Google’s user intent extraction research is often misread. The claim goes that Google Search can now guess your goal from how you use your phone. The earlier version of this article leaned that way too. The research says something narrower, and more interesting.
On 22 January 2026, Google Research described a new method. It lets small AI models work out a person’s goal from the screens they tap through [1]. The paper behind it was presented at EMNLP 2025 [3]. It is a research result, not a product, and it says nothing about Search rankings.
Updated 18 November 2026: rebuilt from the Google Research post and the full paper. We corrected how the error figures were described. We removed a dataset size and an industry claim we could not find in any source. We also separated the research from speculation about SEO.
Key takeaways
- User intent extraction here means turning a sequence of app or web screens into one sentence describing the user’s goal [1].
- Google’s method splits the job in two: summarise each screen, then extract the intent from the summaries [1].
- A small Gemini model using it beat Gemini 1.5 Pro on one test set and matched it on the other [2].
- The aim is on-device use, which avoids sending screen data to a server [1].
- Google’s post does not say this is used in Search, AI Overviews or any product [1].
Table of Contents
ToggleWhat is user intent extraction in Google’s research?
Picture someone booking a trip on their phone. They search for festivals in Europe, then look at flights to London. Google’s example is an assistant that notices this and offers to find London festivals on those dates [1].
To do that, a model must read the trail of screens and actions and state the goal. The researchers call that trail a trajectory. The output is one sentence. Google’s examples of the details such a sentence holds include “a one-way flight” and “a flight from London to Kigali” [1].
This is harder than it sounds. The paper cites earlier work on this. People writing intents for the same trajectories agreed only 80% of the time on web tasks, and 76% on phone tasks [2]. The authors treat that as a practical ceiling for any model.
The old approach against the new one
Large multimodal models can already do this well. Google’s post names the catch: they usually run on a server. That can be slow and costly, and it risks exposing sensitive information [1].
Small models fit on a phone, but they struggle when given a whole trajectory at once. Google’s answer is to break the task into two steps that a small model can handle [1].
A trajectory: screenshots and actions, one per step [1]
Three-screen window. Screen context, user action, and a guess at the goal [1]
Facts kept: 3,600 of 4,280 (16% missed) [2]
Fact counts from the paper's error analysis (recall side). Percentages use different bases and do not add up.
Stage one looks at each action with a sliding window of three screens: the one before, the current one and the next. The model answers three questions. What matters on this screen? What did the user just do? And what might they be trying to achieve [1]?
Stage two passes those summaries to a fine-tuned small model, which writes one intent sentence. Two details made a difference. The team removed facts from the training labels that the summaries did not contain, so the model would not learn to invent them. They also dropped the stage one guesses before stage two, which improved results [1].
Expert quote: “on-device intent understanding can become a building block for many assistive features” — Danielle Cohen and Yoni Halpern, Software Engineers, Google [1]
How well did small-model user intent extraction do?
| Model and method | Mind2Web (web) | AndroidControl (apps) |
|---|---|---|
| Gemini 1.5 Flash 8B, step-by-step prompting (CoT) | 0.659 | 0.594 |
| Gemini 1.5 Flash 8B, end-to-end fine-tuning | 0.653 | 0.611 |
| Gemini 1.5 Flash 8B, two-stage method | 0.752 | 0.630 |
| Qwen2 VL 7B, two-stage method | 0.623 | 0.608 |
| Gemini 1.5 Pro, step-by-step prompting (large model) | 0.730 | 0.634 |
Source: arXiv 2509.12423, Table 1 [2]. The score uses an AI judge; the paper reports a 0.016 standard deviation on repeat runs.
On Mind2Web, the small Gemini 1.5 Flash 8B model using the two-stage method scored higher than the much larger Gemini 1.5 Pro. On AndroidControl the two were level [2]. The method also beat both usual alternatives: step-by-step prompting and end-to-end fine-tuning. It did so with Qwen2 VL 7B as well as Gemini [2].
A human rater also compared 20 Mind2Web answers. They preferred the two-stage method 12 times, the prompting method 4 times, and rated 4 as equal [2]. That is a small check, and the paper presents it as one.
Where does user intent extraction still go wrong?
The paper tracks every fact through both stages. Out of 4,280 facts in the reference intents, 680, or 16%, never made it into the screen summaries. Of the 3,600 that did, 663, or 18%, were then lost when the intent was written [2].
The model also added things. Of 3,191 facts in its answers, 253 were not in the summaries at all. Of the 2,938 that were, 576, or 20%, did not match the reference intent [2]. Note the different bases: these percentages cannot simply be added up.
The authors list clear limits. Their data is mostly English and US-centric, covers only Android and the web, and Mind2Web tasks stay on one website [2]. Real people switch apps and change their minds mid-task.
Worth knowing: For the Gemini tests, the team filtered Mind2Web by domain to comply with Google’s Google-Extended policy. That cut the test set from 1,005 tasks to 681 [2]. Google-Extended is the robots.txt control that lets publishers keep content out of Gemini training and grounding [4]. Our guide to AI training opt-outs explains how to set it.
What does user intent extraction mean for SEO?
This is our analysis. Directly, very little today. The post and the paper describe research on device assistants, not search ranking [1] [2]. Nothing in them links the method to Google Search or AI Overviews.
The longer-term point is about where intent comes from. Search has always inferred search intent from the words people type. This research infers it from what people do across screens. If assistants built this way reach phones, some tasks may start from an assistant’s suggestion rather than a typed query.
That is a direction, not a forecast. Google’s own wording is a hope that it “can become a building block” as models and phones improve [1]. We would not change a content plan because of this paper.
One practical link does exist. The Google-Extended filter shows Google’s researchers respected publisher opt-outs when testing [2]. That control does not affect Search inclusion or ranking, according to Google [4].
Professional tip: When a client sends you a headline about a Google “patent” or “paper”, check two things first. Is it research or a launch? And does the source mention Search at all? Here, the answers are research, and no.
What should you do now?
- Keep writing for clear tasks. Pages that state what a user can do, step by step, are easier for any system to match to a goal.
- Make key actions obvious. Agents and assistants read screens, so an agent-accessible site helps. Clear buttons and labels help them as much as people.
- Review your Google-Extended choice. Decide whether you want your content used for Gemini training and grounding. It will not change your Search ranking [4].
- Read the source. The paper is free on arXiv [2]. Its limits section is two paragraphs long and worth the time.
What is still unknown?
- Whether Google will ship it. No product announcement links to this method (EXPECTED only as a hope in Google’s post [1]).
- How it does with real users. The tests used public datasets, not live phone use [2].
- Other languages and platforms. The authors say their data is mostly English and covers Android and the web only [2].
Frequently asked questions
What is user intent extraction?
In this research, it means reading a sequence of screens and actions in an app or website and writing one sentence that states what the user was trying to do.
Does Google use this in Search?
Google has not said so. Its research post and the paper describe the method for on-device assistants, with no mention of Search, AI Overviews or rankings.
Why does Google use small models for this?
Small models can run on a phone. Google says sending screen data to large server models can be slow, costly and risky for sensitive information.
How accurate is the method?
With Gemini 1.5 Flash 8B it scored 0.752 Bi-Fact F1 on Mind2Web, above Gemini 1.5 Pro at 0.730, and was level with it on AndroidControl.
Where can I read the paper?
It is on arXiv as 2509.12423 and in the ACL Anthology as part of EMNLP 2025, pages 18780 to 18799.
Where to go next
For intent in everyday SEO, read our guide to user intent optimisation. For where assistants and agents are heading, start with agentic web fundamentals, then see what the agentic web is On Google-Extended, see our report on Google and Bing’s shared training rule.
Sources
How we checked this: on 18 November 2026 we read Google’s research post, the full paper on arXiv and its ACL Anthology record, and Google’s crawler documentation. Every figure above comes from the paper or the post. We have no relationship with the authors.
- Google Research — Small models, big results: Achieving superior intent extraction through decomposition, Danielle Cohen and Yoni Halpern, 22 January 2026
- Cohen, Halpern et al. — Small Models, Big Results: Achieving Superior Intent Extraction through Decomposition, arXiv 2509.12423, submitted 15 September 2025
- ACL Anthology — Proceedings of EMNLP 2025, pages 18780–18799, Suzhou, November 2025
- Google — Google’s common crawlers (Google-Extended)







