What was the state of play before Gemini 3?
Before January 2026, Google’s AI answers ran on earlier Gemini versions with visibly inconsistent citation behavior: the same question could produce different source sets day to day, and the relationship between AI Overviews (the inline panel) and AI Mode (the conversational surface) was ambiguous — two surfaces, overlapping but not identical source pools. Publishers optimized for a moving target and often measured the wrong one.
What changed with Gemini 3 as default?
Three shifts consolidated over the first 90 days and have held since:
- Fan-out went deeper. Query fan-out — decomposing a question into many sub-queries — expanded visibly. Head-term optimization mattered less; coverage of the sub-question tree mattered more. Pages built as answer-block collections around a fan-out map gained citations across whole conversations. See query fan-out.
- Grounding got stricter. Cross-checking hardened: claims contradicted by third-party sources, or dated inconsistently across meta, schema and page, disappeared from answers more reliably. Fact consistency moved from best practice to gate. See grounding fact consistency.
- Citation stabilized. Day-to-day source-set volatility dropped, which made monthly measurement meaningful: a frozen 50-100 question set now produces comparable runs, enabling the four-dimension scoring discipline instead of anecdote collection.
What did the I/O 2026 merger actually merge?
At Google I/O 2026, Google announced AI Overviews and AI Mode as one AI answer system — same model (Gemini 3, default since January), same retrieval-grounding-generation-citation pipeline, two interfaces. The practical consequence, confirmed over the following months: optimizing for one surface is optimizing for both. Alphabet’s Q2 2026 report put usage above 2.5 billion monthly users, so the merged system is simply where search answers happen. The full mechanics are in our AI Mode analysis.
What did not change?
More than the hype suggested. Crawlability and indexation still gate everything — the unfetched page is still the uncited page. Classic organic results still exist alongside AI answers, and the pages retrieval prefers remain the technically sound ones. Google-Extended still governs training only, with no effect on citations. And the Ahrefs February 2026 baseline still frames the economics: where AI answers appear, position-1 CTR drops 58% and roughly 68% of sessions end zero-click. Gemini 3 changed the citation game’s rules of precision, not its stakes — those were set by the interface shift itself.
It is also worth noting what consolidation does for experimentation. With a stable default model and calmer citation behavior, controlled content tests on Google’s AI surfaces finally produce interpretable results: a passage rewritten answer-first in week one shows up — or does not — in the same month’s question-set run. That feedback loop simply did not exist on the volatile pre-2026 surfaces, and it is quietly the most useful thing Gemini 3 gave practitioners.
What does the 90-day evidence say publishers should do?
- Audit your fan-out coverage: take your ten most valuable questions, map their sub-question trees, and check which sub-answers you actually win.
- Run a consistency pass on dates, names and numbers across your top pages — grounding strictness makes this the cheapest citation defense available.
- Re-baseline your question set against the merged system; pre-merger baselines mix two source pools and are not comparable.
- Stop maintaining separate AI Overviews and AI Mode workstreams; one strategy covers both surfaces.
The meta-lesson of Gemini 3’s first months is reassuring: no exotic new tactic emerged. The pipeline became more demanding of exactly the things the discipline already prescribed — clean passages, verifiable facts, consistent entities, honest measurement. Teams that treated those as fundamentals absorbed the model change in a quarter; teams chasing surface quirks are still re-optimizing. The next model update will reward the same boring list.