Google swapped models in AI Mode, and the citations vanished for a day

Google put its new Gemini 3.8 Flash model into AI Mode on September 2, 2026, and within a day practitioners noticed it had all but stopped linking to the sources behind its answers, a behaviour Google called unintended and fixed on September 4. The episode lasted roughly twenty-four hours, it only touched an optional model reserved for paying subscribers, and it still makes the point that matters: whether an AI answer names your business depends on a component you do not control and that Google replaces every few weeks.

What actually happened between September 2 and September 4, 2026?

Google announced Gemini 3.8 Flash on September 2, 2026 and made it selectable the same day inside AI Mode, the conversational mode of Google Search, for Google AI Pro and Ultra subscribers. On the morning of September 3, Search Engine Roundtable reported observations from two practitioners, Gagan Ghotra and Glenn Gabe, who had posted screenshots on X comparing the default AI Mode model with Gemini 3.8 Flash on identical queries. The default model showed its sources. The new one answered with no links and no citations.

On September 4, 2026, Robby Stein, VP of Product for Google Search, replied publicly that the behaviour was not working as intended and that a fix would roll out shortly. Glenn Gabe confirmed the same morning that Gemini 3.8 Flash was showing links again. The nature of the evidence deserves to be stated plainly: it rests on screenshots of a handful of top-of-funnel queries, posted by two well-known practitioners, not on a measurement at scale. No sample size, no percentage and no volume were ever published.

That gap is part of the story. Nobody, Google included, has quantified how many citations disappeared during those twenty-four hours. A site that lost every one of its AI Mode citations that day had no way of noticing at the time, and no way of proving it afterwards.

Why does swapping a model make links disappear?

An answer engine such as AI Mode is not a ranking, it is a model that writes. Which sources get cited, how many of them, and how they are displayed beneath the answer are properties of the model and the product layer around it, not of a ranking algorithm that holds steady over time. Swapping the model therefore changes how sources surface, even when the underlying index and the pages themselves have not moved a byte.

The cadence is the real story. Gemini 3.8 Flash arrived three weeks after Gemini 3.7 Flash, according to 9to5Google, and Search Engine Journal draws the operational consequence for anyone tracking their presence in AI answers: the model options available in an AI Mode test can change inside a single reporting cycle. A month-on-month swing may come from your site, or simply from whichever model answered.

This continues a pattern described here in "Google is putting links back into AI answers. The clicks are not coming back." Google keeps adjusting how much room links get inside generated answers, sometimes deliberately restoring them, sometimes removing them by accident. Either way the decision sits with Google, and it moves fast.

What does this change for a small business?

Almost nothing this week, and quite a lot about how you steer. The September 3, 2026 incident hit an optional model that a small business owner is unlikely to be using: Gemini 3.8 Flash has to be picked by hand inside AI Mode and requires a Google AI Pro or Ultra subscription. Your customers mostly meet AI Overviews instead, the generated summaries sitting above the standard Google results, which reached France on July 22, 2026 according to Abondance and are now live across every major market.

The real issue is dependency, and its cost shows up in traffic. Ahrefs measured across 300,000 keywords, comparing December 2023 with December 2025, a 58% lower click-through rate for the top organic position when an AI overview is displayed. The share of your visibility decided inside the answer rather than inside the list of links keeps growing, while the control you hold over that answer stays at zero.

Does this mean measuring AI visibility is pointless?

No, but it forces you to measure differently and to accept that the signal is noisy. A tracker that queries an AI once a week and hands you a score without naming the model that answered is selling a precision that does not exist. Search Engine Journal puts the useful rule simply: log the model name with every test. A drop in citations that coincides with a model change is not a drop in your site.

The same caution applies to Google's own tooling, as we wrote in "Google just opened its AI visibility report to every site. It counts impressions, not clicks." Those reports count impressions, blend different surfaces together and never say which model produced the answer. They are useful as a trend, never as fine-grained proof of a cause.

One part of the equation does not move when the model does, and that is the part worth your budget. Whether an AI cites three sources or none, it can only draw on pages it was able to reach, read, understand and attribute to an identifiable company. A site that answer-engine crawlers cannot reach, with no content of its own and no page stating plainly who you are and what you sell, is a candidate for no citation at all, whichever model is shipping this month.

What should you do, and what should you leave alone?

• Change nothing on your site because of this incident: it lasted twenty-four hours and Google described it as unintended behaviour.
• If you track your presence in AI answers, record the date and the model name for every reading, or you will credit your own site with swings that came from Google.
• Treat any isolated drop on a single surface as a hypothesis rather than a fact until it holds across several weeks and several engines.
• Spend on what does not depend on the model: crawler accessibility for answer engines, content that answers real buyer questions, and an explicit, verifiable company identity.
• Do not chase a guaranteed citation. No vendor can promise one, and September 3, 2026 shows exactly why.

A closing note that cuts against our own trade: if your site has nothing worth citing, no reputation and no page explaining what you do, measuring your AI visibility will only tell you that you are absent. That budget belongs in content and in technical accessibility before it belongs in tracking.