFERENCE: citation rate and ordering must be measured by signed-in state, region, language, activity settings, and whether Search grounding displayed — a single “Gemini rank” is not portable.

Sources

Three steps to get started.

A couple of hours a month is enough to make real progress.

Enter your website.

We look up your business and the questions your customers actually ask AI.

Review what we found.

Confirm your business, your questions and your competitors. Then we ask the live apps your customers actually use: ChatGPT, Gemini, Google AI Overviews, AI Mode, Copilot and Perplexity. Not the developer API, and not the model's training data. Both give a different answer from the one your customer sees.

Get your free report.

Where you appear, who wins instead, and how AI describes you, in plain English with clear next steps. A one-time snapshot to start from.

See what AI says about you.

Get a free report: no sign-up, no credit card, results in minutes.  ·  See a sample report ›