Google Patents Technology That Builds and Scores Ads Without Human Input
Google has filed a patent describing a system where multiple AI agents divide up the work of writing, scoring, and assembling digital ads entirely on their own. No human copywriter required at any step.
How Google's AI agents would auto-create ads for you
Ever tried to get a decent ad written fast, only to spend hours briefing a team, reviewing drafts, and going back and forth on which version works best? Google is patenting a system that hands all of that to a team of AI agents.
Here is how it plays out for you: one AI agent reads your goal, breaks it into a to-do list, and another agent sorts that list by priority. A third agent actually does the work, pulling on a shared memory bank of everything the system has already learned and produced. Each candidate ad gets a score, and the best pieces get assembled into a finished result.
No single agent handles everything. The design deliberately splits the job so each agent stays focused, which is meant to make the whole process faster and less error-prone than asking one AI to do it all.
a task creation agent configured to receive a user query indicating an objective and generate a plurality of tasks to achieve the objective, wherein the plurality of tasks includes at least a first task for generating a digital component …
Translation: An automated planner breaks down a user goal into smaller jobs, including making an ad.
Inside Google's task-queue and shared-memory agent loop
The patent describes a multi-agent architecture (a setup where several AI programs each handle a specific role rather than one program doing everything).
The three main agents are:
- Task creation agent: takes a user's goal as input and breaks it into concrete sub-tasks, such as generating a headline or scoring a banner image.
- Task prioritization agent: maps out which tasks depend on other tasks finishing first, assigns each a priority level, and routes them into separate queues (think of queues like different lanes at a toll plaza, each moving at its own pace).
- Execution agent: pulls tasks from the queues, checks a shared memory store for relevant context (previous results, user info, prior decisions), runs a generative AI model to produce a candidate ad component, then writes the result back into shared memory so the other agents can use it.
The shared memory is the connective tissue. Because every agent reads from and writes to the same store, a result produced by one execution run automatically becomes available context for the next. The system scores each candidate component and selects the best ones to combine into a final output.
The patent covers both the orchestration logic and the scoring step, meaning the agents are not just creating content, they are also judging it.
… determining, using the one or more autonomous agents, digital component scores of the plurality of candidate digital components …
Translation: Software bots automatically rate how well different ad options perform.
What automated ad creation means for advertisers and publishers
For anyone who buys or runs digital advertising, this describes a potential path to a fully automated creative pipeline. You give the system an objective, and it handles ideation, drafting, evaluation, and assembly. That kind of end-to-end automation could shrink the time between briefing and live ad from days to minutes.
Google's steady filing of multi-agent AI systems is a pattern worth understanding in context: Google operates one of the world's largest digital ad platforms, so a system that lowers the cost and time of ad production has an obvious home there. Whether this becomes a feature inside Google Ads or a tool for agencies, the practical pressure it puts on human creative teams is real.
This is the 17th Google filing we've tracked since May on AI models working in teams, after a self-grading image system and a voice AI that overrides itself.
The shortest path from this patent to a shippable product is actually not that long, because almost everything it describes is software running on existing infrastructure. There is no new hardware required, no novel chip, no physical sensor. The building blocks (large language models, task queues, memory stores) all exist today.
The harder part is the scoring step. The patent describes agents assigning quality scores to candidate ads, but it does not spell out what those scores measure or how reliable they are. If the scoring is weak, the whole selection loop falls apart and you just get a fast way to produce mediocre output at scale.
The design choice to split work across specialized agents rather than using one general-purpose AI is a real architectural decision with tradeoffs. It can improve speed and reliability, but it also adds coordination overhead. The patent is more of a systems blueprint than a research breakthrough, which means the interesting question is execution, not invention.
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 2026/0270224 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →