Microsoft Patents a Way to Serve Personalized Ads Without Knowing Exactly Who You Are
Microsoft is patenting a way to deliver personalized content to web visitors even when the system can't confirm who they are, by placing probabilistic bets on multiple possible user identities at once.
How Microsoft's probabilistic ad system works without cookies
You're browsing a website and you haven't logged in. The site wants to show you a relevant ad, but without a cookie or a login, it has no firm idea who you are. That's a problem that's getting bigger as browsers block traditional tracking tools.
Microsoft's approach here is to stop demanding a certain answer and work with probable ones instead. When your browser asks a website for content, Microsoft's content delivery server assembles a short list of possible user identities you might be, each tagged with a confidence score. Then it picks one and looks up that profile to decide what ad or content to show.
The selection isn't always the most confident guess. Sometimes the system deliberately picks a random identity from the list, giving less-certain profiles a chance to prove their worth over time. That random-exploration step is a way for the system to keep learning rather than doubling down on its first instinct every time.
… determining, by the content delivery server, a set of user-probability pairs, wherein each of the user-probability pairs includes a predicted user identifier and a probability that the predicted user identifier is associated with the request from the web server …
Translation: The server guesses a list of possible users and the odds that each guess is correct.
How the objective function picks a user profile to bet on
The patent describes a content delivery server that handles ad or content requests coming from a web server on behalf of an end user. Rather than looking up a single confirmed user ID, the server generates a set of user-probability pairs: each pair contains a candidate identity and a number representing how likely that identity matches the person making the request.
Each candidate also carries an attribution value, a score that reflects how that identity has performed in the past, essentially a running grade for how profitable or accurate that guess has been. When the server needs to pick one identity to act on, it applies an objective function (a decision rule that weighs the options) in one of two modes:
- Exploit mode: pick the candidate with the highest attribution value, the system's best current guess.
- Explore mode: pick a candidate at random from the list, giving lower-ranked identities a chance to be tested.
A selection parameter controls how often the server explores versus exploits, a classic machine-learning trade-off called an explore-exploit balance. After selecting an identity, the server retrieves that profile's attributes and uses them to choose a content item from a pool of options before sending it back to the web server.
… applying an objective function to the set of user-probability pairs to select one of the user identifiers with a corresponding candidate user profile.
Translation: The system uses a math formula to pick one likely user profile from the list.
What this means for ad targeting after third-party cookies die
Third-party cookies, the backbone of web ad targeting for decades, are being phased out across major browsers. Microsoft's steady investment in post-cookie identity tech means this filing lands in a genuinely contested space. A system that can assemble a reasonable guess about who a visitor is from fragments of signal (device type, timing, browsing context) and still deliver targeted content has real commercial value.
For everyday users, this is a double-edged situation. Ads that rely on probability rather than confirmed identity could be less accurate, which some people might prefer. But the system is also designed to get better over time by learning which guesses pay off, so the targeting could end up nearly as sharp without ever needing a login or a tracker cookie.
This is the fourth Microsoft filing we've tracked in on-device AI privacy since August, following applications on running AI in the browser and catching bad auto-dubbing.
The entire system described here runs on servers Microsoft already operates, with no new devices, sensors, or physical infrastructure required. That makes the path from patent to working feature unusually short.
The core idea is that instead of knowing exactly who a user is, the system makes educated guesses and assigns each guess a confidence score, then picks which guess to act on using a method borrowed from how AI systems learn to balance trying new things against doing what already works. All of that is software running on existing ad and content delivery systems.
Probabilistic identity resolution matters right now because the advertising industry is actively scrambling to replace a decades-old tracking method that browsers are phasing out. A system that works confidently with incomplete information, rather than requiring confirmed user data, fits that gap in a practical and deployable way.
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 Microsoft patent news breaks.
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
16 drawing sheets from US 2026/0267931 A1 · click any drawing to enlarge
Show the other 8 sheets
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →
Be the first to weigh in