What model powers Google AI Mode?

Google Search’s conversational AI tab.

As of July 2026, Google AI Mode answers most questions with Gemini 3.5 Flash — its default globally since Google I/O 2026 in May — and routes complex questions to the heavier Gemini 3.x Pro. The choice is made per query, server-side, and Google never labels which model wrote a given answer. So AI Mode has no single per-answer model: the honest description is the routing, which is exactly the point.

Current model
Gemini 3.5 Flash (default) · Gemini 3.x Pro (complex queries)
Evidence
Vendor-stated
As of
May 2026
What we capture
google.com’s AI Mode answer

The vendor discloses this default; the engine never labels which model produced a given answer, so we report the vendor’s word with its date, not an observation.

How AI Mode picks its model

AI Mode doesn’t use one model per answer — it triages each question. Google’s own description is that most queries are handled by the default Gemini 3.5 Flash, while complex questions are escalated to Gemini 3.x Pro. That decision happens server-side, per query, and is invisible to you. Because of that, asking “which model wrote this AI Mode answer?” is the wrong question in principle: the same prompt can be routed differently on different days.

Under the hood, AI Mode answers through query fan-out: one question is decomposed into many sub-queries that run against Google’s index in parallel, and the answer is synthesized across all of them with link cards to the sources it used. Google changes the underlying models on its own schedule — it has already moved AI Mode from a custom Gemini 2.0 at launch to Gemini 3, then Gemini 3.5 Flash.

Model history

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

Custom Gemini 2.0
LaunchVendor-stated

Google launches AI Mode as an experiment in Labs, powered by a custom version of Gemini 2.0.

Google ↗ (opens in a new tab)
Gemini 3 (Flash / Pro)
Model changeVendor-stated

Google brings Gemini 3 (Flash and Pro) to Search and AI Mode on day one.

Google ↗ (opens in a new tab)
Gemini 3.5 Flash
Model changeVendor-stated

Gemini 3.5 Flash becomes the default model in AI Mode for everyone globally.

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

    Gemini 3.5 Flash becomes the default model in AI Mode for everyone globally.

    Google ↗ (opens in a new tab)
  2. Gemini 3 (Flash / Pro)

    Google brings Gemini 3 (Flash and Pro) to Search and AI Mode on day one.

    Google ↗ (opens in a new tab)
  3. Custom Gemini 2.0

    Google launches AI Mode as an experiment in Labs, powered by a custom version of Gemini 2.0.

    Google ↗ (opens in a new tab)

What decides whether Google AI Mode names your brand

RAG plus “query fan-out”: one question is split into concurrent sub-queries against Google’s index, always Search-grounded — so the unit of competition is the sub-query, not the head term.

How answers form & what drives brand mentions

RAG + query fan-out. VENDOR-DOCUMENTED AI Mode uses Google’s core Search ranking to retrieve relevant current pages, plus “query fan-out” — concurrent model-generated related sub-queries — then synthesizes (Google’s worked example fans a lawn-weeds question into herbicide, chemical-free, and prevention sub-queries). Beyond web pages, inputs include the Knowledge Graph, Shopping Graph (commerce), and Maps/local data (per query/market/account).

Eligibility. VENDOR-DOCUMENTED A supporting page must be indexed and snippet-eligible; no special markup, no AI files (Google states Search ignores llms.txt); preview controls (nosnippet/max-snippet/noindex) and a Search Console generative-AI inclusion setting govern appearance.

Vendor myth-busting. VENDOR-DOCUMENTED Google’s 2026-05 guide explicitly lists what does not work: llms.txt/AI files (ignored), content chunking (unnecessary), rewriting for “AI phrasing” (unnecessary), over-focus on structured data (not required for AI features), and inauthentic mention-seeking (spam-adjacent) — and warns that mass-producing pages to chase fan-out variations violates the scaled-content-abuse policy.

What fan-out changes. INDEPENDENT STUDY The citation pool extends well beyond the head query’s top-10. Moz’s study of 40,000 AI Mode queries found only ~12% of AI Mode citations match the organic top-10 URLs — ~88% fall outside it; SE Ranking and Ahrefs separately put head-query top-10 overlap near 12–14%. OUR INFERENCE: the practical unit of competition is the sub-query, not the head query.

Brand-signal coupling. INDEPENDENT STUDY Ahrefs’ 75,000-brand update found AI Mode’s strongest correlations with branded-authority signals of the three Google surfaces (branded anchors ~0.63, branded search volume ~0.47), YouTube the top single correlate (~0.737), and high output overlap across ChatGPT/AI Mode/AIO (~0.78) — correlation, not causation.

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 index + snippet eligibility (preview controls)High (binary gate)HighVENDOR-DOCUMENTED
Retrievability via core Search ranking across fan-out sub-queriesHighHighVENDOR-DOCUMENTED mechanism; sub-query emphasis OUR INFERENCE + INDEPENDENT STUDY
Head-term top-10 rank for the original queryMedMed-HighINDEPENDENT STUDY (Moz ~12% overlap; SE Ranking/Ahrefs ~12–14%); correlated, no longer decisive
Knowledge Graph / entity data (entity questions)High (entity)HighVENDOR-DOCUMENTED
Shopping Graph / Merchant Center + Business Profile (commerce/local)High (vertical)HighVENDOR-DOCUMENTED
Helpful, non-commodity, first-hand contentHighMed-HighVENDOR-DOCUMENTED (2026-05 guide’s central recommendation)
Branded authority + YouTube presenceMed-HighMedINDEPENDENT STUDY (Ahrefs 2025-12); correlational
Freshness (freshness-sensitive fan-out queries)High (conditional)MedOUR INFERENCE from documented freshness systems
Generic structured data as a direct boostLowHighVENDOR-DOCUMENTED (not required; must match visible text)

What you can do (ranked)

  1. Cover the fan-out tree, not the head termengine-specific emphasisVENDOR-DOCUMENTED mechanism + boundary

    Build genuinely strong pages for the sub-questions (comparisons, setup, pricing, alternatives, how-it-works) — as consolidated quality coverage, because Google warns scaled page-per-variation production is spam.

  2. Hold eligibility basicscross-engine within GoogleVENDOR-DOCUMENTED

    Indexed, snippet-eligible, no accidental nosnippet/max-snippet, and check the Search Console generative-AI inclusion setting (it gates AI Mode and AIO together).

  3. Invest in unique, first-hand, non-commodity contentcross-engine · vendor-endorsedVENDOR-DOCUMENTED

    The 2026-05 guide makes a unique viewpoint its lead recommendation; for an SMB: first-party data, real implementation detail, named expertise.

  4. Build branded footprint off-site, incl. YouTubecross-engineINDEPENDENT STUDY + VENDOR-DOCUMENTED boundary

    Earned coverage and genuine video; Google warns against inauthentic mention-seeking.

  5. Commerce/local feed + profile hygieneGoogle-specificVENDOR-DOCUMENTED

    Current Merchant Center feeds and Business Profile.

  6. Measure with the Search Console Generative AI reportengine-specificVENDOR-DOCUMENTED tooling

    Pair with question-sample mention tracking (the report covers your site’s citations, not your brand’s mentions on others’ pages).

Grounding, citation & locality

VENDOR-DOCUMENTED Always Search-grounded; responses embed prominent supporting links, and Google frames link diversity as a design goal. Locality: runs inside Search’s market/language/location systems; local intents draw on Business Profile/local results; connected-app actions were added July 2026 (perishable). Verify per target market before regional claims.

INDEPENDENT In one Sept-2025 US paired sample ~3% of captured AI Mode answers lacked a visible citation — grounding and visible display are different measurement fields.

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 ›