TL;DR
- Search Engine Roundtable reported on 3 September 2026 that AI Mode running Gemini 3.8 Flash is not link or citation friendly, citing and linking to sources less often than previous models.
- Gagan Ghotra posted numerous AI Mode searches on X for top-of-the-funnel queries in which none showed any links or citations, and Glenn Gabe posted a side-by-side of the default model against 3.8 Flash on 3 September 2026.
- Robby Stein, VP of Product for Google Search, replied on X on 4 September 2026: "This isn't working as intended, and we'll roll out a fix soon." No date, scope or mechanism for that fix was given.
- No measured citation rate has been published for either model, and 3.8 Flash is selectable only by Google AI Pro and Ultra subscribers, so free-tier AI Mode answers are outside what was reported.
Google's AI Mode is showing far fewer links and citations when it runs on Gemini 3.8 Flash than it did on the model it replaced, and Google has said the behaviour is a bug. Search Engine Roundtable reported the pattern on 3 September 2026, and on 4 September 2026 Robby Stein, VP of Product for Google Search, replied on X: "This isn't working as intended, and we'll roll out a fix soon."
The reports concern the model that paying subscribers can select for themselves, not the default answer that every AI Mode user receives. Stein announced the 3.8 Flash rollout on X, saying the model is available in AI Mode for Google AI Pro and Ultra subscribers around the world and is chosen from the "+" icon in AI Mode's "Ask anything" bar. Free-tier AI Mode users get no model choice at all. They receive the default, and nothing in this story says the default changed.
What the reports actually describe
Barry Schwartz wrote at Search Engine Roundtable on 3 September 2026 that early feedback on AI Mode running Gemini 3.8 Flash was that it "is not link or citation friendly", meaning the model is less likely to cite sources and link to them than previous models did. Google had rolled the model out for subscribers in AI Mode and some users had started testing it, which is how the pattern surfaced within days rather than weeks.
Two sets of observations sit behind that summary. Gagan Ghotra posted numerous examples on X of searches run in AI Mode for top-of-the-funnel queries in which none of them showed any links or citations. Glenn Gabe posted a side-by-side on X on 3 September 2026 comparing the default model with 3.8 Flash and said the difference between the two was visible. Schwartz wrote that he was not sure why this would be, and that he hoped Google could fix it soon.
Nobody in that reporting offered a cause. Neither Google nor the people running the queries said what produced the drop in links and citations, and no explanation has been attached to it since.
Google's reply came the next day
The update on the Search Engine Roundtable story is dated 4 September 2026 and carries Stein's reply on X, thanking the reporter for flagging it and stating that this is not working as intended and that a fix will roll out soon. That is a short reply, and it is worth reading for exactly what it contains. It confirms that Google considers the reduced linking a defect rather than a design decision. It commits to a fix. It does not give a date, it does not describe the scope of the fix, and it does not say what the fix changes underneath. "Soon" is the only timing anyone has.
That matters when you are deciding whether to act. A confirmed defect with an uncommitted release date is not something a publisher can plan a content change around, and there is no version number to watch for.
The sequence, as reported
The whole story fits inside two days, which is worth laying out before anyone builds a narrative on top of it. Every row below comes from the Search Engine Roundtable report and its update.
| Date | Who | What was reported |
|---|---|---|
| Reported 3 September 2026 | Barry Schwartz, Search Engine Roundtable | Early feedback that AI Mode on Gemini 3.8 Flash "is not link or citation friendly", less likely to cite and link than previous models |
| Reported 3 September 2026 | Gagan Ghotra, on X | Numerous examples of AI Mode searches on top-of-the-funnel queries in which none showed any links or citations |
| 3 September 2026 | Glenn Gabe, on X | Side-by-side of the default model against 3.8 Flash, with the difference described as visible |
| 4 September 2026 | Robby Stein, VP of Product, Google Search, on X | "This isn't working as intended, and we'll roll out a fix soon." |
Screenshots are an observation, not a measurement
This is the part most summaries of the story will skip. Nobody has published a measured citation rate for Gemini 3.8 Flash against the previous model. The evidence in the source is screenshots of individual queries, posted by people who ran those queries. That is an observation, and it is a useful one, but it is not a measurement and it carries none of the properties a measurement would have.
There is no sample size, no query set anyone else can rerun, no split by query type beyond the note that Ghotra's examples were top-of-the-funnel, and no before-and-after figure. A percentage will circulate anyway, because that is what happens to stories like this one, and when it does it will not have come from the reporting. If you repeat a number about this, check where it came from first.
The honest version of the claim is the one the source made: several experienced observers ran queries, saw an absence of links, and one of them compared the two models side by side and said the gap was visible. Google then agreed something was wrong. That is enough to justify checking your own data. It is not enough to quantify anything.
Who this affects, and who it does not
Model selection in AI Mode is a paid-tier feature. Google AI Pro and Ultra subscribers see the menu behind the "+" icon; everyone else gets the default. So the population running Gemini 3.8 Flash is a subset of AI Mode users, and AI Mode is itself a subset of Google search traffic. Most search traffic is not coming from this surface, and saying so is not a way of dismissing the story.
The source says nothing about whether free-tier or default AI Mode answers are affected, and it should not be read as implying they are. It also says nothing about Thailand, and nothing about whether Thai-language AI Mode answers behave the same way as the English ones the observers were running.
Where the story does hit hard is in testing. Marketers who track AI visibility tend to do it from accounts that have the model menu, because those accounts are the ones that let you control the variable. The surface being tested and the surface carrying the traffic are therefore not guaranteed to be the same, and this week they demonstrably behaved differently.
Why a citation is the whole mechanism
For anyone doing generative engine optimisation or answer engine optimisation work, this is the analysis worth carrying away, and it is analysis rather than reporting. A citation is the only mechanism by which an AI answer sends a click or a visible brand mention back to a site. Strip the citation out and the answer still uses what it read, but the site gets nothing it can see in its own analytics.
That means a model swap which suppresses links changes measured referral traffic without anything changing on the site at all. The pages are the same, the crawl is the same, the ranking is the same, and the number in the report moves. Anyone reporting AI-referred sessions as a performance metric is reporting on a system whose linking behaviour can shift between one week and the next for reasons that live entirely inside Google.
The strategic conclusion does not change because of one buggy model. Content still has to be retrievable, clearly attributed and specific enough to be quoted, which is what generative engine optimisation has been about from the start. What changes is how much weight a single month of citation data can carry.
The measurement trap, and what to check in your own reporting
A reading taken on one model and a reading taken on another are not comparable. That sentence is the whole operational lesson. If your August test ran on the previous model and your September test ran on 3.8 Flash, the difference between them includes a model change you did not make and did not log.
So log which model answered. Every AI-visibility test you run should record the model name in the same row as the prompt, the date, the account tier and the result, because the model in that menu changes on Google's release schedule and not on yours. Without that field, a citation log records two variables moving at once and cannot separate them afterwards.
Then check the shape of any drop before you explain it. If AI-referred sessions fell, look at whether the fall lines up with the dates in this story or with something you changed on the site. Those are different explanations with different fixes, and only one of them is yours to solve. Do not attribute a fall to a bug you cannot confirm affected your queries: you would need to know that your traffic came from subscribers who had selected 3.8 Flash, and standard analytics does not tell you that.
The same discipline applies to the recovery. When Google ships the fix, citation counts taken during the affected window will look artificially low next to whatever follows, and a rebound will look like a content win. It will not be one. Mark the window in your reporting notes now, while you still remember why.
What this means for Thai marketers
The source says nothing about Thailand, and it does not say whether Thai-language AI Mode answers behave the same way, so treat any assumption in either direction as untested. The rollout Stein described covers Google AI Pro and Ultra subscribers around the world, which places Thai subscribers on those tiers inside the described rollout, but that is a statement about availability rather than about behaviour in Thai.
The practical work for a Thai team is measurement hygiene rather than a content change. If you run AI visibility checks in Thai and English, run them as separate series, record the model for each, and stop treating a month-over-month move in citations as a signal until you can rule out a model change underneath it. Teams building AI SEO reporting this year are measuring a system that is being rebuilt while they measure it.
Keep the AI-answer reporting separate from the organic reporting as well. Rankings and impressions move on a crawl and indexing cycle that you can reason about, while AI citations move on a model release cycle that you cannot see. Blending them into one chart hides the difference and makes ordinary SEO in Thailand performance look unstable when it is not.
FAQ on the Gemini 3.8 Flash citation drop
Is Google going to fix this?
Yes, Google has said it will. Robby Stein, VP of Product for Google Search, replied on X on 4 September 2026 that this is not working as intended and that a fix will roll out soon, but no date, no scope and no mechanism for that fix has been given.
How much did citations actually drop?
Nobody has published a figure, and that is the honest answer. The evidence in the reporting is screenshots of individual queries posted by Gagan Ghotra and a side-by-side comparison posted by Glenn Gabe on 3 September 2026, which is an observation rather than a measured citation rate.
Does this affect free AI Mode users or the default answers?
The source does not say. The reports concern Gemini 3.8 Flash, the model that Google AI Pro and Ultra subscribers can select from the "+" icon, and free-tier AI Mode users get no model choice at all.
Is this happening in Thailand, or in Thai-language answers?
The source says nothing about Thailand and nothing about Thai-language AI Mode answers. Stein described the 3.8 Flash rollout as available to Google AI Pro and Ultra subscribers around the world, but availability of a model is not evidence about how it behaves in a given language.
Do I have to change anything on my site because of this?
No, there is nothing in this story that points to a site change. The useful response is to log which model answered in every AI visibility test you run, and to check whether any fall in AI-referred sessions lines up with the dates in this story rather than with something you did.
Where to go from here
The thing to watch is whether the fix arrives quietly, which is how these usually land. There is no version number attached to it and no announcement promised, so the only signal most teams will get is their own tracking returning to whatever it was before. That is a good reason to have the tracking in place this week rather than after the fact.
If your AI visibility reporting does not currently record which model produced each answer, that is the change worth making today, and it takes an afternoon. Anyone who wants a second pair of eyes on how their AI-answer measurement is structured is welcome to get in touch.
