Google
Google · AI / ML · Filed Nov 25, 2025 · USPTO published Nov 12, 2026 · This week's release
Google Patents a Recommendation Engine That Treats Every Clue About You Differently
Your phone suggests a food delivery app at lunchtime because the system learned that the hour and your country matter in different ways for different apps. A new Google patent application describes a way to read each of those clues differently depending on what it is being paired with.
- Publication number
- US 2026/0359308 A1(US20260359308A1)
- Applicant
- Google LLC
- Filed · Published
- Nov 25, 2025 · Nov 12, 2026
- Inventors
- Huan Gui, Lichan Hong, Ed Huai-hsin Chi, Taibai Xu, Ruoxi Wang, Ke Yin, Maciej Kula, Long Jin
- Status when we published
- Examiner raised objections; the company can respond
- Document
- 20 claims
- US classification
- 707/741
- Examiner
- KUDDUS, DANIEL A (Art Unit 2154)
- Parent application
- National Stage Entry of PCT/US2024/030365 (filed May 21, 2024)(PCTUS2024030365)
What Google's recommendation patent actually does
Ever notice your app store suggesting a food delivery app right around lunchtime? Behind that nudge is a system guessing what you will tap, using clues like the app, the hour, and the country you are in.
Google's filing points out a snag. Everything the system knows about an app, say Zomato, gets bundled into one summary: what the app is, where it is popular, when people use it, what language its users speak. But when the system asks does this app fit this country?, only the where-it-is-popular part matters. When it asks about the hour, the when-people-use-it part matters.
So Google describes a design where each clue gets its own way of being read, and every pairing of clues gets its own tailored comparison. The goal is picks you are more likely to download, produced fast enough to run live.
… learns individual query, key, and value projection matrices for each individual feature embedding in each pair of feature embeddings; …
Translation: The system builds custom mathematical weights for every single piece of data it collects about you.
How Google gives every clue its own lens
The patent covers a recommendation system built on a transformer, an AI model that uses 'attention' (a way of scoring which pieces of input deserve the most weight). It starts with records of how a person reacted to items recommended before. Those records include categories (like product type), plain numbers (like price or age), and task embeddings (lists of numbers standing in for a goal, such as a click or an install). All of it becomes a list of embeddings, which are numeric fingerprints.
A standard transformer runs every item through the same math. Here the model learns separate query, key and value matrices (three sets of numbers used to ask a question, match it, and carry the answer) for each individual clue. That gives every pair of clues its own transformation matrix, which moves both into a shared space so they can be compared. The comparison is lopsided on purpose: how an app looks from the hour's side differs from how the hour looks from the app's side.
The rest of the pipeline:
- The per-clue matrices are joined into one big matrix, and a linear transformation (a simple mathematical reshuffle) produces a full projection matrix, so clues can mix more freely.
- A one-layer interaction transformer lets the task embeddings act as the questions and the combined clues act as the answers, producing weighted scores.
- A neural network turns the weighted task embeddings into the final list of recommended items.
… generates a set of recommended items based on applying a projection function to the weighted task embeddings.
Translation: It finally produces your personalized suggestions after crunching all those customized data weights together.
Why Google says its bigger model stays fast
If you have ever been shown the same irrelevant suggestion again and again, this is the part of a system meant to fix it. The filing says handling how clues combine more precisely should mean picks you are more likely to tap, download, or watch. Clicks, installs, and watch time are the examples it names for what the system can be trained to predict.
Speed is the business case. Recommenders answer huge numbers of people in a blink, so the filing stresses that the design needs the same amount of math as a standard layer and trims it further with low-rank shortcuts (compressed versions of the matrices). The examiner's first review rejected the claims as filed, which is routine at this stage and usually followed by amendments.
Google files its 28th patent in the AI recommendation space we've tracked since May, building on earlier applications like one that reads sites for keywords and one that picks ad layouts.
Source: US 2026/0359308 A1 at the USPTO (PDF). Full patent text and figures are in the official publication.
Get our take in your Top Stories
Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time Google patent news breaks.
Add Patentlyze on Google
One click.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings · 5 sheets
5 drawing sheets from US 2026/0359308 A1 · click any drawing to enlarge
More from Google
All 1,079 Google patents →
More in AI recommendation
All 86 AI recommendation patents →
Intel
Intel Patents a System That Connects Similar Browsing Sessions to Sharpen Recommendations
Microsoft
New Microsoft Patent Tracks Cross-App Behavior to Sharpen Recommendations
AMD
AMD Patents a System That Moves the Most-Used AI Data to Faster Storage Automatically
Sony
No comments yet