Answer before the query

Google Instant (2010–2017) rendered result pages as you typed; it died on mobile — a results page on top of a phone keyboard was a mess.

Google's intelligent Search box (May 2026) predicts intent as you type — but its output is a better question. Every current engine, AI Mode included, still waits for a submitted query.

Prediction is spent on the wrong side. A model that knows the time, the location, what you did a minute ago and the four letters typed so far doesn't need the full query to know the outcome. "amer…" at 8:40 next to a café is an americano, not Vespucci.

Generate the outcome, not the sentence.

1. Branch tree instead of a list. Each keystroke narrows a small set of live outcomes — three or four cards ranked by context. Wrong branches fade; the page never reloads.
2. Every card is an answer plus one action. Ratio · how to make · cheapest nearby · order, with the delivery partner resolved per city.
3. Mobile-first this time. Cards live above the keyboard, in the space autocomplete already occupies. This is exactly where Instant broke in 2010.
4. Monetization. Advertisers bid on intent branches, not keywords. "Order" under a drink is an inventory slot with conversion data from day one.

Whether pre-generation is affordable at Google's volume, or only for high-frequency heads. How much context users tolerate before it feels like surveillance. Whether a wrong prediction costs more trust than no prediction — the "not this" control in the prototype exists to measure that.

These are the things to test before any bet.

Put a live model behind the prototype on a narrow slice — food and drink, one city. Measure time-to-first-useful-card against Enter-based search, the cancel rate of predictions, and tap-through on the action card.

Status: concept, scripted prototype. Next: live scoring.