Microsoft Patent: How AI Search Compresses Your Content

Estimated reading time: 12 min

I’ve been doing SEO since before Google had a name. And let me tell you: the patent Microsoft published on August 13, 2026, is the clearest window yet into how generative search actually handles your content.

Here’s what actually happened. A patent application (US20260236548A1) from Microsoft Technology Licensing describes a system that compresses web pages before feeding them to a language model. No hype. Just a detailed look at the selection logic that determines whether your content ever reaches the AI that answers the user’s question.

Let me show you the data. The document outlines how a generative search engine retrieves pages, reduces them to their essentials, and passes a compressed version to the model. For publishers, this is a roadmap of what gets kept and what gets discarded.

The Technical Problem: Too Many Tokens, Too Little Value

The patent starts with a familiar engineering challenge. When you feed a language model more context than it can handle, you get two well-documented failures:

• Lost in the middle: Information buried in the middle of a long prompt gets overlooked.
• Recency bias: Models overweight the last pieces of information in a prompt.

The solution? Compress the grounding data—the retrieved snippets, metadata, and direct answers—before sending it. This ensures everything fits in a single call, saving time and compute.

The patent describes two compression techniques: query-focused summarization and, optionally, conversion to embeddings. But before any of that, the system can evaluate freshness and credibility.

Ian Sorin, who spotted the patent and shared it on LinkedIn on October 26, 2026, framed it as a description of how Microsoft’s Copilot and Bing chatbots work. The patent itself speaks of “generative search engines” in general, citing models like GPT-3.5, GPT-4, and GPT-4o—but not naming Bing or Copilot specifically.

How Sentences Get Selected: The PageRank of Phrases

The most actionable part for content teams is the summarization process. The system cleans the page—stripping HTML, special characters, encoding errors—then splits it into sentences and performs extractive summarization. That means it selects existing sentences rather than rewriting them.

Here’s where it gets interesting. The first method builds a graph where each sentence is a node, and edges between sentences are weighted by cosine similarity. The importance of each sentence is calculated using eigenvector centrality—essentially, a sentence is important if it’s strongly connected to other sentences that are themselves important.

Ian Sorin called this a “PageRank for sentences.” The analogy holds. It’s similar to TextRank or LexRank, though the patent doesn’t cite those algorithms. A synthesis sentence that pulls together arguments developed elsewhere in the article becomes linked to each of them and gains centrality.

One nuance: the links are weighted by similarity, so it’s not a simple count.

Keywords Aren’t Dead—They’re Just Part of the Equation

The patent describes two other selection methods:

• Clustering: Sentences are grouped by topic based on semantic proximity to the query, then a representative sentence is extracted from each cluster to cover different aspects of the subject.
• Frequency-based approach: The system prioritizes sentences containing frequent or relevant terms relative to the query.

That second point should give pause to anyone claiming keywords no longer matter. The patent explicitly states that initial retrieval is based on keywords extracted from the query. The semantic analysis is layered on top—not a replacement.

A final method, called discourse-based, analyzes the rhetorical structure of the text to identify sentences that carry the main message and the arguments supporting it. This is likely what Ian Sorin referenced when he mentioned priority given to the main argument and its supporting evidence, like numbers.

Introductions and Conclusions Carry More Weight

Once sentences are selected, the summary is refined. The system applies positional weighting: sentences in key positions, like introductions and conclusions, are favored because they’re more likely to contain essential information. It then removes redundancies and checks that the summary remains readable, varied, and relevant to the query.

I’ve seen this play out before. The old advice to “front-load your key points” hasn’t changed—it’s just been encoded into a new layer of the stack.

Old Content Gets Demoted—Unless It’s Evergreen

Freshness evaluation relies on timestamps: creation date or last modification date. Content is ranked from oldest to newest, and older items receive lower relevance scores.

Ian Sorin noted that these dates are extracted from server headers, HTML metadata, search indexes, or the text itself. That detail, however, doesn’t appear to be in the patent description.

Several safeguards are built in:

• Exemption for foundational knowledge: Content with lasting value—historical references, core knowledge—can escape age-related demotion.
• Dynamic expiration: Content exceeding an age threshold, or ranked beyond a certain position (the patent gives the example of the top 1,000 items), is removed.
• Domain-specific thresholds: A breaking news article has a shorter shelf life than an academic paper or technical manual.
• Update detection: The system monitors changes to articles or datasets to use the most recent version.

In other words, updating content isn’t just a signal for traditional search engines. In this system, it directly determines the content’s place in the data passed to the model.

Reputation, Neutrality, and Corroboration Filter Sources

Credibility evaluation starts with a reputation score that aggregates three weighted criteria:

• Historical accuracy
• Domain authority
• Peer validation

The patent doesn’t specify how these are measured—Ian Sorin also flagged the peer validation element. Notably, the text focuses on source and data reliability, without explicitly mentioning authors.

The system then cross-references information across multiple independent sources to reduce error and bias. It also uses bias detection models that analyze linguistic patterns and framing to identify subjective or slanted content, then takes corrective measures to keep the final answer neutral.

Finally, source diversity. The system aims to avoid relying on a single perspective. It flags information supported by fewer than a threshold number of sources and attempts to find complementary ones. The text doesn’t explicitly say that such information is penalized if no additional sources are found.

If the volume is still too large after these steps, the patent describes removing the least relevant, least fresh, or least credible data until it fits under the token limit—possibly by alternating between these criteria. It also mentions reusing a previously compressed version stored for a similar query.

The Limits of Patents—and Why They Still Matter

Let’s be clear: patents are not recipes. The text describes many variations introduced with “in some implementations,” and its publication doesn’t prove this system is running behind Bing or Copilot today. But it does illuminate the selection logic a generative engine can apply to the pages it retrieves.

Nobody talks about this part: the patent leaves an open question about exclusive information. A scoop or proprietary data—by definition, reported by few sources at publication—triggers the search for complementary sources. The text doesn’t say what happens when none are found.

But the practical takeaways for content teams are clear:

• Synthesis sentences: A standalone sentence that summarizes arguments developed elsewhere in the article is more likely to be judged central and retained.
• Placement: Key information benefits from appearing in introductions and conclusions of sections.
• Dating: Visible, up-to-date publication and modification dates influence freshness evaluation, especially for news.
• Tone: A factual, hype-free style reduces the risk of being flagged as biased content.
• Verifiability: Numeric data and claims that can be verified elsewhere carry more weight than isolated information.

The Playbook Changed. Again.

I’ve watched every major shift in search since 1997. Each time, the fundamentals of good writing—clarity, structure, verifiability—held up. The difference now is that the machine reading your content is looking for different signals than the ranking algorithm of 2010.

This isn’t a take—it’s a pattern. The patent tells us that generative search cares about centrality, freshness, reputation, and corroboration. Those are the levers. Pull them deliberately.

Slow down. Think. The future of search isn’t about tricking the model—it’s about writing in a way that survives compression and still carries meaning. That’s a game worth playing.