Microsoft Patents a Way to Compress Your Browsing History Into Sharper Content Picks
Every time a content feed guesses what you want to read next, it has to sort through everything you've read before. Microsoft has filed a patent for a system that compresses that history into a much smaller summary before making its pick, aiming to get recommendations right without the computational weight of processing everything at full size.
How Microsoft's system decides what content to show you next
Imagine you open a news app and it has to decide, in a split second, which article to show you next. To make a good guess, it wants to know what you've been reading lately, but scanning your entire history every time is slow and expensive.
Microsoft's approach is to first squeeze your recent reading history into a compact summary, then use that summary alongside a candidate article to predict how likely you are to click. The summary step is guided by the candidate article itself, so the system focuses on the parts of your history that are actually relevant to what it's considering showing you.
Embedding summarization is the name for this compression trick. Instead of feeding a full, sprawling record of your past behavior into the prediction model, the system distills it into a tight representative snapshot. The goal is faster, more accurate recommendations across products like Microsoft News, Bing, or ad placements.
obtaining a text input, the text input including a basic input and a context input corresponding to the basic input, the basic input including at least a candidate content item …
Translation: The system collects what you are looking at along with your past browsing history.
How the pooling and summary steps shrink context into a prediction
The system starts by converting two things into numerical representations called embeddings (think of these as coordinate points in a mathematical space where meaning is encoded as position). The first is the basic input, which contains the candidate content item, such as a headline being considered for recommendation. The second is the context input, which encodes the user's recent history or surrounding behavior.
The basic input's embeddings get compressed into a single pooling embedding (a pooling operation averages or selects across a sequence to produce one representative vector). That pooled value then guides a summary operation on the context embeddings, selecting the parts of the user's history that are most relevant to the candidate article, rather than treating all history equally.
The outputs of these two steps are combined into a text input representation, a single unified signal fed into a model that predicts the probability the user will click the candidate item.
- Basic input embedding + pooling step
- Context input embedding + guided summarization step
- Combined representation fed to click-probability predictor
The key claim is that the summary operation is conditioned on the candidate content, so the history compression is always tailored to what's being evaluated, not generic.
A representative embedding sequence corresponding to the context input may be obtained through performing a summary operation on the embedding sequence corresponding to the context input.
Translation: It condenses your browsing history down into a shorter summary of your interests.
What this means for Microsoft's news, search, and ad products
Microsoft runs recommendation surfaces across Bing search, the MSN news feed, and the advertising systems that sit behind many websites. Any improvement in how efficiently those systems process user context can translate to lower server costs at enormous scale, and potentially to recommendations that feel more on-target.
For you as a reader or user, the practical effect would be a feed that accounts for your recent interests without the lag or battery drain that comes from heavier processing. Microsoft's filing activity around AI-driven personalization suggests the company is investing in making its recommendation infrastructure leaner, not just more capable.
That makes this Microsoft's 53rd filing we've tracked in Language AI since May, a corpus that includes one turning fuzzy dates precise and one writing captions from photos.
The system decides what to remember about your reading history based on what it's already thinking about showing you. That circular logic has a real cost: clues pointing toward something unexpected get discarded before they ever get a chance to influence the result.
The tradeoff is defensible. Checking your full history against every possible article, millions of times per second, is not feasible, so compression is unavoidable, and letting the candidate article guide that compression is smarter than compressing randomly.
The honest question is whether the compression is good enough. If it consistently drops the weak, wandering signals that would have led somewhere surprising, the system gets faster without getting broader, and that is how recommendation feeds slowly narrow without anyone deciding they should. This patent makes the tradeoff workable but does not claim to solve it, and that distinction matters.
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
7 drawing sheets from US 2026/0267894 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →