Google never documented how AI Overviews use entities — so the industry cited Microsoft's patents instead

Executive summary

Google has a public, verifiable patent record on how it ranks entities, builds knowledge panels, and disambiguates queries — four granted patents, all confirmed by direct inspection of the assignee field. What Google does not have, anywhere in its public documentation, is a description of how (or whether) AI Overviews and AI Mode use its Knowledge Graph as a grounding mechanism. That gap has been filled by the SEO industry with citations to patents that are not Google's at all. The most consequential of these is the patent behind GraphRAG, Microsoft Research's framework for extracting entity graphs from documents and using them to guide retrieval-augmented generation — filed by Microsoft Technology Licensing LLC, not Google LLC (Confidence: High — primary patent record confirmed by direct fetch). Three other patents commonly presented in SEO content as Google entity-resolution prior art belong to Microsoft and IBM respectively.

This matters because Google's own consumer-facing Knowledge Graph API tells developers not to build anything critical on top of it: "This API is not suitable for use as a production-critical service. Your product should not form a critical dependence on this API" (Confidence: High — verbatim, directly fetched from developers.google.com). Meanwhile the growth numbers are real and large — Sundar Pichai told I/O 2026 that "AI Overviews now has over 2.5 billion monthly active users" and that AI Mode "has already surpassed 1 billion monthly active users" in its first year (Confidence: High — verbatim keynote quote) — but that same keynote post makes zero mention of Knowledge Graph or entity mechanics (Confidence: High — confirmed by full-text search of the fetched post).

The honest state of the evidence: Google has real, narrow patents on entity ranking and disambiguation dating back to 2012–2021, a Knowledge Graph that underwent a measurable contraction in mid-2025 according to one analyst's independent tracking (Confidence: Medium — single-source, not independently replicated), and a stated preference for not depending on its public entity API in production. What it does not have — at least not in anything publicly findable — is a documented, current explanation of how AI Overviews decide which entities to trust.

A timeline / comparison table

Part 1 — What Google has actually filed, in Google's own words

Strip away the secondary commentary and four things remain on the public record, all confirmed by directly pulling the patent text and assignee metadata.

US10,235,423 B2 describes weighting knowledge-graph entity metrics — relatedness, notability, "contribution," prizes and awards — differently depending on entity type, so that, for example, award-winning is weighted more heavily for a Film-domain entity than for other types, producing a composite ranking score (Confidence: High — primary patent record). US9,268,820 B2 (published as US20130110825A1) describes aggregating content from multiple sources into a knowledge panel, ranking the facts shown by how often users search for them, and generating disambiguation panels when a query maps to more than one entity — the patent's own example is the ambiguous query "Phoenix" (Confidence: High — primary patent record, elements confirmed verbatim). US9,336,211 B1 scores candidate query-to-entity associations using ambiguity, popularity, and result-quality signals, illustrated with the query "sting" resolving to either the musician or the wrestler (Confidence: High). US11,210,289 B2 segments a query into terms, matches them against entity metadata, and filters out low-confidence or ambiguous matches — its own worked example is disambiguating "action movie with tom cruise," not the "Lincoln" (president vs. car vs. city) example that has circulated in secondary SEO writeups (Confidence: High — patent confirmed directly; the Lincoln example does not appear in the patent text and should not be attributed to it).

Separately, Google's commercial Cloud offering — Enterprise Knowledge Graph — includes an "Entity Reconciliation API" that converts relational BigQuery data into RDF triples, clusters entities using fuzzy matching across text, relationships, types and attributes, assigns each a stable Machine ID, and can optionally link out to Google's global Knowledge Graph (Confidence: High — directly fetched from Google Cloud documentation). This is a real, documented reconciliation pipeline. It is enterprise tooling for customers building their own knowledge graphs on GCP — not a description of how Search or AI Overviews resolve entities at query time.

And the consumer-facing Knowledge Graph Search API — the one product most SEO practitioners actually interact with — carries an explicit disclaimer that it is not production-grade, and a second disclaimer that it "returns only individual matching entities, rather than graphs of interconnected entities," recommending third-party Wikidata dumps for anyone who actually needs graph structure (Confidence: High — verbatim). That is a striking admission for a product whose name is literally "Knowledge Graph API": Google is telling developers, in its own docs, that the graph part isn't really there for public consumption.

Part 2 — The patents SEO content wrongly calls Google's

Four patents that keep surfacing in SEO commentary about "Google's entity resolution" are not Google's.

US9,443,021 B2, "Entity based search and resolution," describes merging equivalent entities discovered across separate vertical data sources — the patent's own examples are IMDB, Netflix, and Rovi — into a single aggregate-ranked result. It is assigned to Microsoft Technology Licensing LLC (Confidence: High — confirmed against both the pre-grant publication and the granted patent, same family). US9,251,473 B2, "Identifying salient items in documents," derives training labels for salience scoring automatically from search-log click-through ratios rather than manual annotation — a genuinely clever technique — and is also Microsoft's (Confidence: High). US11,379,754 B2, covering entity resolution across disparate data sources using tokenization, normalization, and supplementary web search to consolidate records representing the same real-world entity, is Microsoft's as well (Confidence: High). And US12,159,224 B2, "Entity disambiguation using graph neural networks," which describes a heterogeneous graph neural network with metapath-guided attention and a Siamese-network matcher, is assigned to IBM, with named inventors Chuan Lei, Alina Vretinaris, Vasilis Efthymiou, and Fatma Ozcan (Confidence: High).

The most consequential misattribution is GraphRAG. US2025/0131289 A1, "Knowledge Graph Extraction," is the patent filing behind Microsoft Research's GraphRAG framework — the technique of using an LLM to extract entities, relationships and claims from documents into a graph, running Leiden-algorithm community detection over it, and then answering queries with local, community-level, or global retrieval over that structure, including LLM-based merging of duplicate entity variants. It is assigned to Microsoft Technology Licensing LLC, and its inventor list overlaps directly with the GraphRAG paper's authorship — Darren Edge, Jonathan Larson, David Koleczek, Thu Ha Trinh, Newman Cheng among them (Confidence: High — confirmed against the patent record and the paper's inventor list). The paper itself, "From Local to Global: A Graph RAG Approach to Query-Focused Summarization," was posted to arXiv on April 24, 2024, revised in February 2025, and is hosted on Microsoft's own official research publications page — confirming Microsoft Research authorship independently of the patent filing (Confidence: High — two independent primary sources agree).

GraphRAG is, by a wide margin, the most technically detailed public blueprint that exists for how an entity graph might ground an LLM's answers. That is precisely why it keeps getting pattern-matched onto Google's AI Overviews in SEO commentary — it's the only fully-specified mechanism anyone has published. But specificity is not evidence of adoption. Nothing in Google's own I/O 2026 materials, patent filings, or developer documentation describes an equivalent pipeline for AI Overviews, and treating Microsoft's architecture as a stand-in for Google's is an inference, not a citation.

Design note for the article layout: consider a simple two-column visual here — left column "Filed by Google LLC" listing the four confirmed Google patents with grant dates; right column "Filed by Microsoft / IBM" listing the four misattributed ones with their actual assignees in bold. The visual contrast (same subject matter, different assignee) does more work than another paragraph of prose.

Part 3 — Scale is documented; the mechanism isn't

What Google will say, loudly, is how many people use the AI-powered surfaces. At I/O 2026, Pichai stated: "AI Overviews now has over 2.5 billion monthly active users. And AI Mode has been a revelation, our biggest upgrade to Search ever. People love it, and in just a year, it's already surpassed 1 billion monthly active users." (Confidence: High — verbatim.) What that post does not contain, anywhere, is a reference to Knowledge Graph, entity resolution, or how answers are grounded.

Google's own research has, in the past, published concrete numbers about its knowledge base: the 2014 Knowledge Vault paper reported 1.6 billion extracted facts, of which 271 million met a confidence threshold of 90% or higher (Confidence: Medium — the figures were independently corroborated via secondary summaries of the paper, but direct text extraction from the original KDD PDF failed during verification; treat as reported, not re-confirmed from primary text). That is more than a decade old and predates any AI Overview product.

The most recent independent data point on the graph's current state comes from a single analyst, Jason Barnard of Kalicube, who reported a two-stage "Clarity Cleanup" in mid-2025: over 3 billion entities removed on June 13 and June 20 (a 6.26% reduction), followed by an August 11 cleanup targeting corporate and brand entities. His tracking shows event entities dropping 76.91% in count (average lifespan falling from 839 days pre-COVID to 124 days post-March 2020), the generic "Thing" category shrinking by roughly 8 billion entities (15.27%) as more entities received specific types, and confidence in unambiguous person-typing rising from 70.16% to 76.78% (Confidence: Medium — every figure matched the source article verbatim, but this is one analyst's proprietary tracking dataset, not a Google disclosure, and no second analyst has independently replicated it with separate methodology). The directional story — fewer, more confidently typed entities — is plausible and consistent with a system trying to reduce noise before LLM consumption. It is not proof of what AI Overviews actually query.

The claim that gets repeated most often in SEO content — that a large majority of AI Overview citations go to pages already ranking in the top 10 organically — turns out to be the wrong kind of number to build a strategy on. It is volatile, and now measured: Ahrefs' March 2026 study of roughly 4 million AI Overview citation URLs across 863,000 keywords found the top-10-organic share of citations fell from approximately 76% in July 2025 to 37.9% in March 2026 (37.1% counting organic blue links only), attributing the fall to query fan-out — Google generating sub-queries whose own separate result sets, increasingly populated by non-traditional sources, supply citations independent of the primary query's top-10 ranking (Confidence: High — primary Ahrefs source, precise dual figure; see also this site's companion piece on the same data). Presenting last year's number as this year's fact would be actively misleading, so it is best understood as evidence that the citation-rank relationship is currently in flux, not as a stable statistic to plan around.

One serious, if unconfirmed, hypothesis worth taking on its own terms: industry analysis from iPullRank describes AI Mode's "Deep Search" as decomposing a prompt into parallel sub-queries run against the web index plus Google's internal graphs — Knowledge Graph, Shopping Graph, Maps — then synthesizing a single cited answer. It's a coherent theory of the mechanism. It is also explicitly not confirmed by Google, and iPullRank's own phrasing hedges it as something AI Mode would "likely leverage" rather than something documented (Confidence: Low — labeled explicitly as industry inference by its own source, not a documented Google mechanism).

What everyone is missing

The industry is matching patents by keyword, not by assignee. A patent titled "entity resolution" or "entity disambiguation" gets folded into "how Google does entities" content regardless of whose name is on the filing. Four of the most-cited examples in this space belong to Microsoft or IBM. That's not a minor footnote — it means a meaningful slice of published "Google entity resolution" analysis is actually a description of Microsoft's and IBM's research programs, wearing Google's name because the topic overlaps.

GraphRAG's popularity as an SEO framing device is inversely correlated with the evidence that it's what Google runs. The more technically elegant and fully-specified a public architecture is, the more it gets treated as ground truth for a system that has published nothing comparably detailed. GraphRAG deserves attention on its own merits — it's a real, useful pattern for LLM-plus-graph retrieval, documented by its actual creators — but its clarity has made it a stand-in for Google's undocumented internals, which is a category error.

Google's own product documentation quietly undercuts a decade of "get into the Knowledge Graph" advice. If the company that owns the graph is telling developers not to build production dependencies on its public API, and that the API doesn't even expose interconnected graph structure, the idea that a business should chase "Knowledge Graph presence" as a discrete, gameable objective was always resting on a product Google itself won't vouch for.

The one number everyone wanted to hold still — the AI Overview citation/rank correlation — moved, and the move is more informative than the original figure. Ahrefs' own measurement fell from roughly 76% to 37.9% within about eight months, attributed to query fan-out behavior, which tells you the underlying system is not a static ranking-to-citation pipeline (Confidence: High — see Part 3; primary Ahrefs source). Treating any single snapshot of that number as durable guidance is a mistake regardless of which direction it currently points.

Future predictions

• Google will continue to avoid publishing a detailed technical account of how AI Overviews select and weight entities, for the same competitive reasons it has never fully documented core ranking. (Confidence: Medium — consistent with a decade of Google's disclosure pattern, not a stated commitment.)
• Third-party analysts (Barnard/Kalicube-style tracking) will remain the primary public source of Knowledge Graph size and composition data, and single-source claims from this space will keep circulating as if they were Google disclosures. (Confidence: Medium — extrapolated from the current absence of a second independent tracker.)
• SEO industry content will continue misattributing Microsoft's and IBM's entity-resolution patents to Google, because the pattern-matching (title keyword → assumed Google authorship) has not been corrected at scale. (Confidence: Medium — based on the volume of existing misattributed content found during this research.)
• The AI Overview citation/top-10-rank correlation will keep moving as Google's query fan-out and citation logic evolves, meaning any fixed percentage published in 2026 should be assumed stale within a year. (Confidence: Medium — based on Ahrefs' documented swing from ~76% to ~37.9% in eight months.)
• Google Cloud's Enterprise Knowledge Graph / Entity Reconciliation API tooling is more likely to see incremental documentation updates (it's an active commercial product) than the consumer Knowledge Graph Search API, which reads as a legacy, deprioritized surface. (Confidence: Low — inference from current documentation emphasis, not a stated roadmap.)
• Expect continued growth disclosures (user counts, adoption) from Google's AI Overviews/AI Mode teams without matching technical disclosure on entity or retrieval mechanics, mirroring the I/O 2026 pattern. (Confidence: Medium — consistent with the I/O 2026 keynote's own framing.)

Practical takeaways

1. Stop citing "Google's entity resolution patent" without checking the assignee field on Google Patents first — verify it says Google LLC, not Microsoft Technology Licensing or IBM, before it goes in a deck or a client report.
2. Treat GraphRAG as a useful mental model for graph-grounded retrieval in general, not as documentation of Google's specific pipeline — if you're building entity data architecture, our Entity & Knowledge Architecture approach starts from what's actually verifiable about a given engine, not borrowed blueprints from adjacent vendors.
3. Don't build a strategy around a single snapshot of "AI Overview citations correlate with top-10 rank" — that number moved by roughly half in about a year; monitor it on a rolling basis rather than treating any published figure as fixed, which is the kind of tracking our LLM Visibility Monitoring work is built around.
4. If you're relying on Google's consumer Knowledge Graph API for anything customer-facing, read its own disclaimer again — it explicitly says not to build production dependencies on it, and to use Wikidata for anything requiring real graph structure.
5. Prioritize the four confirmed Google mechanisms — entity-metric ranking, knowledge panel aggregation and disambiguation, query-entity association, and query interpretation via entity metadata — as the actual, evidenced targets for structured data and disambiguation work, which is the core of our Technical Foundations engagements.
6. When a client asks "should we go after Knowledge Graph presence," reframe the question: fewer, higher-confidence entities appear to be the current direction (per one analyst's tracking), so breadth-first entity tactics are a weaker bet than depth and disambiguation clarity for entities you actually own.