Liz Reid, VP of Search at Google, announced at I/O 2026 that AI Mode hit 1 billion monthly users in its first year, with queries doubling every quarter. Source: Google's announcement post. The headline narrative — that AI is replacing Search — turned out to be wrong. AI is what is growing it.
The product reveal that matters is not the user count. It is what Search now does with them.
Mini apps, generated on demand
Google brought Antigravity, their agentic coding platform, into Search. You can ask Search to build you a fitness tracker. It writes one. Pulls real-time data, the local weather, reviews, live maps — gives you a custom dashboard that persists across sessions. Not a one-shot answer. A returnable workspace. Generative UI for one-off questions (visualizing how a watch works) and persistent dashboards for ongoing tasks (wedding planning, home moves).
US-only this summer for AI Pro and Ultra subscribers. Free generative UI for individual queries rolls out globally the same window.
I have been writing backend code for two decades. The first reaction to apps generated on the fly is skepticism — that phrase has lied to me before. The second reaction, after reading the announcement carefully, is that this is the most interesting Search product change in a decade. Not because the generated apps will be good. They will not be, often. Because the existence of generated apps changes the math on what is worth shipping yourself.
What the SERP can fake now
Three product categories just got harder to justify as standalone apps:
- Lightweight personal trackers. Habits, light fitness, simple budgets. If Search can spin up a workable version on demand, you need a 10x differentiator to charge for one.
- Comparison and decision-support tools. Search already handled vs queries. Now it generates side-by-side live dashboards in the answer.
- Personalized monitoring. Information agents — also announced today — do this natively. Alert me when this rental matches my criteria. Notify me when these sneakers drop. No app needed.
Three categories that got more defensible, not less:
- Deep workflows. Search can ship a tracker. It cannot ship a CRM with role-based access control, audit trails, and provisioning.
- Regulated software. Search-generated UI does not sign BAAs. Healthcare, finance, legal — the compliance overhead is the moat.
- Proprietary systems of record. When the UI layer commoditizes, the data layer underneath gets more valuable, not less.
What breaks
Generative UI is impressive in demos. Production engineering is where it gets hard:
- State persistence. Where does the wedding-planning dashboard actually live? On Google's infrastructure, presumably. What happens when your data is in someone's prompt history.
- Source of truth. The generated tracker pulls live data. From where? With what update cadence? What happens when sources disagree.
- Failure modes. A real app has error handling. A generated app has I am sorry, let me try again.
- Integration. Real apps connect to Stripe, Calendly, Plaid, your existing systems. Mini-apps in Search do not. Yet.
These are not reasons to dismiss the announcement. They are the seams where products built by humans still win.
What I would change on a backend project this week
- Design APIs for agent consumption, not just browsers. Predictable pagination, idempotent endpoints, structured responses, explicit rate limits. The next caller of your API is going to be an LLM, not a user.
- Ship semantic HTML and structured data on everything. Same conclusion the marketing team reached for a different reason. For backend engineers it is: if Search is going to cite your product, it has to read your product first. Server-rendered, machine-readable, no critical content trapped in client-side hydration.
- Audit your roadmap against what Search now does for free. If Search can ship the feature you are scoping in a query, the feature is not a moat. The moat is the system underneath.
The honest take
This is not AGI in the search box. It is Google taking an explicit position on what it wants to own: the entry point, the agentic layer, and now the lightweight app layer too. Everything above the data is in play.
Build the data. Ship the integrations. Make sure the system underneath the UI can survive when the UI gets generated by someone else.
Last updated May 20, 2026