Google patent: query answers with constraints

In May 2022, Google was granted the patent "Generating Query Answers" (US 11,321,331). It describes how search systems answer factual user questions—such as via voice assistant—not only as link lists but as grammatically correct sentences. Central to this are constraints: rules that define which data fits answer templates. For SEO professionals, the patent offers deep insight into entity attributes, semantic triples, and the bridge between the Knowledge Graph and natural language.

From database facts to natural answers

Classic search results often deliver document lists. For voice search or dialog systems, Google expects a direct, spoken answer. The patent explains the path: facts from a graph-based data store are converted into sentences in the user's language. Instead of random triples, an answer engine selects information that precisely answers the query—using constraints as quality filters.

Constraints as quality filters for query answers

Each field in an answer template can carry multiple constraints. Typical types include type constraints (date, number, entity name), temporal constraints (past or future), gender constraints, relationship constraints (e.g., spouse), singular/plural rules, units of measure, and determinants such as "the." The system selects the template with the most satisfied constraint fields—the most data-rich, grammatically fitting formulation.

• Type constraints for date, number, and entity fields
• Temporal constraints for past and future
• Relationship constraints for marriage, origin, or alma mater
• Singular/plural and unit-of-measure constraints

Entity-attribute-value triples in the Knowledge Graph

Facts appear as triples: subject (entity), predicate (attribute), and object (value). Example: Woody Allen – acted in – Annie Hall. Billions of such triples feed index clusters and query resolvers. For complex questions with multiple attributes—"Where is Woody Allen's hometown and alma mater?"—the answer engine combines several phrase templates in meta-templates bound to entity types such as "person."

Patent facts at a glance

Template selection and the answer engine

For each attribute in a query—"age," "marriages," "hometown"—multiple candidate templates exist. For "How old is Woody Allen?" variants are available: with birth date and age, age only, or birth and death dates. The answer engine parses the query, extracts attributes, loads matching templates in the user's language, and fills fields with facts. Heuristics check gender, tense, and whether the number of answer triples matches template fields.

Voice search and dialog systems

Users can ask questions as spoken sentences. Client devices convert audio via speech recognition to text; the search system formats the query structurally and sends it to index clusters. The generated answer can return as text on a SERP, as transcription for text-to-speech, or directly as an audio signal. The patent thus links classic graph search with conversational search—a building block for featured snippets, knowledge panels, and AI overviews.

Meta-templates for complex queries

Simple questions with one attribute produce a sentence from one template. Complex queries use meta-templates: frame structures with fields for multiple phrase templates. For marriage questions about Woody Allen, the system selects the most data-rich phrase template per triple—such as "has gotten married to since" or "was married to from to"—and inserts phrases into a person meta-template. Result: "Woody Allen has gotten married to Soon-Yi Previn since 1997 and was previously married to Louise Lasser from 1966 to 1970."

Link to semantic SEO and related patents

The article references GRIP, QA-by-Dossier-with-Constraints, and the Google Knowledge Vault as research context. Closely related patents cover Knowledge Graph reconciliation and entity extraction. Those optimizing content for entity-based search should focus on structured facts, consistent attributes, and complete triple coverage—missing birth dates or incomplete relationship data lead to weaker templates and less visibility in direct answers.

Practice for SEO teams

The patent underscores: Google prefers not arbitrary facts but the best-fitting, constraint-compliant formulation. Schema.org markup, consistent entity names, precise dates, and relational facts (marriages, locations, education) support selection of data-rich templates. Internationalization is built in: templates load in the language of the original query. For multilingual sites, facts must be correctly linked per language, not only translated.

Data graph search and index clusters

The patent outlines a search engine for data stores with indexing system, search system, and index cluster. The query resolver accesses the index and delivers results that serve the answer engine as a fact base. Triples link entities via edges; the graph can be updated in real time. For SEO, Google separates data storage from sentence generation—clean, current facts in structured form are the prerequisite for visibility in direct answers and voice results.

Inventors Engin Cinar Sahin, Vinicius J. Fortuna, and Emma S. Persky, with assignee Google LLC, anchor the invention at the core of factual answer generation. Teams that read patents as roadmaps see: visibility increasingly arises where entity knowledge is machine-readable, complete, and constraint-ready—whether the query is typed or spoken.