Google put a new model inside AI Mode on Tuesday and by Wednesday morning practitioners were posting answers with no links and no citations at all.

Google added Gemini 3.8 Flash to AI Mode on 2 September. Robby Stein, VP of Product for Google Search, announced it on X with "Gemini 3.8 Flash has landed in Search", and Nick Fox, who runs Google Search, noted "Less than three weeks after 3.7, we're introducing Gemini 3.8 Flash." It is opt-in, not the default: AI Pro and Ultra subscribers pick it from the (+) icon in the model dropdown, and free users get the default model with no choice. Search Engine Land expects Google will "likely roll this mode out to all searchers, even free plans, in the coming weeks or months", which is the reporter's expectation and not a Google statement.

The next morning Gagan Ghotra posted multiple examples on X showing the new model is, in Search Engine Roundtable's words, "less likely to cite sources and link to sources than previous models", including top-of-funnel queries answered with no links or citations whatsoever. Glenn Gabe confirmed seeing the same gap between the default model and 3.8 Flash. Two named practitioners posting screenshots is early warning, not measurement, and Google has said nothing.

Google extended Merchant Center AI performance reports to Australia, Canada, India and New Zealand, on top of the United States where they launched in July. The reports cover AI Mode and AI Overviews for "conversational queries with shopping intent", English only. Google's wording: "With these insights, you can better understand how your brands show up in conversational queries, evaluate your visibility across different shopping stages, and view trends to help optimize your product data."

Older material, four months old, never carried here, and the most useful thing this radar has read this week. Method as stated on the page: 82,619 prompts producing 1,548,213 snapshots across Google AI Overviews, Google AI Mode and ChatGPT Search, in Germany, the USA, the UK, Italy, Spain and France, measured at weekly intervals over 17 weeks from 17 December 2025 to 8 April 2026.

The headline figures are the share of cited sources replaced from one week to the next: AI Overviews 5%, AI Mode 56%, ChatGPT Search 74%. Content type survival inside AI Mode: video and YouTube 24%, big tech 16%, news and media 1.4%. The author's own caveat, verbatim: "The ChatGPT data only includes prompts with consistent source attribution. The drift rates cited are conservative; at the URL level, they are 15% higher."

Read it carefully before it reaches copy. This measures which URLs are cited, not which brands are named, and those are different quantities. A business can stay named while the URL supporting it gets swapped. It is also a different question from the one we asked on 1 September, which was run-to-run variance within a single day with nothing changed. Both matter and neither substitutes for the other.

The 68% headline was carried on 13 August via a secondhand reference, and the ledger dated it to July 2026. That date was wrong. The study is from May 2025, which makes it sixteen months old, and the prevalence percentage should be treated as historical. Correcting it here so the ledger does not carry a bad date.

What was never carried is the part that has aged better, because it is structural rather than a percentage. Method: 540 queries, manually run, three US cities (Houston, Phoenix, Denver), six industries (plumbers, personal injury lawyers, dentists, optometrists, medical clinics, real estate agents), searched from outside the target geographies. The split by intent: a local-intent query like "personal injury lawyers in Phoenix" returned a local pack 93% of the time and an AI Overview only 15%, while an informational query like "how much do personal injury lawyers charge in Phoenix" returned an AI Overview 92% of the time and a local pack 6%. Hybrid queries such as "should I get a lawyer after a car accident in Phoenix" returned an AI Overview 97% of the time.

The relevance is direct: how a question is phrased decides whether an AI Overview exists to be measured at all. If a chunk of our 25 questions are phrased as local-intent lookups, we may be returning zeroes on our differentiator surface that mean "no AI Overview was shown", not "the client was not named".