New Google Patents · Filed Oct 21, 2024 · Published Jul 2, 2026 · verified — real USPTO data

Google Patent Reveals Hidden AI Signals That Could Shape Ads You See

Google has figured out a way to read the hidden internal state of an AI chatbot mid-conversation and use that signal to decide which ad to show you, without waiting to fully analyze what you typed.

Google Patent: AI Chatbot Ad Targeting via Neural Network Signals — figure from US 2026/0187172 A1
FIG. 1A — rendered from the official USPTO publication PDF.
Publication number US 2026/0187172 A1
Applicant GOOGLE LLC
Filing date Oct 21, 2024
Publication date Jul 2, 2026
Inventors Nicholas James Reid
CPC classification 707/732
Grant likelihood Medium
Examiner TO, BAOQUOC N (Art Unit 2154)
Status Non Final Action Mailed (May 7, 2026)
Parent application is a National Stage Entry of PCTUS2023028041 (filed 2023-07-18)
Document 21 claims

How Google's AI picks ads from inside a chatbot conversation

Imagine you ask an AI assistant a question like "what's the best running shoe for knee pain?" Google wants to know more than just the words you used. It wants to understand where you are in your decision process: are you just starting to research, or are you ready to buy?

This patent describes a system that reads a kind of internal snapshot from the AI model while it is generating your answer. That snapshot captures subtle signals about your intent that the words alone might not reveal. Based on that, the system maps your position onto a "user journey" (think: awareness, consideration, purchase) and picks an ad that fits that stage.

The result is that the ad appearing alongside the AI's response isn't just matched to your keywords. It is matched to what the AI's own internal workings suggest about where you are in making a decision. Google is essentially letting the chatbot's brain, not just its output, drive ad targeting.

How the intermediate embedding drives ad selection

The patent describes a pipeline that intercepts a signal called an intermediate embedding from inside a language model neural network. An embedding is a dense numerical representation that a neural network builds up internally as it processes text. An intermediate embedding is captured partway through that process, before the model finishes generating its response, and it tends to encode richer contextual meaning than the raw input text alone.

That embedding is then fed into a separate classification step that maps it to one or more stages in a user journey. A user journey here is a structured sequence of states a person moves through when completing a task online (for example: researching a product, comparing options, deciding to purchase). Google maps candidate digital components, meaning ads or other content units, to each stage in advance.

Once the system determines which stage best matches the embedding, it selects the corresponding digital components (the patent's term for ads or sponsored content) and surfaces them alongside the AI-generated response.

  • The language model processes the user's prompt as normal
  • An intermediate embedding is captured mid-process
  • That embedding is classified against predefined user-journey stages
  • Ads mapped to the matching stage are selected and displayed with the response

What this means for ads inside Google's AI products

For users, this means the ads shown next to an AI response could be significantly more context-aware than what keyword-based targeting delivers. A query about "knee pain running shoes" at the research stage might surface an informational comparison tool; the same query phrased as "buy Brooks Ghost size 10" might surface a direct retailer ad. The system is designed to make that distinction automatically, based on the AI's internal read of your intent rather than a rules-based keyword match.

For Google's business, this is about making AI-generated search results monetizable in a precise way. As AI answers displace traditional search result pages, the old click-on-a-blue-link ad model loses relevance. This patent describes one concrete mechanism for keeping ad revenue tied to AI-driven interactions, which is the direction Google's core products are clearly heading.

Editorial take

This is one of the more technically specific ad-monetization patents Google has filed in the AI era, and it matters because it shows a real engineering path, not just a strategy slide, toward making AI search pay for itself. The intermediate-embedding approach is clever: it gives the ad system a richer signal than keyword matching without requiring a second full inference pass. Whether it ships as described is another question, but the underlying problem it solves (how do you sell ads next to an AI answer?) is very real and very urgent for Google.

Which company should we read for you?

We track 17 companies here. Pro is the same weekly breakdown for any company you choose, delivered privately. Type a name and we'll scope it and send you a quote.

Get one Big Tech patent every Sunday

Plain English, intelligent commentary, no hype. Free.

Source. Full patent text and figures from the official USPTO publication PDF.

Editorial commentary on a publicly published patent application. Not legal advice.