On 3 September 2026, users who picked Gemini 3.8 Flash in Google's AI Mode reported answers with barely any source links. Google's Robby Stein said it "isn't working as intended" and promised a fix, and by the next morning Glenn Gabe reported links were back. The episode lasted roughly a day, affected a paid-subscriber model option, and carries one durable lesson: your citations depend on a model layer you do not control.
(A note on dates: some internal planning notes placed this in early October. The primary reporting I opened dates the bug to 3-4 September 2026. I use the verified dates. Facts are as of 4 October 2026.)
Key Takeaways
- Gemini 3.8 Flash arrived in AI Mode on 2 September 2026 for Google AI Pro and Ultra subscribers, selectable from the plus icon.
- On 3 September, Gagan Ghotra and Glenn Gabe posted examples of answers with few or no links; Search Engine Roundtable reported it.
- Stein acknowledged the problem on 3-4 September; Roundtable reported links returned by the following morning.
- No measured citation rates were published. Evidence was screenshots, not a dataset.
- Treat model changes like algorithm changes: log the model, re-test, and do not rewrite content on one day of data.
The timeline
| Date (2026) | Event | Source |
|---|---|---|
| 2 Sep | Gemini 3.8 Flash added to AI Mode for AI Pro and Ultra | Search Engine Roundtable |
| 3 Sep | Reports of few or no links; Stein: "This isn't working as intended, and we'll roll out a fix soon" | Search Engine Roundtable |
| 4 Sep | Gabe: Gemini 3.8 Flash "is now providing links" | Same report, updated |
| 15 Sep | Search Live moved to Gemini 3.8 Live, which Google says includes web links | Search Engine Roundtable |
Roundtable's monthly recap, the October 2026 Google Webmaster Report, lists the Flash rollout, the link issue and the Live follow-up together.
What we know and what we do not
Known
Google's VP of Product for Search publicly acknowledged the behaviour and called it unintended. The model was available only to paying subscribers who chose it manually; free users received the default model. These details are in the Roundtable coverage and in an independent write-up from Relevant Audience.
Not known
- How many queries were affected. No citation-rate data was published. Relevant Audience makes the same point: there were screenshots of top-of-funnel queries with no citations, not sample sizes.
- Why it happened. Google gave no technical explanation or timeline detail.
- Whether the fix changed anything beyond restoring links. Nobody has published a before/after study.
Be wary of any article that quantifies the drop. If it does not show a method and sample, it is guessing.
Why this matters even though it was brief
1. Citations are model behaviour, not a fixed feature
AI Mode answers come from a model, and models differ in how they attribute. When Google changes the model, the link behaviour can change overnight. I tested the opposite problem in my earlier post on Gemini 3.8 Flash and visibility: the model name can shift under your measurements.
2. Your visibility reports have a hidden variable
If you track AI Mode citations weekly, a dip on 3 September could have been a bug and not a content problem. Teams that reacted by rewriting pages would have fixed nothing. Add a "model/version" column to every AI visibility log.
3. Referral value depends on links existing at all
Publishers have been watching AI Mode traffic closely; see my analysis of AI Mode and publisher traffic. When links vanish, even for a day, so does the click path. Google's documentation on AI features in Search says AI features show "a wider and more diverse set of helpful links" than traditional search, and the Search Live announcement stressed "web links to dive deeper" for the new Live model.
The same documentation is blunt about optimisation: there are "no additional requirements to appear in AI Overviews or AI Mode," and you do not need special AI text files or markup. To be eligible you need pages that are indexed and snippet-eligible. Traffic from these features is reported in the standard Search Console Performance report under the Web search type, so you cannot isolate AI Mode clicks there, which is another reason a one-day dip is hard to diagnose. Google's Search Central documentation updates page is the place to watch for any official change to those rules, and my notes on opting out of AI features cover the controls (nosnippet, max-snippet, noindex) if you want to limit exposure.
A bug-vs-decline decision test
Use this before changing any content after a citation dip.
- Is it one surface or many? If only AI Mode dropped and AI Overviews and classic results are steady, suspect a surface-level change.
- Is it one model? Compare the same prompts on the default model and the optional one.
- Is it across competitors too? If rivals lost citations on the same queries, it is probably not you.
- Does it persist 48 hours? Short dips that recover are noise or fixes.
- Did Google or a reliable reporter acknowledge anything? Check Search Engine Roundtable and Google's product leads' posts.
If you answer yes to the first, third and fifth, hold your content steady.
What I changed in my own testing
I now record, for every AI Mode test: date, model selector state, location, logged-in or out, query wording, and whether links were present. It takes a minute per query and prevented me from misreading exactly this kind of event. For the broader testing approach, my post on why a competitor ranks in AI answers has the checklist.
What to do if you rely on AI Mode visibility
- Use Search Console's generative AI reporting where available, and see my explainer on the report for what it does and does not show.
- Diversify. Do not let one surface carry your pipeline. Newsletter, community and direct traffic cushion model swings.
- Keep your pages answer-ready. Clear definitions, specific facts and visible authorship help in any model, and they match Google's people-first content guidance. Also keep an eye on the wider search weather: the September spam update began three weeks after this bug, so do not confuse the two events in your data. That is durable work; bug-chasing is not.
A sample log you can copy
| Date | Query | Model selected | Logged in | Location | Links shown | Your page cited |
|---|---|---|---|---|---|---|
| 3 Sep | "what is X" | 3.8 Flash | Yes | City, country | 0 | No |
| 4 Sep | "what is X" | 3.8 Flash | Yes | City, country | 5 | Yes |
| 4 Sep | "what is X" | Default | Yes | City, country | 6 | Yes |
The third row is the useful one. Running the same query on the default model tells you whether a change is about the model or about you. Without that control, you cannot separate a bug from a ranking issue. Even five or ten queries logged this way for a month gives you a baseline that survives the next model launch, and Google is clearly shipping them quickly: Roundtable noted Gemini 3.8 Flash arrived less than three weeks after Gemini 3.7 Flash reached AI Mode.
The honest takeaway
Nothing in the record shows this was a policy decision about publishers. It was a bug, acknowledged and fixed within about a day. The takeaway is operational: build your measurement so one bad model day cannot trick you into bad decisions.
FAQ
What was the Gemini 3.8 Flash citation bug?
Users of AI Mode who selected Gemini 3.8 Flash saw answers with few or no source links on 3 September 2026. Google's Robby Stein said it was not working as intended and promised a fix.
Was it fixed?
Yes, according to Search Engine Roundtable, Glenn Gabe reported by the following morning that Gemini 3.8 Flash was providing links again.
Who could use Gemini 3.8 Flash in AI Mode?
Google AI Pro and Ultra subscribers, via the plus icon model option, per Roundtable's report of the 2 September launch. Free users did not get it at launch.
How many citations were lost?
Unknown. Only screenshots of individual queries were shared; no citation-rate study has been published.
Does this mean Google is removing links from AI Mode?
There is no evidence of that. Google called it unintended and fixed it. Search Live's Gemini 3.8 Live upgrade was announced with a promise of web links.
Should I change my content because of the bug?
No. A one-day, acknowledged bug is not a content signal. Log it as an event and compare before and after.
How do I test AI Mode citations properly?
Record date, model selection, location, query wording and link presence for a fixed query set, repeated weekly. Compare across models if you can.
Does this affect AI Overviews?
Reports concerned the AI Mode model option. Check AI Overviews separately; do not assume the same behaviour.
Work with me
I'm Younus Fardeen, an AEO and organic growth marketer with 4+ years of marketing experience in edtech and startups, including Masai School's social growth (Instagram 26K to 117K, LinkedIn 50K to 200K). If you want AI search measurement that separates real signals from bugs, see my work and use the contact form at younusfardeen.in.