On 13 August 2026, Microsoft Technology Licensing LLC published patent application US20260236548A1, which describes a system for compressing data from search results before transmitting it to a generative AI model. To summarize a page, this system retains the sentences most connected to the rest of the text and to the query. However, the document does not prove that Bing or Copilot apply this exact process.
- Microsoft published a patent application describing how search results are compressed before reaching a generative AI model.
- The system keeps the sentences most connected to the rest of the text and to the query.
- The patent does not prove that Bing or Copilot use this process.
- The text does not say what happens to exclusive information when no complementary source exists.
Compressing Sources Before Calling the Model
The document is titled ‘Generating and utilizing compressed grounding data for search engines that utilize generative artificial intelligence models’. It claims priority from an Indian provisional application filed on 13 February 2025. It concerns ‘grounding data’, i.e., content retrieved from search results (links, snippets, metadata, direct answers) and provided to a language model to base its response on verifiable facts. This mechanism is known as RAG (retrieval-augmented generation).
The patent starts from a technical problem. When this data exceeds the token limit (the units of text a model can process at once), multiple calls are required, costing time and resources. Conversely, providing too much at once leads to generic or incoherent answers. The text cites two known flaws of large language models: the ‘lost in the middle’ phenomenon, where information placed in the middle of a long prompt is poorly considered, and recency bias, which gives too much weight to elements at the end of the prompt.
The proposed solution involves compressing this data before sending it, to transmit everything in a single call. The system can produce a query-focused summary, then convert it into embeddings, i.e., vector representations of a text’s meaning. It can also assess the freshness and credibility of content beforehand.
In a LinkedIn post, Ian Sorin presents this patent as a description of how Copilot and Bing chatbots work. The text itself, however, refers to ‘generative search engines’ in general and cites models like GPT-3.5, GPT-4, or GPT-4o as examples, without naming Bing or Copilot.
Central Sentences, Freshness, and Credibility
For summarizing pages, the system first cleans the content by removing HTML tags, special characters, and encoding errors. The text is then split into sentences and summarized extractively by selecting existing sentences rather than rewriting content. The main method represents each sentence as a node in a graph, connected to others by their cosine similarity, a measure of semantic proximity. A sentence is deemed important if it is strongly connected to other sentences that are themselves important, which Ian Sorin calls a ‘sentence pagerank’. The patent also describes clustering by topic and selection based on terms frequent relative to the query. Sentences located in introductions or conclusions are then favored.
Freshness is assessed from content timestamps, i.e., their creation or last modification date. Older content receives lower relevance, with exceptions for foundational knowledge, variable lifespan thresholds depending on the domain, and detection of updates. Credibility relies on a reputation score combining historical accuracy, domain authority, and peer validation. The system also cross-references information across multiple independent sources and checks for diversity of perspectives. The patent does not specify how these criteria are measured.
Guidelines for Writers, No Proven Application
The document describes many variants, introduced by the phrase ‘in some implementations’. Its publication therefore does not prove that this system currently operates behind Bing or Copilot. However, it sheds light on the selection logic a generative engine can apply to the pages it retrieves. For editorial teams, Ian Sorin draws several lessons: self-contained summary sentences, placement of essential information in introductions and conclusions, visible publication and update dates, a factual tone, and verifiable data.
The patent, however, leaves one question open: that of exclusive information. A scoop or proprietary data, covered by few sources at the time of its publication, triggers a search for complementary sources. The text does not specify what happens to this information when the system finds none.
