Google Turned Its Search Bar Into an AI Agent
Gemini 3.5 Flash now runs Google AI Mode. What it does for writers, where its accuracy breaks, and how to get your work cited inside AI answers.
TL;DR: Google made Gemini 3.5 Flash the default model behind Google AI Mode in May 2026 and pushed it to nearly 200 countries. For writers, it turns Search into a fast research engine that fans a single question into a dozen-plus parallel searches and answers with citations. The speed is genuinely useful for research and fact-checking. The catch: it scored a 61% hallucination rate on one accuracy benchmark, so treat every answer as a lead to verify, not a source to trust. Below: how it works, where it breaks, the Deep Search and Personal Intelligence power moves, and how to get your own writing cited inside AI answers.
If you write for a living and you still picture Google as a list of blue links, your mental map is 6 months stale.
In May, Google made Gemini 3.5 Flash the default brain behind AI Mode. Then rolled it out to nearly 200 countries.
Most writers I know haven’t touched it. They’re still pasting research questions into a chat window in another tab.
This piece tells you three things. What AI Mode actually is now. Whether it’s good enough to pull into your writing day. And where it quietly falls on its face.
Short verdict up front: worth opening for research. NOT worth trusting blind.
The speed is real. The accuracy is shakier than Google wants you to believe.
What Google AI Mode actually is now
AI Mode is the tab inside Google Search that stopped handing you links and started handing you answers.
You type a question. Gemini reads across a pile of sources. It writes you back a synthesized answer with citations you can click. Ask a follow-up and it remembers what you just asked.
Google flipped the default to Gemini 3.5 Flash in May at I/O 2026. The number they keep repeating: AI Mode crossed 1 billion monthly users, with queries more than doubling every quarter (blog.google).
The part that matters for you is under the hood.
When you ask a question, the model doesn’t run one search. It uses query fan-out, splitting your question into a dozen or more smaller searches that all run at once, then stitching the answer together (Search Atlas).
So you ask “how have newsletter open rates shifted since Apple Mail Privacy Protection.” It isn’t reading one blog post and paraphrasing. It’s fanning out across a dozen sources and merging them.
For the messy, multi-part questions writers actually have, that’s useful.
Why Gemini 3.5 Flash’s speed changes how you research
Gemini 3.5 Flash is built to be fast. Google says it pushes out tokens about 4 times faster than other frontier models (blog.google).
You feel it. Ask a real research question and the answer lands in seconds. No spinner. No alt-tab to go check email while it thinks.
Speed changes behavior.
When research is slow, you batch it and cut corners. When it’s instant, you actually check the thing you were about to assert. You ask the follow-up. You chase the second source.
The friction that used to talk you out of the work just isn’t there.
That’s the real pitch. Not that it’s smart. That it’s fast enough to stay in your flow.
The Gemini 3.5 Flash cost reality nobody mentions
For you, typing questions into Search, AI Mode is free. No subscription. 98 languages, almost 200 countries (blog.google).
But there’s a number worth knowing.
Google tripled the token price. Gemini 3.5 Flash now runs $1.50 per million input tokens and $9.00 per million output, triple what the old Gemini 3 Flash charged at $0.50 and $3.00 (The Decoder).
Why care if you’re not a developer?
Because “free” search products tend to get monetized. Universal Cart already wires shopping straight into AI Mode across Search, Gemini, YouTube, and Gmail, sitting on a catalog of more than 60 billion product listings (blog.google).
The tool is free because you’re being walked toward a checkout. Use it knowing that.
Where Gemini 3.5 Flash falls on its face on accuracy
This is the part the launch posts skip.
Gemini 3.5 Flash scored a 61% hallucination rate on one Artificial Analysis benchmark. That’s an improvement, down 31 points from the older Flash model. It still trails the leaders, MiMo-V2.5-Pro and Grok 4.3, both sitting around 25% (Artificial Analysis).
A 61% rate on that test means it still invents something on most of those questions. Sit with that before you trust an unsourced claim it hands you.
It also isn’t a clean upgrade. The Flash model still loses to the older Gemini 3.1 Pro on harder reasoning tests like Humanity’s Last Exam and the long-context MRCR v2 benchmark (Appwrite).
Google made the faster model the default. Not the smartest one.
For quick lookups, fine. For a 12-source piece where one wrong fact tanks your credibility, slow down and verify.
My rule: treat every answer as a lead, not a source. Click the citation. No citation, assume it might be inventing the detail.
That one habit separates writers who use AI Mode well from the ones who publish a confident lie because a search box told them so.
Want the part that actually changes your week?
Everything above is the honest review. What’s good, where it breaks, what to watch for.
What’s below is how I actually USE it.
The power-user moves most people never find. Plus the AI Mode features that change how your own writing gets discovered now that Google answers questions instead of listing links.
Paid subscribers get the rest:
The Deep Search trick that turns AI Mode into a research analyst running hundreds of searches and handing you a cited report
How to wire Gmail into AI Mode so it answers using your own receipts, bookings, and threads
The content structure changes that get your writing cited inside AI answers, backed by AirOps’ 2026 analysis
The one freshness number that decides whether AI Mode ever quotes you
My verify-or-kill checklist for using AI Mode research without publishing a hallucination
[Upgrade to read the rest →]
