TL;DR: On 13 August Google released a new AI model called Gemini 3.7 Flash. A day later it appeared inside AI Mode, the tab in Google that writes you an answer instead of showing a list of links. It is not on for everyone: only people paying for Google’s AI Pro or Ultra plans can pick it, and only in English. The model itself is a routine upgrade. What matters is the menu it arrived in. AI Mode used to give one answer to one question, and now the answer depends on which engine the person selected.
First, what a “model” is here
An AI model is the engine that writes the answer. Google has a family of them, all called Gemini, and they get replaced every few weeks.
AI Mode is where that engine sits inside Google Search. You type a question, you get a paragraph or two of prose, and a small number of links to the pages the answer drew on. Those links are called citations, and winning them is what people mean by AEO or GEO — optimising so an answer engine names your site as a source, rather than optimising for position three in a list of blue links.
Until now, most people using AI Mode got whatever engine Google had set as standard, or an “Auto” setting that picks one for you based on how hard your question looks. What changed on 14 August is that a paying subscriber can override both and choose from a list.
Two engines reading the same web pages do not necessarily pull out the same sentences or credit the same sites. Google publishes no mapping of which engine cites what.
What actually happened
Gemini 3.7 Flash shipped on 13 August, about three weeks after the model it replaces, according to 9to5Google, at roughly half the launch price of that predecessor. Google’s own model card is modest about it, describing the model as based on its predecessor with “algorithmic improvements to its core reasoning foundation” rather than a rebuild from scratch.
The Search part came the next day. Robby Stein, a VP of product at Google, posted that it was rolling out “today globally in AI Mode for Google AI Pro & Ultra subs in English”. That was picked up by Search Engine Land and, separately, by Search Engine Journal. You reach it by clicking the “+” icon in the ask-anything bar and choosing from the model list.
Search Engine Journal flags two things Google has not said: whether 3.7 Flash becomes the default for everybody, and whether the Auto router can now send your question to it without telling you. Google’s support pages hadn’t been updated to mention the option at all.
Why an agency should care about a menu
If you check what Google’s AI says about a client, you now have to answer a follow-up question: which Google AI?
That’s a nuisance rather than a catastrophe, but it’s the kind of nuisance that quietly invalidates a reporting habit. A screenshot of an AI Mode answer, taken once, from one account, is evidence about one engine on one day. The prospect reading about your client may be a Pro subscriber who has pinned a different model. When two people’s answers disagree, nobody in the room can say who is right, because everybody is right about a different thing.
Treat AI Mode as a distribution rather than a fact. Ask the same question repeatedly and look at how often your client shows up, not whether they showed up that once. That was already the right way to track AI Overviews and AI Mode, which vary between users anyway. A visible model picker makes the variance impossible to ignore.
The cutoff line, and what it does to your pages
Buried in the model card is a number that matters more than the benchmarks. The knowledge cutoff — the date after which the model simply has no memory of the world — is March 2026, and the card warns that for some subjects its knowledge stops as far back as January 2025.
So take anything about your client that changed after March 2026. A price. A new branch. A product they stopped selling. None of it is inside the model. If it appears in an AI Mode answer, it got there because the system fetched a page while the person was waiting, and read it.
That puts the weight back on the page. The fact has to be there as text, not baked into an image or a PDF. The page has to be reachable by the crawler, the automated reader that fetches pages on the engine’s behalf: no stray noindex tag telling engines to skip it, no blocking rule in robots.txt, the small file at the root of a site that tells automated readers where they may go. And it has to render without waiting for JavaScript, because most of these fetchers don’t wait.
What to check this week
- Run your standard client queries in AI Mode more than once, on different days, and record whether the client appeared rather than what the paragraph said.
- If you have a Pro or Ultra account, run the same query on 3.7 Flash and on Auto, and note where the cited sources differ. That’s free evidence about how much model choice moves your results.
- For every fact that changed since March 2026, confirm it exists as plain text on a crawlable page. Prices in images are the usual offender.
- Stop treating a single flattering AI answer as a deliverable. Its shelf life is about one model release.
Where this gets expensive
The work above is fine for one client and one query. Across a portfolio, run weekly, against a surface that keeps swapping engines, it stops fitting in a retainer. Preferium measures four AI engines on every plan — ChatGPT, Claude, Perplexity and Gemini — with Google AI Overviews and AI Mode tracked as separate surfaces, so a shift shows up as a change in the trend rather than a surprise in a client call. On the page side, 47 automated checks crawl every URL and score the site 0–1000, then the system fixes what it finds, deploys, and re-checks the live page in a real browser. More on how the system works, and the underlying method is in measuring AI citations.
Key takeaways
- Gemini 3.7 Flash reached AI Mode on 14 August, one day after release, for AI Pro and Ultra subscribers only, English only, worldwide.
- Google has not said whether it becomes the default or whether the Auto router uses it, and the support docs are behind.
- AI Mode is now a menu of engines, so a one-off screenshot of an answer is weak evidence. Track appearance rates over time instead.
- The model’s own knowledge stops in March 2026, so anything newer about a client has to be read off a live page: text, crawlable, no JavaScript required.
- Run one query on two models and see how far the citations move. That gap tells you what your current reporting is worth.
