On 2 September Google released Gemini 3.8 Flash and, unusually, put the new model into AI Mode in Google Search on launch day. Nick Fox, who runs Google Search, announced on X that the model was “available today in AI Mode in Google Search and the Gemini app for Google AI Pro & Ultra subscribers”. Robby Stein, VP of Product for Search, called it “our most intelligent workhorse model yet”.
Within a day the SEO community noticed something else: answers from the new model were coming back without links. Google’s Stein confirmed on 3 September that “this isn’t working as intended” and promised a fix, which Search Engine Roundtable reports was in place by the morning of 4 September. The episode is small, but it shows how much of a site’s visibility in AI Mode now depends on a model that changes every few weeks.
Context: three Flash models in six weeks
Google’s own announcement, written by Tulsee Doshi and Raluca Ada Popa, calls 3.8 Flash “our third Flash release in only six weeks”. Search Engine Journal dates the previous model, Gemini 3.7 Flash, to 14 August, three weeks earlier. The line of Flash models is the one that matters for search: Search Engine Land reported in May that Gemini 3.5 Flash, unveiled at I/O, became the model behind AI Mode globally, because a distilled Flash model is what Google can run at search scale with acceptable latency and cost.
What changed with 3.8 is the timing. Search Engine Journal points out that AI Mode used to receive a new model days after the Gemini app; this time Search was part of the launch itself. That makes Google’s model release calendar a search calendar too.
What Google says the model does
According to the announcement, 3.8 Flash delivers “significant improvements from 3.7 Flash across software engineering, agentic tasks, and critical, multi-step reasoning in specialized domains”. Google says the model “works harder” on complex requests by “executing extra reasoning steps, and calling tools iteratively”, which is a description of an answer engine that runs more searches before it writes. On Google’s HLE-Verified reasoning benchmark the company reports a score of 54.9%.
The Register, citing the Artificial Analysis Intelligence Index, puts the model at 59 points, three above its predecessor and tied with GPT-5.6 Sol and Grok 4.6 on that index, while noting it remains behind the top Anthropic models. Google is selling it on price: the announcement lists introductory API rates of $0.75 per million input tokens and $3.75 per million output tokens until 31 December 2026, and The Register describes it as the cheapest model at its level of intelligence.
A second variant, Gemini 3.8 Flash Cyber, is restricted to vetted defenders through Google’s Fairwind Program and does not appear in Search.
Who actually sees it
The rollout is narrower than the headlines suggest. Per Search Engine Land, 3.8 Flash is available in AI Mode to Google AI Pro and Ultra subscribers worldwide, in English. Subscribers choose it from the ”+” icon in the “Ask anything” bar, where Search Engine Journal locates it between the “Auto” and “Pro” options. Free users keep the default model and have no selector.
Search Engine Land expects Google to extend the model to all searchers “in the coming weeks or months”, as it did with previous Flash versions. Until then, two different AI Mode users can get two different answers, from two different models, to the same query.
The missing links
On 3 September Gagan Ghotra posted examples on X of top-of-funnel queries answered by 3.8 Flash with no citations at all, and Glenn Gabe confirmed the pattern with side-by-side screenshots against the default model, asking “What is going on here?”. Search Engine Roundtable collected the reports. Stein replied the same day: “Hey — thanks for flagging. This isn’t working as intended, and we’ll roll out a fix soon.” By 4 September the links were back.
Google has spent the past year adding link cards, inline citations and subscription labels to AI answers, so a model that stopped linking was a regression rather than a policy. But it went to paying subscribers for about a day, and nobody outside the SEO community would have known.
What it means for your business
Your AI Mode visibility now has a version number. Search Engine Journal’s practical advice is the one we would repeat: if you track how AI Mode describes your brand or your category, log which model answered. A comparison between a test run on 3.7 Flash in August and one on 3.8 Flash in September measures the model change, not your content. In our SEO work we treat every Google model release as a reset point for AI visibility tracking, the same way we treat a core update for rankings.
Citations are a feature Google chooses to ship. The bug was fixed within a day, and Google’s public direction is towards more links, not fewer. But the episode is a reminder that a citation in an AI answer is not a ranking you earned once and keep. The way to stay cited is to be the page the model needs: a clear, specific answer to the question, with the facts the model would otherwise have to assemble from three sources. That is what a search-focused content programme produces, and it is also what earns the classic blue-link click.
Measure it in Search Console, not by eye. Google’s AI performance report went live worldwide last week, as we covered on 5 September. It is the only first-party source for how often your pages appear in AI Overviews and AI Mode. If a future model drops your citations, the impressions line will show it before anyone’s screenshot does.
Expect the next model in weeks. Three Flash releases in six weeks is the cadence Google has chosen. Our view is that businesses should stop treating AI Mode as a fixed surface and start treating it as software with a release cycle: check the AI report after each release, and keep the content itself stable rather than rewriting pages to chase the answer style of the current model.
The more interesting question is not whether 3.8 Flash cites less or more than 3.7. It is that a model swap can change who gets cited overnight, and that Google has made Search part of the launch itself. That is the environment SEO now lives in.
