AI answers and retrieval-augmented generation: patents and papers
AI answer patents and papers describe how search systems combine retrieval with large language models (LLMs): fetching relevant documents, generating a summary grounded in them, and citing sources. Five Google patent documents (including one pending application) and ten research papers are listed. They explain why retrievable, citable, well-structured pages are the entry ticket to AI answers.
What does this theme cover?
Retrieval-augmented generation (RAG) is a method where a language model first retrieves relevant documents and then generates an answer using them. Google says SEO best practices “remain relevant” for its AI features, with no special schema or files required.[1][2] Officially documented The documents below show why: an AI answer can only draw on pages its retrieval step finds. For the full picture see from SEO to AEO, entities in AI search, and AEO.wiki’s how answer engines work, Google AI Overviews and AI Mode and research pages.
Which patents describe generative search and information gain?
Generative summaries for search results
What it describes: Selectively using a large language model to generate a natural-language summary in response to a query, processing additional content (such as search result documents) along with the query to reduce inaccuracies and over- or under-specified summaries.
Why it matters for SEO: Often discussed as a patent behind generative search summaries such as AI Overviews; Google has not said so. Its design grounds the summary in retrieved documents, so being among the retrieved documents is what counts. Practitioner practice
Generative summaries for search results
What it describes: A later patent in the same family, with the same title and abstract.
Why it matters for SEO: Shows Google continued to pursue claims on generative summaries in 2024.
Search with stateful chat
What it describes: Augmenting search with stateful chat (a “generative companion”): using a generative model with the query and context to create synthetic queries, select search result documents, and keep state across turns.
Why it matters for SEO: This is a published application, not a granted patent. The idea of synthetic follow-up queries resembles the query fan-out AEO.wiki explains (see its fan-out planner): cover the related questions, not just the head term. Practitioner practice
Retrieval-augmented language model pre-training and fine-tuning
What it describes: Pre-training and fine-tuning a language model together with a neural knowledge retriever that learns to fetch relevant documents from a corpus.
Why it matters for SEO: Google’s patent on the approach published as REALM (below). Retrieval quality shapes answer quality.
Contextual estimation of link information gain
What it describes: An “information gain” score: how much additional information a document contains beyond documents the user has already seen, estimated with a machine learning model and used to decide what to present.
Why it matters for SEO: Rewriting what already ranks adds little in this design. Original data, experience and new angles add information. This patent’s family has a 2018 priority date. Practitioner practice
Which research papers explain RAG and AI answers?
REALM: Retrieval-Augmented Language Model Pre-Training
What it describes: A Google model that learns to retrieve Wikipedia passages during pre-training and uses them to answer questions.
Why it matters for SEO: An early blueprint for grounding language models in retrieved documents.
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
What it describes: The paper that named RAG: combining a dense retriever with a generator so answers draw on retrieved text.
Why it matters for SEO: The core pattern behind most AI answer engines: no retrieval, no citation.
WebGPT: Browser-assisted question-answering with human feedback
What it describes: A model that searches and browses the web, collects references and writes answers citing them.
Why it matters for SEO: An early demonstration of an LLM choosing which web pages to cite.
Rethinking Search: Making Domain Experts out of Dilettantes
What it describes: A Google Research proposal for model-based search that answers directly, stressing the need for authoritative, attributed answers.
Why it matters for SEO: Frames why attribution and source authority matter in AI search.
Transformer Memory as a Differentiable Search Index
What it describes: Google research that stores a corpus in model parameters and retrieves document IDs directly from a Transformer.
Why it matters for SEO: A research direction in which the index itself becomes a model.
Lost in the Middle: How Language Models Use Long Contexts
What it describes: Evidence that language models use information at the start or end of a long context better than information in the middle.
Why it matters for SEO: Supports putting the answer first. Independent study
Enabling Large Language Models to Generate Text with Citations
What it describes: ALCE, a benchmark for measuring whether LLM answers cite sources that actually support their claims.
Why it matters for SEO: Citable pages state specific, checkable claims that a model can quote as support.
Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection
What it describes: A model that decides when to retrieve and critiques whether retrieved passages support its output.
Why it matters for SEO: Passages that plainly support a claim are more useful to models that check support.
GEO: Generative Engine Optimization
What it describes: A study of content changes (such as adding citations, quotations and statistics) that increased visibility in generative engine answers on a test benchmark.
Why it matters for SEO: The first widely cited academic study of optimizing for AI answers. Results come from a research benchmark, not a production engine. See AEO.wiki’s AEO vs SEO, GEO and LLMO. Independent study
Retrieval-Augmented Generation for Large Language Models: A Survey
What it describes: An overview of RAG methods, from simple retrieve-then-read pipelines to modular designs.
Why it matters for SEO: A good single reference for how retrieval, chunking and generation fit together.
What should you do with this?
- Make sure pages are crawlable, indexable and fast; retrieval comes first. See SEO fundamentals.
- Put the answer first and keep passages self-contained and specific.
- Add original information: data, experience, examples.
- Build a clear entity so models can connect your content to you. See entities in AI search.
Frequently asked questions
Is there a patent for AI Overviews?
Google has not named one. US 11,769,017 B1, 'Generative summaries for search results', describes LLM summaries grounded in search results and is often discussed in that context.
What is information gain in SEO?
It refers to US 12,013,887 B2, which describes scoring how much new information a document adds beyond what a user has already seen. Google has not confirmed using it.
Do I need special markup for AI answers?
Google says no special schema or files are required for its AI features; standard SEO best practices apply.
See also
References
Pages accessed September 29, 2026 unless a date is given. See all sources and our editorial policy.
- ^ "Optimizing your website for generative AI features on Google Search". Google Search Central. Added May 15, 2026; last updated July 10, 2026.
- ^ "AI features and your website". Google Search Central. Last updated December 10, 2025.