What made Google's AI understand users better wasn't more data

The question: What actually improves what an AI can tell you about your users: more data, better input or a bigger model?

Dataset & sample

Source

Wang et al., “LLM-Based User Personas for Recommendations at Scale”, arXiv 2606.12198 (Google, Google DeepMind)

Sample

Offline evaluation against topics clicked by hundreds of users; user survey with thousands of US users

Primary metric

BLEURT score against clicked topics; instruction-following rate

In June, a team at Google and Google DeepMind published how they built an LLM that writes, in plain text, what each user of a video platform is interested in, built to run for billions of users. In the last field note I wrote about what happened when it went live.

This one is about what happened before. The team ran a set of tests that most companies skip when they plug AI into their data. Not "which model should we use?", but "what exactly do we put in front of it?"

The answers are the most useful thing I've read this year about feeding data to AI. And they apply to your analytics almost line by line.

How they tested it

The model reads someone's watch history and returns a few short interest labels. To score it, they used a ground truth that doesn't depend on the model: topics that hundreds of real users had clicked on the platform. A good summary should match what those people actually chose.

Then they changed one thing at a time. Which text describes each video. How the history is arranged. Which model reads it. That's the part worth copying, before any of the results.

1. The cleanest field beat the richest one

They tried three ways to describe each watched video: its description, a list of weighted keywords, and its title.

The title won. The description has the most text, but much of it isn't about the video: promotional copy, links, copyright notes, background music credits. The keywords failed the other way: too broad to say anything specific.

Your analytics has the same three options. The "descriptions" are what gets collected just in case: full URLs with tracking parameters, auto-captured clicks, event properties nobody defined. The "titles" are the few fields someone named on purpose: page type, product category, a campaign name that follows a convention.

When you hand data to an AI, or to a person, hand it the titles.

2. Group by meaning, not by appearance

Same users, same model. The only change was how the watch history was arranged before the model read it.

Source: Wang et al., arXiv 2606.12198, Table 1.

The first answer is decoration. The second is something you can act on. It took zero extra data.

There's a subtler result next to it. They compared two ways of grouping. One clustered videos by what they look and sound like (audio-visual similarity). The other clustered them by what they're about. Grouping by topic produced better summaries, because a cluster of "videos that look alike" has nothing in common that you can name.

Most analytics reports group by appearance: URL folder, page template, device, channel. Those groups are easy to build and hard to reason about. An AI, like a person, gives specific answers when the rows are grouped by intent: discovering, comparing, buying, coming back.