Get the newsletter
Microsoft · AI / ML · Filed May 19, 2025 · USPTO published Nov 19, 2026 · This week's release

Microsoft Files Patent for AI That Stops Advertisers From Overpaying for Web Ads

Every time a web page loads, advertisers fight in a split-second auction for the ad spot, and Microsoft's new patent application tries to help them win without overpaying. It uses one AI model trained on winning bids and another trained on losing ones.

A tablet feed displaying sponsored content sent from remote servers over the web. Drawing from patent filing US 2026/0368781 A1.
A tablet feed displaying sponsored content sent from remote servers over the web.All 10 drawings ↓
Publication number
US 2026/0368781 A1(US20260368781A1)
Applicant
Microsoft Technology Licensing, LLC
Filed · Published
May 19, 2025 · Nov 19, 2026
Inventors
Ruby Tu, Shengjun Pan, Min Liu, Xi Xiong, Yi Zhang, Yuan Gao, Haipeng Gao
Status when we published
Rejected again; the company can still amend or appeal
Document
19 claims
US classification
705/14.71
Examiner
CIRNU, ALEXANDRU (Art Unit 3622)

What Microsoft's two-model ad bidding does

Imagine you're at an auction where you have to guess the lowest price that still wins, but you only ever get told what the winning bids looked like. Nobody shows you the bids that came up short. You'd probably pay too much just to be safe.

That is the situation behind many of the ads on the pages you load. Each time a page opens, an automatic auction runs in milliseconds for the ad slot, and the advertiser's software has to name a price instantly. Microsoft's patent application describes two AI models: one learns from past winning bids and sets a ceiling, the other learns from past losing bids and sets a floor.

The final price lands somewhere in between, and it shifts based on how recent auctions went. Lose one and the next bid creeps up; win one and it creeps down. The goal is to pay just enough to win, and no more.

From the filing · CLAIM 1
… generating a bid upper boundary value for a prediction interval based on the input vector using a first machine learning (ML) model, the first ML model trained on a first set of training datapoints, wherein each first training datapoint comprises an accepted bid from a first set of auctions …

Translation: The system uses one AI model trained on past winning bids to figure out the maximum amount you should consider paying.

How the floor and ceiling prices get set

The patent application covers a method for an advertiser's bidding software (called a demand-side platform, or DSP, a service that buys ad space on behalf of companies). When a web page loads, the software receives a bid request: details about the page, the ad slot, and the visitor's device and context, bundled into a list of numbers (an input vector).

Two machine learning models then each produce one end of a price range:

  • Accepted bid model: trained on bids that won past auctions. It sets the upper boundary, a price likely high enough to win.
  • Rejected bid model: trained on bids that lost. It sets the lower boundary, a price below which a bid probably fails.

Bid selection logic then picks one price from inside that range. An exploration parameter (a dial that decides whether to lean toward the high end or the low end) is adjusted using live auction results and performance numbers. If the last bid lost, the dial moves up; if it won, it moves down. The claims also cover the last steps: getting notified of a win and sending the ad to appear as the page loads.

The filing also describes training the models with a custom scoring rule (a loss function, the formula that tells a model how wrong its guess was) that punishes some kinds of wrong guesses more heavily than others. It says the two models could also be merged into one.

Why ad overbidding costs advertisers real money

If it works, advertisers would stop paying more than they have to for the same ad spots. The filing's own example imagines a day with over 2.3 billion auctions, where an advertiser won about 141.9 million of them and paid roughly $6.2 million. Winning the same number in cheaper tiers could have cost about $1.25 million, a cut of over 70%. That is an illustration from the company, not a measured result.

The patent office has issued a final rejection of this application, so nothing here is approved. Microsoft can still amend the claims or appeal. Even if it never becomes a granted patent, the filing shows how Microsoft approaches the problem of overpaying in automated ad auctions.

That makes this Microsoft's 33rd filing we've tracked under AI training and infrastructure since May, following work on adding specialties to one model and stopping game sharpening AI.

Source: US 2026/0368781 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 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 · 10 sheets

10 drawing sheets from US 2026/0368781 A1 · click any drawing to enlarge

How web ads are auctioned off in real time, drawing sheet 1 from US 2026/0368781 A1
Sheet 1 How web ads are auctioned off in real time
Using winning and losing auction history to price bids, drawing sheet 2 from US 2026/0368781 A1
Sheet 2 Using winning and losing auction history to price bids
Combining auction data to estimate price ranges, drawing sheet 3 from US 2026/0368781 A1
Sheet 3 Combining auction data to estimate price ranges
The AI model that predicts bid outcomes, drawing sheet 4 from US 2026/0368781 A1
Sheet 4 The AI model that predicts bid outcomes
Training the AI on past accepted and rejected bids, drawing sheet 5 from US 2026/0368781 A1
Sheet 5 Training the AI on past accepted and rejected bids
Server components that generate ad bids, drawing sheet 6 from US 2026/0368781 A1
Sheet 6 Server components that generate ad bids
Steps for picking an ideal bid amount, drawing sheet 7 from US 2026/0368781 A1
Sheet 7 Steps for picking an ideal bid amount
Refining the bidding AI to improve price accuracy, drawing sheet 8 from US 2026/0368781 A1
Sheet 8 Refining the bidding AI to improve price accuracy
Show the other 2 sheets
How ad servers deliver targeted content to a device, drawing sheet 9 from US 2026/0368781 A1
Sheet 9 How ad servers deliver targeted content to a device
Hardware chips that run the bidding and inferencing servers, drawing sheet 10 from US 2026/0368781 A1
Sheet 10 Hardware chips that run the bidding and inferencing servers
Patent filing page

More in AI training & infrastructure

All 323 AI training & infrastructure patents →
Reader comments

No comments yet

Start the discussion

Real name or a handle, either is fine. Comments are read by a person before they appear, so allow a little time. Keep it about the filing.