Discovery Science · established evidence
The Substrate Shift: From Ranked Links to Entity-Based Search, 2012 to 2026
For most of its history, search did one job: it matched the words in a query against the words in a library of documents and ranked the documents that matched best. Entity-based search does something structurally different. It resolves the query to a thing, a person, a place, a business, a concept, and retrieves what it holds about that thing, increasingly as a passage rather than a whole page. This is not a new interface bolted onto an old system; it is a change in the underlying unit of work, and it has arrived in stages since 2012. Google's Knowledge Graph reframed the index around entities. Retrieval-augmented generation, introduced in 2020, made a passage the retrievable unit that answer engines now assemble. A 2024 academic result gave the practice of optimizing for those systems a name and a first measurement. Reading that shift precisely, rather than reacting to the marketing around it, is what tells you what optimizing for search now actually means.
A ranked list of links was always a proxy
The ten blue links were a user interface, not the mechanism underneath. To produce that list, an engine matched query terms against an inverted index of documents and scored each document for relevance and authority. The document was the atomic unit: the thing retrieved, ranked, and returned. Everything a practitioner called "SEO" was, at bottom, an effort to make a document score higher in that matching process.
That design carried an unspoken assumption, that the best way to answer a question is to hand back the documents most likely to contain the answer and let the reader do the final work of extraction. It was a proxy. The reader wanted a fact, a business, a decision; the engine returned a shelf of pages and trusted the reader to find it. The substrate shift is the story of that proxy being progressively removed, one layer at a time, until the engine retrieves the thing itself.
Things, not strings: the 2012 turn
The first public marker of the shift is precise and datable. In May 2012 Google announced the Knowledge Graph under an explicit thesis: index things, not strings. At launch it described roughly 500 million entities and about 3.5 billion facts about the relationships between them, assembled in part from Freebase, Wikipedia, and the CIA World Factbook.
The claim embedded in that announcement is the one that matters here. A "string" is a sequence of characters to be matched. An "entity" is a modelled real-world thing with a stable identity and a web of attributes and relationships. Moving the unit of representation from the string to the entity means the engine no longer has to find a document that happens to contain your words; it can retrieve a structured record about the thing your words refer to. Structured entity representation via knowledge graphs predates generative AI by more than a decade, and it remains the substrate the newer answer systems reason over.
What is a knowledge graph?
A knowledge graph is a database of entities and the typed relationships between them: nodes for things, edges for facts. Rather than storing "the document at this URL contains these words," it stores "this clinic is located in this city, is led by this practitioner, and offers these services." That structure is what lets an engine answer a question about a thing without first ranking a page about the thing, which is the capability every downstream answer engine inherits.
Retrieval-augmented generation and the passage as the unit
The second marker is the architecture that sits under almost every generative answer engine today. In 2020, Lewis and colleagues introduced retrieval-augmented generation, or RAG: a pretrained language model coupled with a separate, non-parametric index that is queried at inference time and whose retrieved text is fed to the model as it generates. Their reported result was that RAG models produced more specific, diverse, and factual language than a comparable model relying on its parameters alone, and could update their knowledge without being retrained.
The consequence for our subject is subtle but decisive. In a RAG system the retrievable unit is not the ranked document; it is the passage, a chunk of text pulled from an index because its meaning is close to the query, then handed to a model that composes an answer from several such chunks. The visible answer is assembled, not selected. A page can contribute a passage to an answer without being the top-ranked page, and a top-ranked page can contribute nothing at all.
What retrieval actually does
Retrieval in this sense is not keyword lookup. The query and the candidate passages are encoded as vectors, and the system fetches the passages whose vectors sit closest to the query's, a similarity in meaning rather than an overlap in words. This is why "entity-based search" and "retrieval" belong in the same discussion: both replace string matching with a representation, an entity record or a semantic vector, that stands in for the underlying thing.
What it means to optimize for a substrate of entities and passages
If the unit of work has moved from the ranked document to the retrieved entity and passage, then "optimizing for search" cannot mean only "rank a page higher." The first controlled evidence for what it does mean arrived in 2024. Aggarwal and colleagues coined the term Generative Engine Optimization and tested it on GEO-bench, a benchmark of roughly ten thousand queries across nine datasets.
Their finding is specific and worth stating exactly. The strongest content-level interventions, adding citations to credible sources, including direct quotations, and replacing vague claims with concrete statistics, produced a relative lift in the range of 30 to 40 percent on a position-adjusted visibility metric, versus unoptimized baselines, in the engines they tested. Citing authoritative sources was consistently the single strongest lever. Note what the object of optimization has become: not a rank in a list, but the probability that a passage from your source is retrieved and cited inside a generated answer.
geo vs seo: the same page, two different jobs
This is the practical heart of the substrate shift. Classic search optimization improves a document's position in a ranked list. Generative and answer-engine optimization improves the odds that a passage is retrieved and a source is cited. The two share tactics, clear structure, credible sourcing, extractable claims, but they act on different mechanisms and are measured differently. A rigorous read treats them as separate jobs on the same page, not as a rename of one job.
The identity problem underneath the entity
A substrate of entities is only as reliable as its ability to decide that two references denote the same thing. That problem is older than any AI answer engine and is not fully solved. A foundational Semantic Web analysis by Halpin and colleagues found that owl:sameAs, the formal mechanism for declaring that two identifiers refer to one real-world entity, is used inconsistently across the Linked Data web, with publishers applying at least four looser, non-equivalent senses of "same" in practice.
The looseness has never been fixed in the underlying standard; it has been worked around with platform-specific heuristics. This resurfaces for an ordinary business as a mundane but costly question: is this listing, review profile, or citation about the same business as that one? Entity consistency, the same name, same location, same practitioner, same claims, expressed the same way across the web, is not tidy bookkeeping. It is how you help imperfect resolution systems conclude that the scattered references to your business are one entity, which is a precondition for being retrieved as that entity at all.
Structured data is a standard, not a trick
The vocabulary businesses use to declare entities to machines, schema.org, is a shared standard, founded jointly by Google, Bing, Yahoo, and Yandex in 2011 to standardize machine-readable markup. That governance is the point: it is durable infrastructure, not a proprietary loophole, which is why it belongs in a technical foundation rather than in a list of tactics that expire.
The limits of the substrate shift
The established spine of this story is strong: the Knowledge Graph in 2012, RAG in 2020, its formal taxonomy in 2023, and the first controlled GEO measurement in 2024 are all documented, primary, and mutually consistent. What the substrate shift does not license is confident prediction about how any single engine will weight or cite on a given day.
Two cautions belong here. First, some of the most-cited practitioner claims about how today's engines decouple ranking from citation, that a page far outside the top organic results can still be cited, rest on secondary analysis of platform disclosures rather than a single verifiable primary paper, and should be read as emerging rather than settled. Second, a large 2025 controlled study reported that AI answer engines skew toward earned, third-party media over brand-owned pages more sharply than classic Google does, a striking result, but a single study not yet independently replicated. The disciplined posture is the same one the evidence itself models: state what is established, tier what is emerging, and measure your own position rather than assert it. Classic rank and AI citation are now different enough that they are worth measuring as separate things, which is exactly why a blended rank number no longer describes where a business actually stands.
The evidence
Key findings, with their sources
-
Google's Knowledge Graph launched in May 2012 with roughly 500 million entities and about 3.5 billion facts, under the explicit thesis of indexing "things, not strings."
established Singhal, A., "Introducing the Knowledge Graph: things, not strings," Official Google Blog, May 16, 2012.
-
Retrieval-augmented generation coupled a pretrained model with a queried-at-inference index and produced more specific, diverse, and factual language than parametric-only generation, updatable without retraining.
established Lewis, P. et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks," arXiv:2005.11401, NeurIPS 2020.
-
The RAG literature is organized into a Naive / Advanced / Modular taxonomy, now the shared vocabulary for how AI Overviews, ChatGPT Search, Perplexity, and Gemini construct answers.
established Gao, Y. et al., "Retrieval-Augmented Generation for Large Language Models: A Survey," arXiv:2312.10997, 2023-24.
-
Adding citations to credible sources, direct quotations, and specific statistics produced a 30 to 40 percent relative lift on a position-adjusted visibility metric across roughly 10,000 queries in the GEO-bench benchmark; citing authoritative sources was the single strongest lever.
established Aggarwal, P. et al., "GEO: Generative Engine Optimization," arXiv:2311.09735, ACM SIGKDD 2024.
-
The owl:sameAs mechanism for declaring two identifiers denote one real-world entity is used inconsistently across the Linked Data web, with publishers applying at least four looser, non-equivalent senses of "same."
established Halpin, H. et al., "When owl:sameAs Isn't the Same: An Analysis of Identity in Linked Data," ISWC 2010.
-
Schema.org, the vocabulary for declaring machine-readable entities, was founded jointly by Google, Bing, Yahoo, and Yandex in 2011 as a shared standard rather than a proprietary signal.
established Schema.org founding history (Google, Bing, Yahoo, Yandex), 2011.
-
A large 2025 controlled study reported AI answer engines are systematically biased toward earned, third-party media over brand-owned content, more sharply than classic Google.
emerging Chen, M. et al., "Generative Engine Optimization: How to Dominate AI Search," arXiv:2509.08919, 2025.
Calibration
What is proven, what is promising, what is unproven
| Evidence tier | Tactics | What the evidence says |
|---|---|---|
| established | Treat the entity, not the document, as the unit; keep entity representation consistent; structure content so credible, quotable, source-cited passages are extractable. | Google Knowledge Graph 2012; Lewis et al. RAG 2020; Gao et al. RAG survey 2023-24; Aggarwal et al. GEO 2024; Halpin et al. 2010; schema.org 2011. |
| emerging | Weight earned, third-party corroboration alongside owned pages; watch for ranking and citation diverging on the same query. | Chen et al. 2025 (single large study, not yet replicated); secondary analysis of AI Overviews two-stage retrieve-then-cite architecture. |
| contested | Do not treat AEO and GEO as identical, and do not forecast any single engine's weighting or citation behavior as fixed. | AEO-to-GEO lineage is an industry-consensus narrative, not a primary-sourced claim; per-engine weighting figures are practitioner estimates. |
Reference
Glossary
- Entity-based search
- Search that resolves a query to a modelled real-world thing (a person, place, business, concept) and retrieves structured information about it, rather than matching query strings against documents.
- Knowledge graph
- A database of entities and the typed relationships between them, storing facts about things rather than the words contained in pages. Google's launched in 2012.
- Retrieval-augmented generation (RAG)
- An architecture that pairs a language model with a separate index queried at inference time, feeding retrieved text into the model so it can generate a grounded answer. The pattern underneath most generative answer engines.
- Passage
- A chunk of text retrieved from an index because its meaning is close to the query and assembled into a generated answer. In RAG systems the passage, not the ranked page, is the retrievable unit.
- Things, not strings
- Google's 2012 framing for the Knowledge Graph, signaling the move from matching character strings in documents to representing real-world entities and their relationships.
- Entity resolution
- Deciding that two references (two listings, two profiles, two citations) denote the same real-world thing. A known-hard, imperfectly solved problem, which is why consistency across the web matters.
Straight answers
Frequently asked questions
What is entity-based search?
It is search that treats your query as a reference to a thing, a business, a person, a place, a concept, and retrieves structured information about that thing, instead of only matching your words against a library of documents. Google's 2012 Knowledge Graph was the first large public step, and retrieval-augmented generation extended the idea to passages assembled into answers.
What is a knowledge graph?
A knowledge graph stores entities and the typed relationships between them, nodes for things, edges for facts. Google announced its own in 2012 with roughly 500 million entities and about 3.5 billion facts, under the thesis of indexing "things, not strings." It is the structured substrate that lets an engine answer a question about a thing without first ranking a page about it.
What is retrieval-augmented generation?
Retrieval-augmented generation, or RAG, pairs a language model with a separate index that is queried while the answer is being written, so the model composes a response from retrieved passages rather than from memory alone. Introduced by Lewis and colleagues in 2020, it is the architecture underneath almost every generative answer engine, which is why a passage from your site can appear in an answer even when your page is not the top-ranked result.
Is entity-based search just SEO with a new name?
No. Classic search optimization improves a document's position in a ranked list. Optimizing for entity-based and generative answers improves the odds that a passage from your source is retrieved and cited inside an assembled answer. The two share tactics, clear structure and credible sourcing, but they act on different mechanisms and are measured differently. Treating them as one blended job is the common mistake.
How would I know where my business stands in entity-based search?
You have to measure it directly, because no engine publishes it and classic rank no longer describes it. A structured read samples your real buyer questions across each surface and records whether your business is resolved as an entity and named in the answer. A free Machine-Readiness Score is that starting read.
Provenance
Sources
- Lewis, P. et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks," arXiv:2005.11401, NeurIPS 2020 (established)arxiv.org
- Gao, Y. et al., "Retrieval-Augmented Generation for Large Language Models: A Survey," arXiv:2312.10997, 2023-24 (established)arxiv.org
- Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., Deshpande, A., "GEO: Generative Engine Optimization," arXiv:2311.09735, ACM SIGKDD 2024 (established)arxiv.org
- Singhal, A., "Introducing the Knowledge Graph: things, not strings," Official Google Blog, May 16, 2012 (established)blog.google
- Halpin, H., Hayes, P.J., McCusker, J.P., McGuinness, D.L., Thompson, H.S., "When owl:sameAs Isn't the Same: An Analysis of Identity in Linked Data," ISWC 2010 (established)link.springer.com
- Schema.org, structured-data vocabulary jointly founded by Google, Bing, Yahoo, and Yandex, 2011 (established)
- Chen, M., Wang, X., Chen, K., Koudas, N., "Generative Engine Optimization: How to Dominate AI Search," arXiv:2509.08919, 2025 (emerging, single study not yet replicated)arxiv.org
Every figure above is attributed to a real, dated source and tagged with its evidence tier. Where a claim could not be verified to a primary source, it is not stated as fact.