Gemini in Google Search got another upgrade on September 2, 2026: Google’s Robby Stein, VP of Product for Search, announced on X that Gemini 3.8 Flash is now selectable in AI Mode for Google AI Pro and Ultra subscribers worldwide. The same day, Google published its Gemini 3.8 Flash launch post, listing AI Mode alongside the Gemini app, Sheets and developer tools as launch surfaces. It replaces Gemini 3.7 Flash in the picker, which only arrived on August 14. AI Overviews are not mentioned in either announcement.
- Confirmed: Gemini 3.8 Flash is live in AI Mode as an opt-in model for paid Google AI Pro and Ultra subscribers, English first, globally.
- Not confirmed: any change to AI Overviews or to the default (free-tier) AI Mode model, which Google set to Gemini 3 in January 2026.
- Google says 3.8 Flash beats 3.7 Flash on multi-step reasoning, agentic tasks and coding at the same speed and price tier.
- Within a day, SEOs spotted the new model returning far fewer links; Stein called it a bug and a fix rolled out by September 4.
- This is Google’s third Flash model in Search in six weeks, so treat “which model answered” as a variable in every AI Mode test you run.
What Google announced about Gemini in Google Search
Stein’s post says 3.8 Flash delivers “significant improvements from 3.7 Flash across software engineering, agentic tasks, and multi-step reasoning.” Subscribers reach it by tapping the plus icon in AI Mode’s “Ask anything” bar and picking it from the model dropdown, where it sits under “Gemini 3 models” between Auto and Pro. Free-tier users still get the default model with no picker.
The launch post from Tulsee Doshi and Raluca Ada Popa frames the model as Google’s “best reasoning and coding model yet, at the same speed and low cost of 3.7.” The headline mechanism is that 3.8 Flash “works harder”: it runs extra reasoning steps and calls tools iteratively, which, per The Register’s analysis, means roughly 40 percent more tokens per task than 3.7 Flash. Benchmark deltas over 3.7 Flash are modest but consistent: HLE-Verified 54.9 percent vs 53.6, Vals Finance Agent V2 61.4 vs 59.0.
Google’s DeepMind model page lists a 1M-token input window, text, image, video, audio and PDF inputs, and “search as a tool” support.
Here is how the model line-up behind Search looks as of September 2026, based on Google’s announcements to date:
| Surface | Model | Who gets it | Source / status |
|---|---|---|---|
| AI Overviews (default) | Gemini 3 | Everyone, global | Google blog, Jan 27, 2026 (confirmed) |
| AI Mode (default / Auto) | Gemini 3 | Everyone, global | Google blog, Jan 27, 2026 (confirmed) |
| AI Mode (picker: Pro) | Gemini 3 Pro | AI Pro and Ultra | Nov 2025 (confirmed) |
| AI Mode (picker: Flash) | Gemini 3.8 Flash, replacing 3.7 Flash | AI Pro and Ultra, English first | Stein on X, Sept 2, 2026 (confirmed) |
| AI Overviews on 3.8 Flash | Not announced | — | Unconfirmed |
The citation bug
On September 3, Gagan Ghotra and Glenn Gabe posted screenshots showing AI Mode on 3.8 Flash answering top-of-funnel queries with few or no links, next to the default model showing a normal set. Stein replied: “This isn’t working as intended, and we’ll roll out a fix soon.” Search Engine Roundtable reported links back to normal by the morning of September 4. Nobody published a measured citation rate, so treat the whole episode as an observation, not a dataset.
What the AI Mode Gemini update means for SEOs
My honest read: the model itself is not the story. The story is the cadence. Gemini 3.5 Flash-Lite in July, 3.7 Flash on August 14, 3.8 Flash on September 2. Search is now on the model release schedule, announced from a VP’s personal X account rather than Search Central.
Second, the citation bug is the most useful thing that happened this week, because it proved how thin the link layer is. A model swap made links vanish for a day and no crawl, index or ranking signal on your side could have changed that. Everything I wrote about AI Overviews reshaping SEO still holds, with one addition: your visibility inside Gemini in Google Search now depends on which model a user selected and whether that model’s grounding is behaving that day.
Third, a model that “works harder” and calls tools iteratively should, in theory, fan out more sub-queries per question. That rewards pages that answer one narrow sub-question cleanly over pages that try to cover everything. That is the same argument behind generative engine optimization and the AEO vs GEO playbook, and I would not change tactics for this release. I would change how I measure.
Finally, this is paid-tier only. In the client accounts I look at, AI Mode traffic is already a rounding error in Search Console, and the subscriber slice of it is smaller still. Any “AI overviews model update” panic this week is premature; AI Overviews did not change on September 2 as far as anyone has shown.
What to do this week
- Log the model name with every AI Mode test. If you are tracking brand mentions or citations, record Auto, Pro, or 3.8 Flash alongside the date, or your September data will be uninterpretable.
- Re-run your top 20 tracked queries on 3.8 Flash and Auto side by side and note citation counts. You now have a baseline for the next swap, which on current form is three to four weeks away.
- Check the Search Console Performance report for September 3 and 4 with the AI Mode filter, if you have enough volume to see it. Look for a link-click dip on paid-heavy audiences (dev tools, B2B SaaS).
- Audit your most-cited pages for extractability: a one-paragraph direct answer at the top, an H2 per sub-question, and clean schema markup for AEO.
- Update your audit report template with a “model version” field in the AI visibility section so clients understand why numbers move without any change on their site.
What we don’t know yet
Google has not said whether 3.8 Flash will become the default AI Mode model or reach AI Overviews. Stein said only that Google is “looking to bring these capabilities to more people as we can.” There is no stated timeline for languages beyond English, no root-cause explanation for the citation bug, and no published data on whether the fixed version cites the same number of sources as the default.
FAQ
Which Gemini model powers Google AI Overviews right now?
Gemini 3, per Google’s January 27, 2026 announcement that made it the default for both AI Overviews and AI Mode globally. Gemini 3.8 Flash has only been announced for the AI Mode model picker, not AI Overviews.
Can free Google users try Gemini 3.8 Flash in Search?
No. As of September 2026 the model picker in AI Mode is limited to Google AI Pro and Ultra subscribers, and the rollout is English first. Free users get the default model with no selector.
Does Gemini 3.8 Flash show fewer citations in AI Mode?
It did for roughly a day after launch. Google’s Robby Stein confirmed it was not working as intended, and reports on September 4 said links had returned. No one has published measured before-and-after citation rates, so ongoing differences are unproven.
