Google Research Signals a Shift Toward Pre-Query Search Intent - Adil Raseed

Search is moving beyond keywords. According to new Google research, the future of search may involve understanding user intent before a query is ever typed.

In a paper presented at EMNLP 2025, Google researchers outline how small, on-device AI models can infer user intent from behavior such as taps, scrolling, and app usage — without relying on large cloud-based models or explicit search queries.

This signals a major evolution in how search systems may work in the years ahead.

From queries to inferred intent

Traditional search assumes intent begins with a query. A user types something, and the system reacts.

Google’s research flips that assumption.

Instead of waiting for a query, search systems could infer intent by observing how users interact with apps and websites over time. Actions like scrolling patterns, repeated visits, screen transitions, and clicks become signals that reveal what a user is trying to accomplish.

This approach enables search to move from reactive to anticipatory.

Why Google is focusing on small, on-device models

Large language models already demonstrate strong intent inference capabilities, but they typically run in the cloud. That introduces three major challenges:

Latency: Cloud-based models are slower
Cost: Large models are expensive to run at scale
Privacy: User behavior data is sensitive

Google’s solution is to push intent understanding onto the device itself using small multimodal models that are faster, cheaper, and privacy-preserving.

The key innovation: decomposing intent understanding

The research shows that intent extraction becomes significantly more accurate when it’s broken into smaller, structured steps.

Instead of asking a single model to reason over an entire session at once, Google separates the task into two focused stages:

Step one: Interaction summarization

Each screen interaction is summarized independently, capturing:

What appeared on screen
What action the user took
A tentative hypothesis about why

Step two: Intent consolidation

A second model reviews only the factual summaries and ignores speculative guesses. It then produces a concise explanation of the user’s overall goal for the session.

This decomposition prevents small models from failing under long or noisy interaction histories — a common limitation in compact AI systems.

How Google measures success differently

Rather than judging whether intent summaries “sound similar” to the correct answer, researchers use a method called Bi-Fact evaluation.

This approach breaks intent into discrete facts and evaluates:

Which facts were captured correctly
Which were missing
Which were hallucinated

Using F1 score as the primary metric, the decomposed system consistently outperformed other small-model approaches.

Notably:

Gemini 1.5 Flash (8B) matched the performance of Gemini 1.5 Pro on mobile behavior data
Hallucinations dropped sharply because speculative reasoning was removed
The system ran faster and cheaper than large cloud-based models

Why this matters for the future of search

This research points to a post-query future, where keywords become just one of many signals used to understand intent.

If Google wants AI agents that proactively suggest actions, answers, or next steps, it must understand what users are trying to do — not just what they type.

That requires:

Interpreting real user journeys
Understanding behavior across apps and screens
Inferring goals before explicit input

Search, in this future, becomes ambient and contextual.

What this means for SEO and digital strategy

Keywords aren’t disappearing — but they’re no longer the starting point.

In a world where intent is inferred:

Clear, logical user journeys matter more
UX and information architecture become ranking signals
Content must support understanding, not just discovery

Optimizing for search will increasingly mean optimizing how users move through experiences — not just the words at the end of the journey.

The big takeaway

Google’s research makes one thing clear:
Search is evolving from answering questions to anticipating needs.

As intent extraction moves closer to the device and earlier in the user journey, brands and marketers will need to think beyond keywords and focus on clarity, structure, and usefulness across every touchpoint.

The query won’t disappear — but it won’t be the first signal anymore.