AI Mode stopped citing sources on Gemini 3.8 Flash, and only paid subscribers could see it

Google's AI Mode briefly stopped citing sources at all, and the cause was a model release rather than anything a publisher or a marketing team did. On September 2 Google added Gemini 3.8 Flash to the AI Mode model menu on the same day it shipped the model, offering it to Google AI Pro and Ultra subscribers worldwide from the plus icon in the search bar. Robby Stein, Google's VP of product for Search, announced on X that it was "available today in AI Mode for Google AI Pro & Ultra subs around the world". Within a day, people running AI Mode on that model were getting answers with almost no links attached.

Search Engine Land reported the failure on September 3, working from screenshots Gagan Ghotra posted publicly on X, with Glenn Gabe posting his own. The missing citations showed up most on top-of-the-funnel queries. Stein replied to the report on X, saying "this isn't working as intended, and we'll roll out a fix soon". Gabe posted the next morning that links had returned. The bug leg of this story rests on that one outlet, which is owned by Semrush, a company that sells AI visibility tooling; the Google response it quotes was public.

The failure was in the plumbing, not the ranking

For a software company, a citation in AI Mode is not a nice-to-have mention. It is the only mechanism by which a category query about, say, incident management or invoice reconciliation turns into a person landing on your site. When the citation layer goes dark, every brand in the category goes dark simultaneously, and no amount of schema, entity work or content refresh does anything about it, because the answer generation never got to the part where it attributes.

Most AI search advice aimed at software marketers treats citations as an earned outcome: publish the comparison page, get the third-party mention, structure the data, get cited. This episode says the supply of citations is also a build artifact. It can go to roughly zero overnight and come back the next morning, on a release calendar you have no visibility into.

The release cadence is faster than your reporting cycle

Gemini 3.8 Flash is Google's third Flash release in six weeks, arriving about three weeks after 3.7 Flash joined the AI Mode menu on August 14, according to Search Engine Journal's account of Google's launch post. The model underneath AI Mode has moved at a similar pace for most of a year: Gemini 3 Pro appeared as an option in November, Gemini 3 Flash became the default in December, and Gemini 3.5 Flash took over as the global default at I/O in May.

Set that cadence against how AI visibility actually gets reported inside a software company. The share-of-citations number goes in a monthly marketing review, or a quarterly board deck. A month-over-month movement in that number now spans at least one model change, and possibly two. The delta you are presenting as the result of your content programme is partly the result of a different model answering the same prompts. If you cannot say which model produced each reading, the comparison is not measuring what the slide says it is measuring.

The tier that broke is the tier your buyer is on

The outage only reached people who had actively selected 3.8 Flash, which means Pro and Ultra subscribers. Free-tier AI Mode users get whatever the default is and no model picker at all, so nothing about their experience changed. That inverts the usual assumption behind visibility monitoring. Most tools that track AI answers sample the free, logged-out, default experience, because that is cheap and repeatable to automate. For a day, that sampling would have reported business as usual while the version of AI Mode most likely to sit in front of a paying, enterprise-adjacent buyer returned answers with no way out to a vendor site.

There is also no clean audit trail. Search Engine Journal notes that Google's support page for the AI Mode model menu still lists only "Fast" and "Pro", unchanged since 3.7 Flash arrived. The dropdown in front of the searcher is the record of which model answered.

Your AI visibility number is now a reading taken from a specific model on a specific day, not a rank you hold.

What a software marketer should change

• Record the model name and tier alongside every AI visibility measurement, the same way you would record the device and country for a rank check.
• Add a paid-tier spot check to whatever tool you use. If your monitoring only sees the free default, it is blind to the surface your enterprise buyers are on.
• Before you explain a drop in AI referrals to your leadership, check whether a model shipped in the window. Trade press is currently faster at flagging these than any dashboard.
• Stop reporting single-month movements in citation share as a programme outcome. Use a longer window, or report the underlying queries you are cited on instead of the aggregate.

None of this argues against working on AI visibility. It argues against treating it as a position you occupy. It behaves more like uptime on a dependency you do not control, and deserves the same honesty about error bars.