Google claims training a machine-learned model to fold third-party content, such as advertisements, into its answers by weighing a learned satisfaction score against predicted interaction with that content

2026-11-09Filings

Listen · Thinking Makes It So, 2026-11-09 · 9 min

All episodes · Subscribe: Apple Podcasts · Spotify

United States Patent Application Publication No. 2026/0350464 A1, Google LLC, "Multi-Stage Training Process for Sequence Models," Application No. 19/196,371, filed May 1, 2025, published November 5, 2026

Claim 1 trains two models. A reward model is trained with reinforcement learning from human feedback to predict a user satisfaction metric for an output that contains a third-party content item, an output the claim says the reward model itself generates from a prompt holding a user query and one or more such items. A second model is then trained with a loss function that combines the reward model's satisfaction metric with predicted interaction metrics for a third-party content item placed in the second model's output, and the trained model answers a later query with natural language output combined with one or more third-party content items. The specification says third-party content items can include sponsored content such as advertisements, gives predicted revenue and click-through rate as examples of interaction metrics, and states the technical problem as balancing user satisfaction and third-party content presentation without one taking over the other. The USPTO's Open Data Portal shows the application awaiting its first examination.

What it changesMy reading is that Google is claiming the training that teaches a model where third-party content belongs in an answer, not only the act of inserting it, and the revenue and click-through examples make advertising the obvious use. A training-method claim will be hard to enforce if it issues, because training happens on the other party's servers and only the last step of claim 1, an answer that combines text with third-party items, can be seen from outside; for now the claims are pending, and nothing can be enforced unless a patent issues. The published claim set also has mechanical problems an examiner can be expected to raise. Claims 1, 8 and 16 list the loss function's components as "(i)" and "(iii)" with no "(ii)"; claim 16 covers "transitory or non-transitory" computer readable media, which the Office rejects under 35 U.S.C. § 101 because a transitory signal is not within the statutory categories (In re Nuijten); and claims 17 and 19, and through them claims 18 and 20, refer to "the one or more non-transitory computer readable media of claim 16," a narrower term than the "transitory or non-transitory" media claim 16 recites. Each is a routine fix by amendment, and the fix to claim 16 narrows it to non-transitory media, which is where it belonged from the start.

More on What the AI companies are filing · Eligibility of software and AI claims

← All recent developments · Search all developments

Burbage Law is the practice of Jon Burbage, a registered patent attorney. 342 United States patents have issued from applications he wrote; the full list is on the Patents page. More about Jon Burbage.

This is general information about a published decision, agency guidance or piece of legislation. It is not legal advice and does not create an attorney-client relationship. Items concerning offices outside the United States are reported for information only and are not counsel on the law of those countries.