A newly granted Google patent describes a fast, purpose-built AI model that rewrites one vague query into several targeted searches, fires them off at once, and arranges the answers into themed sections on a single page.
On March 17, 2026, the U.S. Patent Office granted Google patent US12579158B1, “Dynamic organization of search results.” Filed in December 2024 by eight Google inventors, it describes how a search engine could handle open-ended questions, the kind with no single right answer, by returning a page organized around the different things a user might actually want.
In short: Google has patented a system where a small, fast AI model reads a broad query, splits it into several specific sub-searches, sends each to a specialist service such as Maps, recipes, videos or forums, and shows the answers as themed sections. For SEO and AEO, visibility increasingly depends on matching those hidden sub-queries, not just the words people type.
One Query, Many Meanings
Search engines already lean on specialized “topical search services” for places, videos, recipes, forums, restaurants, events and shopping. Today these are switched on by rigid triggers: a keyword match, a regular expression, or a classifier that decides a query looks like a places query. When a trigger fires, the user’s exact words go to each service unchanged.
The patent lists three weaknesses. Open-ended queries like “what to eat before a marathon” often trigger nothing, because they fit no template. Services that do fire work blind to each other, so results overlap. And context not typed into the query, such as time of day or location, is ignored. Users then search again to get the variety they wanted.
Google also explains why it didn’t just hand the job to a general-purpose large language model. Such models, the filing says, are too slow for the sub-second pages users expect, expensive to run, and prone to hallucination because they are built to be creative as well as factual.
Meet the Dynamic Result Model
The core of the invention is a “dynamic result model”: a transformer-based generative model, built like an LLM but trained for one narrow job. Because the task is so specific, the patent says it can answer in milliseconds, and it can be taught with very few examples, sometimes fewer than 50.
The model takes the user’s query plus its context: time of day, nearby holidays or events, weather and a generalized location, and, with permission, search history and interests. It returns a list of “topic objects.” Each pairs a rewritten “topic query” with one specialized service, and can carry a generated title, description and justification.
Figure 1 · How one query becomes several searches
For the marathon question, the model might send “healthy restaurants” to the places service and “pasta carb loading recipes” to the recipe service. It can call the same service twice from different angles, and because it writes each topic knowing the others, it steers away from duplicates.
Context shapes the rewrites. “Restaurants” in the morning can become a breakfast search, and in the evening a dinner search. “Events” can mean tourist attractions for a traveler but local events matching known interests for someone at home. One justification in the patent suggests rooftop restaurants because good weather is forecast for the night of the reservation.
Building the Results Page
Before any searching, weak topic objects can be filtered out: those with low confidence scores, those too far from the original query, near-duplicates of the same type, or ones excluded for policy reasons. The rest are sent to their services in parallel, which keeps the wait short.
Results from each topic query stay together in a “rich result listing,” a section of the page headed by the model’s title and, optionally, its description. Listings can follow the model’s order and sit between ordinary web links and other page elements. A listing is dropped if its service returns too few or too weak results.
The timing rule is central to what Google actually claims. The system sets a time limit after sending the topic queries. Whatever has arrived by then appears on the first page; slower services are added in a later update, so a fast service can appear above one the model ranked higher.
Figure 2 · The time limit that splits the page into two deliveries
The same “fill in later” approach covers extras such as LLM-written summaries of individual results, which are too slow to produce before the page first loads.
The patent’s drawings use the query “anniversary celebration dinner in Dallas.” The page shows a places listing, a forums listing, and a second places listing built from a different topic query that reflects the user’s interests, with a generated description.
Figure 3 · Results page for the patent’s anniversary dinner example
A Model That Learns From Clicks
The system isn’t meant for every search. Simple factual questions, answerable in a sentence, can bypass the model entirely; it is reserved for queries with several aspects, nuanced intent or real ambiguity.
Popular queries get a shortcut. If a query is searched often enough, its listings are saved and reused for later queries that mean roughly the same thing, skipping the model. And with permission, the system records which listings people scroll through or click, then uses that as a reward signal for reinforcement learning: topics that draw clicks are favored next time, ignored ones are demoted.
Figure 4 · Where the model is skipped, and how it learns
The patent contrasts this with today’s search pipelines, which it describes as cascades of hard-coded services that are difficult to tune. A single model can instead be tuned end to end. Privacy controls round it out: opt-in data collection, stripped personal identifiers, and locations generalized to city or ZIP-code level.
What It Means for SEO
- You compete for the hidden sub-queries, not just the typed keyword. A search for “what to eat before a marathon” might never reach your page as typed. It reaches you as “pasta carb loading recipes” or “healthy restaurants.” Keyword research should map the specific follow-up questions behind each broad topic, the same way the model does.
- The same keyword can produce different sub-queries. The model rewrites queries using time of day, location, weather, holidays and personal interests. “Restaurants” at 8 a.m. becomes a breakfast search; at 8 p.m. a dinner search. Content that covers these real-world situations has more chances to match.
- Specialist surfaces are the way in. Every themed section comes from a topical service: places, videos, recipes, forums, events, shopping. If your business isn’t in those services, through a complete Business Profile, indexed videos, recipe or event markup, or real discussion on forums, there is no listing to be picked from.
- Near-duplicate angles get filtered. The system drops sub-queries that are too similar to each other, aiming for variety on the page. A distinct angle, audience or use case is more valuable than another version of the same top-ten list.
- Clicks train the model. Listings people click are favored for similar future searches, and ignored ones are demoted. Titles, images and snippets that earn the click help decide which topics appear next time.
What It Means for AEO
Answer engine optimization (AEO) is about becoming the answer an AI system chooses and presents. In this patent, the AI plans the answer instead of writing it: it decides which questions to ask, then pulls real results to fill each section.
- Be the best result for one clear sub-question. Each themed section answers a narrow question. Pages that answer one specific question directly, near the top, are easier to slot into a section than pages that circle a broad topic.
- Make your details explicit. The model writes titles and justifications such as “rooftop restaurants, because good weather is forecast.” It can only justify you with facts it can find. State attributes plainly: outdoor seating, quiet atmosphere, vegan options, good for groups, open late.
- Write so a summary is easy. The patent describes AI-written summaries added to individual results. Clear headings, short factual paragraphs and plain claims give a summarizer accurate material instead of guesses.
- Use structured data. Schema such as LocalBusiness, Recipe, Event, VideoObject and Product helps the specialist services understand what you are, which is how you become eligible for their sections.
How to Prepare
- List the 5 to 10 specific sub-questions behind each broad topic you target, then check whether you have a page that answers each one.
- Cover the situations that change intent: time of day, season, weather, local events and holidays.
- Keep your Google Business Profile complete and current, including attributes, hours, photos and categories.
- Add accurate structured data for recipes, events, videos, products and local business details.
- Publish video and take part in forum discussions where your audience asks questions.
- Give each page one distinct angle instead of repeating what already ranks.
- Put the direct answer first, then the supporting detail, under clear question-style headings.
- Write titles and thumbnails that earn clicks honestly, since engagement feeds back into what Google shows.
Frequently Asked Questions
What Is Google Patent US12579158B1?
It is a Google patent, granted on March 17, 2026, for a search system that uses a small AI model to split one broad query into several targeted searches and show the results as themed sections.
Does This Mean Google Already Uses It?
Not necessarily. A patent protects an idea; it does not confirm the feature is live. It does match Google’s move toward AI-organized results for broad, exploratory searches.
Which Searches Does It Affect?
Open-ended searches with several possible intents, such as planning a meal, a trip or a celebration. Simple factual questions can skip the system entirely.
Will This Reduce Traffic From Regular Web Results?
The themed sections sit alongside ordinary web links, so classic rankings still matter. But more of the page may go to places, videos, recipes and forums, so sites present in those surfaces gain extra ways to appear.
What Is the Single Most Important Change to Make?
Stop optimizing only for the broad keyword. Identify the specific questions hidden inside it and make sure you have a clear, well-marked-up answer for each one.
Why It Matters
The patent shows how Google thinks about bringing generative AI into search without the drawbacks of large chatbots. Here the model doesn’t write the answer; it acts as a fast query planner, breaking a fuzzy request into precise searches and letting established, fact-based services supply the results.
Google argues this cuts hallucination risk, keeps pages fast, and spares users follow-up searches. As with any patent, it doesn’t confirm Google has shipped or will ship this exact system. But it closely matches the broader move toward AI-organized, themed results pages for broad, exploratory searches.
Source: US12579158B1, Google Patents. Figures are illustrations based on the patent’s description, not reproductions of its drawings.
Also See:
