AI Mode Stopped Showing Citations. The Reassuring Quote Was Five Days Old

What actually happened, in UTC

Every timestamp below was verified this run against the source itself. Publication times come from each page’s own structured data. Post times come from decoding the numeric IDs, a method described in the next section.

When (UTC)	What	Source
2026-09-02 15:00	Google announces Gemini 3.8 Flash, available to AI Pro and Ultra subscribers “across the Gemini app, AI Mode in Google Search and Gemini in Google Sheets”	blog.google
2026-09-02 23:51:08	Gagan Ghotra posts that AI Mode on 3.8 Flash is returning no citations for top-of-funnel queries	x.com/gaganghotra_
2026-09-03 03:52:47	Ghotra publishes his own first-hand write-up	Search Engine Watch
2026-09-03 11:41	Search Engine Roundtable publishes the story	seroundtable.com
2026-09-03 12:06:41	Glenn Gabe confirms independently: “Whoa, Gagan is right. See below between default model and 3.8 Flash.”	x.com/glenngabe
2026-09-04 00:05:51	Robby Stein, Google: “This isn’t working as intended, and we’ll roll out a fix soon.”	x.com/rmstein

Twenty-four hours and fifteen minutes separate the first public report from Google’s acknowledgement. That is a fast turnaround by any standard. It is also the entire window in which the five-day-old quote was doing work it was never able to do.

Where the old quote came from

Nobody made anything up. Ghotra’s write-up is careful, and the sentence in question is doing something reasonable. He was explaining why the citation drop surprised him, and reached for evidence that Google says citations matter. His exact words:

“Less citations is bit of surprising thing because even recent Googler Roby Stein on X said “yes, definitely an area of continued focus!” in a reply to Barry Schwartz.”

The load-bearing word is “recent.” The post embedding that quote carries the ID 2093126718068523508, which decodes to 28 August 2026 at 00:01:12 UTC. Google’s own announcement of Gemini 3.8 Flash went up on 2 September at 15:00 UTC. The quote is five days and fifteen hours older than the model, and just under six days older than the finding it sits next to.

Placed beside a live incident and labelled recent, a five-day-old quote reads as a response. That is not a claim about anyone’s intent. It is a claim about what a reader does with two sentences in one paragraph.

Ghotra also wrote, in the same piece, “I guess it might be a bug for now.” That hedge turned out to be the most accurate statement anyone made about this incident, including this blog. Google confirmed it as unintended behaviour a day later.

How to date a post on X without opening X

Every post ID on X encodes its own creation time. The format is public: the top 41 bits of the ID are milliseconds since a fixed epoch of 1288834974657 in Unix time, documented with a worked example. Shift the ID right by 22 bits, add the epoch, and you have the timestamp. No login, no API key, no scraping, and no reliance on whatever date a screenshot happens to show.

// node: date any x.com post from its numeric ID
const ms = (BigInt('2093126718068523508') >> 22n) + 1288834974657n;
new Date(Number(ms)).toISOString(); // 2026-08-28T00:01:12.658Z

Run against the four IDs in this story, it produces the table below. We validated the method first against the worked example in the linked documentation, which reproduces to the millisecond.

Post ID	Decoded timestamp (UTC)	Who, and what it is
2093126718068523508	2026-08-28 00:01:12	Stein, the “continued focus” quote read as a response
2095298509469810837	2026-09-02 23:51:08	Ghotra, the original finding
2095483619284251044	2026-09-03 12:06:41	Gabe, independent confirmation
2095664601023361501	2026-09-04 00:05:51	Stein, the actual response

Two of those four are by the same person, five days apart, and only one of them is about this. Nothing on the page tells you that. The ID does.

Does this change what the citation drop means?

Yes, and this is the part that matters for anyone running visibility measurement.

If a citation drop is a model behaviour change, it is a signal, and the correct response is to re-baseline and adapt. If it is unintended behaviour that Google intends to fix, it is noise, and re-baselining against it will corrupt your numbers for as long as the fix takes to roll out. Those two readings call for opposite actions, and the only evidence that distinguishes them is what Google actually said and when.

The distinction was not available on 3 September. It was available at 00:05 UTC on 4 September. Anyone who ran a comparison during that window and wrote up a trend was measuring a bug and calling it a model.

There is a narrower lesson too. A citation rate measured against AI Mode is measured against whichever model is currently behind the selector, and that changed mid-window here. Any AI-visibility number that does not record which model produced it is not reproducible.

What we are not claiming

We do not have a rate. No source we opened gives a percentage, a sample size, or a query count for how often 3.8 Flash dropped citations. The evidence is two named practitioners posting screenshots of their own tests, which is real evidence and is not a measurement.

We are not claiming the fix has shipped. Stein said Google would “roll out a fix soon.” We checked the sources listed above and found no confirmation that it has landed, no timeline, and no statement of what caused it.

We are not claiming anyone acted in bad faith. The quote is real, quoted accurately, and never presented by its author as a statement about Gemini 3.8 Flash. Stale context attaches itself to fast-moving stories without anybody deciding that it should.

We are not claiming Google was slow. Twenty-four hours from first public report to an on-record acknowledgement is fast.

What would change this verdict

A Google statement naming what happened and when the fix shipped would turn the whole incident from an anecdote into a datable event, and would let anyone who measured through the window mark and discard the affected period. A published sample from either tester, with query counts and dates, would turn a screenshot into a rate. Until one of those exists, the honest summary of this story is that a widely-shared reassurance was five days old, a real one arrived a day later, and the tester’s own hedge beat both.