On September 2, 2026, Google put Gemini 3.8 Flash into AI Mode for Google AI Pro and Ultra subscribers worldwide. Within roughly a day, practitioners running top-of-funnel queries were getting AI Mode answers that contained no links and no citations at all.
Google confirmed the behavior was a bug and shipped a fix by the morning of September 4 — about two days of exposure on a subscriber-tier slice of AI Mode. Small, and worth more attention than a bigger incident, because of what it reveals about the metric agencies started billing for this year.
What happened to AI Mode citations between September 2 and September 4, 2026?
A model upgrade in AI Mode temporarily stopped the system attaching source links to its answers. Sites were still being read and used to build responses; they were not being credited or linked. Google’s Robby Stein, VP of Product for Google Search, called it unintended and promised a fix.
The sequence, as reported:
- September 2 — Stein announced on X that Gemini 3.8 Flash “has landed in Search,” citing improvements over 3.7 Flash in software engineering, agentic tasks, and multi-step reasoning. Availability: AI Mode, AI Pro and Ultra subscribers, globally.
- September 3, 7:41 a.m. — Search Engine Roundtable published Barry Schwartz’s write-up of testing by Gagan Ghotra, who ran a series of top-of-funnel AI Mode queries and got zero links back. Glenn Gabe posted side-by-side captures of the default model versus 3.8 Flash on the same query: citations in one, nothing in the other.
- September 4 — Stein replied publicly: “This isn’t working as intended, and we’ll roll out a fix soon.”
- September 4, later — Gabe confirmed AI Mode on 3.8 Flash was providing links again.
Two things stand out. Google’s announcement described a capability upgrade and said nothing about attribution. And the regression surfaced through two independent practitioners posting screenshots on X, roughly twenty hours after launch — not through Google’s own release testing.
Why did a model swap remove citations when the ranking system never changed?
Because retrieval and attribution are separate stages, and only one lives in the ranking system. AI Mode first retrieves and grounds, using search infrastructure to decide which pages inform the answer. Then the language model writes the prose and emits the annotations the interface turns into citation cards. Swap the model and you have swapped the component that decides whether those annotations appear at all.
That distinction predicts things. Any model rollout, any gap between the free tier and the Pro or Ultra tier, and any region-staged deployment can move your measured citation rate substantially without a byte changing on your site — and the movement will not be uniform across query types. Citation propensity runs highest on specific, entity-heavy, or comparison queries where the model needs to point at something, and lowest on broad conversational questions it can answer from memory. An attribution regression therefore hits discovery-stage queries hardest, which is exactly the pattern Ghotra reported.
Apply that to a case nobody has covered: if Google ships a cheaper model to the free tier next quarter and keeps a costlier one for subscribers, two audiences searching the same phrase will produce different citation rates for your page, and neither number says anything about your content.
Does this make AI citation share a useless metric?
No — but it has a noise floor set by infrastructure you cannot see, and most people report it at a cadence far finer than that floor. Weekly AI citation share is close to meaningless. Monthly, with a log of known platform events beside it, is defensible.
Here is the part worth pushing back on. Google framed 3.8 Flash in terms of reasoning quality, and by that measure it may well be an upgrade. But reasoning quality and attribution behavior are separately engineered and, on this evidence, separately tested — or not tested against publisher outcomes at all. Vendors benchmark models on tasks that do not include “did it credit its sources.” Read every model announcement as a claim about capability and no claim whatsoever about how it will treat your traffic. Those two are now publicly decoupled, with a dated example.
The cost of getting this wrong is not abstract. A marketing lead watching a citation dashboard would have seen a two-day trough, concluded the content was demoted, and started rewriting pages, reworking schema, or cancelling a program that was working fine — money spent fixing a bug Google had already fixed. Same discipline we’ve written about when a reporting glitch looks like a ranking problem: confirm the platform is stable before you touch the site.
Which Michigan businesses were most exposed to a top-of-funnel citation gap?
The ones whose whole discovery layer is top-of-funnel research: long-cycle B2B. Tier 1 and Tier 2 suppliers, industrial distributors, and specialty manufacturers whose buyers spend weeks in an unbranded research phase before a vendor name enters the conversation.
A machining shop in Sterling Heights or a fluid-power distributor in Livonia does not get found on “buy hydraulic manifold.” It gets found on “what causes cavitation in a hydraulic manifold,” and a procurement engineer clicks the source. That citation is the entire introduction. Strip it for forty-eight hours and the mechanism by which a stranger becomes a lead switches off silently, with no impression drop in any dashboard the business owns.
Two other Michigan segments share that shape. Medical device and life-science firms around Ann Arbor publish technical explainers living almost entirely in the research phase, where YMYL scrutiny makes attribution part of the credibility case. Office-furniture and food-processing suppliers in the Grand Rapids corridor sell into specification-driven purchases where shortlists form months ahead of an RFQ. For all three, the correct reading is not “our GEO strategy failed” — it is that a two-day gap in a months-long buying cycle is statistically invisible, and a two-month gap would not be.
How should a small team audit its AI visibility data after an incident like this?
Seven steps a three-person marketing team can finish this week without buying anything.
- Annotate September 2–4, 2026 in your analytics. Add a note in GA4 and in whatever tracker you use. An unlabeled trough becomes a false trend the moment someone builds a quarterly deck from it.
- Segment referral traffic by source, not by bucket. Split google.com AI surfaces from ChatGPT, Perplexity, and Gemini. If all four dropped together, the cause is yours; if one did, it is theirs.
- Pull the same window in Search Console. If organic impressions and clicks held steady while AI referrals dipped, the retrieval layer never stopped seeing you.
- Build a fixed set of fifteen to twenty-five prompts in your real discovery language, run it manually on the first business day of each month, and log which sources get cited. Free, ten minutes, and the only baseline you fully control.
- Record which tier you tested from — signed into a paid Google AI subscription or not. After this incident, tier is a variable, not a footnote.
- Move reporting to monthly with a platform-events log beside it. Keep weekly collection; stop weekly interpretation.
- Set a 30-day rule before structural changes. No rewrites, schema overhauls, or budget shifts triggered by movement younger than a month.
None of that replaces the work of being worth citing — clear answers near the top of the page, real expertise, sources of your own. It stops you dismantling that work over a two-day platform hiccup. For page-level mechanics, see our notes on building content that AI answer engines actually cite.
What dates should Michigan marketers have on the calendar right now?
Four, and only the last one is speculative.
- September 30, 2026 — Google’s alcohol advertising policy update takes effect, revising country-specific guidelines and expanding allowed advertising locations. Relevant to Michigan’s brewery, distillery, and hospitality advertisers.
- October 2026 — Changes to the Display & Video 360 API and Structured Data Files, per the Google Ads Developer Blog. If your reporting stack reads SDFs, budget engineering time now.
- February 2027 — Automatic upgrades of Dynamic Search Ads to AI Max begin; voluntary in-account nudges started in September 2026. Anyone running DSA has roughly five months to decide on their own terms.
- Ongoing — Expect more AI Mode model rotations without publisher-facing release notes. Google has not committed to versioning this publicly, so assume the next swap arrives unannounced.
Frequently asked questions
Did the citation bug affect everyone using Google?
No. Gemini 3.8 Flash rolled out September 2, 2026 to AI Mode for Google AI Pro and Ultra subscribers. Classic blue-link results and standard AI Overviews were not part of the reported issue. If your audience skews toward non-subscribers, exposure was smaller — though you cannot measure that split precisely.
Should we file anything with Google when we see this?
There is no formal channel for AI Mode attribution bugs. Here the fastest working path was public reporting with reproducible screenshots, which drew a response from Google’s product lead within about a day. Documenting examples with timestamps is worth doing regardless — it makes your internal post-mortems credible.
Is it worth paying for a GEO tracking tool after this?
It depends on whether the tool logs platform events alongside your numbers. A tracker that reports citation share without annotating model changes will manufacture trends that do not exist. A manual monthly prompt set costs nothing and gives a cleaner baseline than an unannotated dashboard.
Does this change what we should be publishing?
Not the content. Direct answers placed immediately under question-shaped headings, verifiable specifics, and clear authorship still make a page citable. What changes is reaction time: separate “is our content working” from “is the platform behaving,” and answer the second one first.
The takeaway for Michigan teams
A forty-eight-hour attribution bug is not a strategy problem. It is a clean demonstration that AI citation volume is partly a property of Google’s deployment schedule rather than your website — and that weekly AI visibility reporting has not priced that in. For a Southfield marketer or a Livonia supplier watching a dashboard, the discipline is the one that separates measurement from expensive guessing: build a baseline you control, annotate the platform’s behavior, and refuse to act on a two-day line.
Comments are closed