Google patent: Feedback in dialog Q&A
Voice assistants and dialog-based search surfaces answer millions of questions every day – from fact lookups to complex research. What matters is not only the first answer, but also how a system recognizes whether users are satisfied or need a better solution. A Google patent granted in 2022 describes exactly this mechanism: question-answering in human-to-computer dialog systems with evaluation of user feedback on previously delivered answers.
Background: Google's dialog patents and the new approach
Google has invested for years in patents around human-computer dialogs. Earlier documents covered general dialog architectures and completing assistant conversations. The patent "Providing answers to voice queries using user feedback" (US 11,289,096) goes further: it describes how inputs after an answered question are classified as feedback and used to improve future answers.
Core idea: after a dialog system outputs an answer, every further user input is checked to see whether it is feedback on that exact answer. If feedback is detected, the system classifies it as positive or negative. With negative feedback, an alternative solution can be provided – without the user having to ask the original question again.
How feedback is detected and scored
The method starts with receiving a voice input after an answer has already been delivered. The system checks several signals at once: does the input fall within a defined time window after the answer? Is it received after the solution was output? Does it resemble the original question? Or is an action associated with it – such as dialing a phone number or sending an email linked to the answer?
If the input is classified as feedback, a predefined feedback score is assigned. This score adjusts a confidence value describing the strength of a question-answer combination. If the score falls below a threshold, feedback is negative and the confidence value drops. If it is above, feedback is positive and the score rises.
Examples from the patent: gratitude versus contradiction
The patent illustrates positive and negative cases with concrete dialogs. If a user asks "Who invented the telephone?" and receives "Alexander Graham Bell," a follow-up utterance like "Thanks" can be rated as positive feedback. The system stores that this answer fits the question well.
In the negative example, the question is "What was the highest-grossing romantic comedy of 2003?" The first answer is "Gigli." If the user responds with "That can't be right," the system classifies negative feedback. The confidence score for the question-answer pair drops. As a follow-up, the system can deliver an alternative – for example "How to Lose a Guy in 10 Days may be a better answer" – signaling that it is a correction.
Confidence scores and question-answer repository
Dialog devices store information about the strength of answers relative to specific questions. Each question-answer combination receives a confidence score expressing answer relevance. The patent describes a repository that manages these scores centrally and can adjust them both online during user interaction and offline.
If no entry exists yet for a combination, a default score can be assigned – for example 0.5 on a scale of 0.0 to 1.0. A scoring engine then adjusts the value based on the feedback score. Possible operations include averaging with other user feedback, addition, multiplication, or other mathematical links. Aggregated feedback from multiple devices is also included.
Technical architecture: devices, servers, and search engine
The described environment includes user devices – smartphones, tablets, laptops, or dedicated assistant hardware – plus several servers: a question-answer score repository, a feedback classifier, a feedback score repository, and a search engine server. User devices and servers connect over networks; in some implementations, devices perform parts of server functions locally.
Answer generation engine and search integration
The answer generation engine converts spoken questions via speech-to-text into text and generates search queries from them. Stop words such as "the," "a," or "is" can be removed; prefixes and suffixes are normalized. From the question "What was the highest-grossing romantic comedy of 2003?" the query "highest-grossing romantic comedy 2003" may result.
Search results flow in with confidence values considering relevance, document quality, traffic, and age, among other factors. Alternatively or additionally, the engine uses stored question-answer pairs from the repository – including semantic similarity, Hamming distance, or edit distance to recognize paraphrased questions. The answer is output via speech or display.
Feedback classifier and follow-up answers
The feedback classifier engine assigns incoming feedback to the last question-answer pair – typically within 15 seconds, 30 seconds, or one minute, or independently of time if no other feedback occurred in between. Repeated questions, follow-up topics, or actions such as searches, calls, or emails can further influence the rating.
If the feedback score falls below a threshold, the answer generation engine delivers a follow-up answer – often the second-best option from candidates with lower but still relevant scores or fresh search results. This creates a learning system that dynamically adapts answer quality to user opinions.
Benefits and strategic significance for SEO
According to patent benefits, the system can estimate likely user attitude toward an answer – satisfaction, dissatisfaction, or ambivalence. From this, question-answering capability improves over time: answers are dynamically optimized, and dissatisfaction triggers an immediate alternative. For SEO teams and voice search strategists, this is more than a technical detail.
It shows how Google links dialog-based search with classic web search, maintains confidence scores for answers, and uses implicit user feedback as a ranking signal. Anyone optimizing content for featured snippets, assistants, and AI search surfaces should understand that answer quality is not static – it is continuously reassessed through aggregated feedback, search results, and semantic question mappings.
Patent US 11,289,096 was granted on March 29, 2022, filed on November 15, 2019, and lists inventors Gabriel Taubman, Andrew W. Hogue, and John J. Lee. It joins Google's series on human-to-computer dialog, voice queries, and automated assistants and again offers insight into the technical direction of Google search beyond classic SERP clicks.
