The Evolution of RAG, Quirks and SEO Actionable Insights

LLMs in AI search rely on Retrieval Augmented Generation (RAG) to ensure there are guardrails and trust around the answers that are provided. Much of RAG comes from search engine index's in the traditional SEO sense, in addition to knowledge bases, data stores and other potential documentation as well as potentially nowadays images and other multi-media types of content. RAG has evolved over time from very basic 'naive RAG' very similar to the very early days of keyword matching in traditional search engines, through to modular, advanced, graph RAG and now Agentic RAG which is able to analyse and review the responses provided and improve upon itself as well as potentially collaborate with other agents. But what has this got to do with SEO and how can SEOs optimise for RAG? Well, there are quirks in the geometric background and landscape of RAG which make some things very different to traditional SEO and search. Exploding spaces mean 'The Curse of Dimensionality' relies on a lean index and the continual need for updated information is dominant too. This deck looks at ways that SEOs can take this into consideration and optimise accordingly.

Transcript

1. PRESENTATION The Evolution of RAG, it's quirks, and actionable insights Dawn Anderson - Bertey 01 02 03 HISTORY GEOMETRY SEO IMPACT Pillar 01 Pillar 02 Pillar 03
2. ORIENTATION We all Remember Ungrounded LLM Credibility EAT ROCKS RAG × SEO GLUE ON PIZZA RUNNING WITH SCISSORS The Evolution of RAG
3. But... Evidence not retrieved can't get generated... nor cited RAG × SEO The Evolution of RAG 3
4. ORIENTATION RAG is the Saviour of LLM Credibility RAG did not replace information retrieval. It made retrieval quality important inside generation. 01 Retrieval is the gate Pillar 01 RAG × SEO 02 03 Geometry has quirks Optimise for retrievability Pillar 02 Pillar 03 The Evolution of RAG 4
5. The Key Shift for SEOs TO (RAG ERA) FROM (TRADITIONAL SEO) “Can my page rank?” Focus on URL-level indexing, SERP position 1–10, and driving direct organic web traffic. RAG × SEO  “Can my evidence become a preferred retrieval unit for this intent?” The Evolution of RAG 5
6. ORIENTATION What we will 'quickly' cover IR roots RAG 2020 Naive RAG Advanced RAG Graph / multimodal FOCUS 01 FOCUS 02 Why RAG emerged from decades of IR research How the stack is changed FOCUS 03 FOCUS 04 The curious geometric quirks of RAG Some practical takeaways RAG × SEO The Evolution of RAG 6
7. The Seven Gates of RAG: A Mental Model Think of modern answer systems as a sequence of gates. Discover Represent Retrieve Rerank Assemble Generate Cite SEO can influence the first five gates directly AND the final two are downstream consequences (WHERE WE CAN WIN). RAG × SEO The Evolution of RAG 7
8. Before RAG: the retrieval lineage RAG is the newest layer on top of a much older problem: selecting the right evidence from a large corpus. Classic information retrieval of old. HISTORY
9. THE LINEAGE Classical IR: lexical matching Search began by matching terms, weighting rarity and ranking documents. 01 02 Bag of words TF–IDF / BM25 Unordered collection of words, ignoring grammar and word order but keeping frequency. Statistical weighting systems prioritizing term rarity and document saturation. 03 SEO inheritance (a lot still matters) Foundational structure, crawlability, and keyword relevance remain fundamental. RAG × SEO The Evolution of RAG 9
10. THE LINEAGE The semantic turn From exact words to learned representations of meaning. 01 02 03 Latent space Word embeddings Contextual encoders Mapping concepts into continuous geometric vector spaces. Dense vector representations capturing semantic similarities. Dynamic embeddings understanding word meaning in sentence context. RAG × SEO 04 Dense passage retrieval Retrieving passage-level evidence directly from semantic space. The Evolution of RAG 1 0
11. THE LINEAGE LLMs arrive - Parametric memory hits a wall PARAMETRIC MEMORY NON-PARAMETRIC MEMORY What LLMs knew • Internalised knowledge stored directly in model weights • • What retrieval added + • External document index for dynamic information access Static and frozen at the time of pre-training • Prone to hallucinations when memory is incomplete Continuously updateable without costly re-training • Verifiable source context to anchor generation RAG = parametric reasoning + non-parametric memory Source: Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks” (2020). RAG × SEO The Evolution of RAG 1 1
12. I repeat... RAG is the Saviour of LLM Credibility RAG × SEO The Evolution of RAG
13. THE LINEAGE 2020: RAG gets its name Lewis et al. formalised a general-purpose recipe combining a seq2seq generator with a dense Wikipedia index. Question Dense retriever Top passages Generator Answer STAGE 01 STAGE 02 STAGE 03 A Question Find Augmenting Information Write The Answer The user provides an input prompt or query requiring domain knowledge or factual accuracy. Dense retriever queries the document index to retrieve top matching passages to enrich prompt context. Seq2seq generator uses both parametric model knowledge and non-parametric retrieved facts to generate the final answer. Source: Lewis et al. 2020, NeurIPS / arXiv:2005.11401. RAG × SEO The Evolution of RAG 1 3
14. THE LINEAGE The Big Idea Behind RAG in 2020 Shifting AI architecture from closed-book static memory to dynamic open-book lookup. THE CORE INSIGHT "An AI Model Can Look Up Information to Add To Its Basic Knowledge" 01 / PARAMETRIC MEMORY 02 / NON-PARAMETRIC INDEX Basic Knowledge Information Lookup Internalized neural weights stored during initial pre-training. Provides core reasoning, language fluency, and broad general facts, but cannot update dynamically without re-training. An external retrieval mechanism querying document indexes at inference time. Feeds specific, verifiable, and up-to-date source passages into the prompt context. Source: Lewis et al. 2020, NeurIPS / arXiv:2005.11401. RAG × SEO The Evolution of RAG
15. THE LINEAGE Why RAG mattered Question Dense retriever Top passages Generator Answer 01 / VALUE 02 / VALUE 03 / VALUE 04 / VALUE 05 / VALUE Freshness Provenance Economics Control Dynamic updates without retraining static model weights. Direct attribution and verifiable source passages. Long-tail Knowledge Cost-effective indexing vs continuous model training. Precise governance over document access and updates. Access to specialised enterprise and domain facts. Source: Lewis et al. 2020, NeurIPS / arXiv:2005.11401. RAG × SEO The Evolution of RAG 1 5
16. STAGE 1 Naive RAG architecture – Seductively simple Documents Chunk Embed Vector DB Top-k Stuffing Prompt + LLM Elegant. Fast to prototype. Full of hidden assumptions. RAG × SEO The Evolution of RAG 1 6
17. STAGE 1 Naïve - Chunking became a ranking decision 01 / NAÏVE APPROACH 02 / NAÏVE APPROACH 03 / OPTIMAL APPROACH Too Small Too Large Context Loss Multi-Topic Dilution Better Architecture Over-granularity breaks semantic coherence, causing retrieval of fragments missing critical context. Overly broad chunks combine unrelated topics, diluting the single vector embedding's focus. Units Can Stand Alone Self-contained information units preserve full semantic context while ensuring precise vector retrieval. SEO implication: page architecture determines the candidate units a RAG system can retrieve. RAG × SEO The Evolution of RAG 1 7
18. STAGE 1 But... Single-vector compression is lossy A 500-word chunk entity A feature B exception C number D comparison E ONE VECTOR Compression preserves useful semantics but not every fine-grained signal equally. RAG × SEO The Evolution of RAG 1 8
19. STAGE 1 Failure modes of naive RAG Ambiguous query Rare entity Long document Conflicting sources Dense-only Fixed chunk Top-k stuffing High Med High High Med Med Med High High Med Low High The recurring pattern: recall and precision fail at different stages, then the generator inherits the error. RAG × SEO The Evolution of RAG 1 9
20. Stage 2: Advanced RAG Production systems added query rewriting, hybrid retrieval, metadata, reranking and smarter context assembly.
21. STAGE 2 Hybrid retrieval: lexical + semantic Sparse / lexical Dense / semantic Hybrid search treats these as complementary signals, not competing methods. RAG × SEO The Evolution of RAG 2 1
22. STAGE 2 Reranking separates recall from precision 1. Retrieve 50–200 2. Rerank shortlist 3. Keep best 3–10 4. Generate STAGE 1 • RECALL STAGE 2 • PRECISION “Could this be relevant?” “Which evidence is best?” Broad retrieval prioritizing coverage to ensure no potential matches are missed. Deep scoring and filtering to surface only the most exact context for generation. RAG × SEO The Evolution of RAG 2 2
23. STAGE 2 Query transformation changed the query itself 01 02 03 04 Rewrite Expand Decompose HyDE Rephrase user queries to improve clarity and retrieval precision. Generate related terms and synonyms to broaden context coverage. Break complex multi-part questions into simpler sub-queries. (Hypothetical Document Embeddings) Generate hallucinated answers to search for semantically similar docs. SEO implication: you may be retrieved for a transformed query you never see in keyword tools. RAG × SEO The Evolution of RAG 2 3
24. STAGE 2 Context assembly became its own ranking layer PIPELINE STRUCTURE 01Deduplicate Remove redundant passages & overlapping tokens Building the Prompt Context 02Diversify Ensure balanced coverage across multiple perspectives 03Order Sequence facts strategically to mitigate attention drop-off 04Compress Maximize information density within context limits Refining raw chunks into optimal LLM evidence RAG × SEO 05Cite Map output claims explicitly to retrieved sources The Evolution of RAG 2 4
25. Stage 3: Structured, graph & multimodal RAG Retrieval expanded beyond flat text chunks into relationships, tables, images, layouts and graph-connected evidence.
26. GraphRAG: retrieve relationships AND passages Topic Entity B Source Entity A Event Claim Useful when the answer depends on multi-hop relationships, communities, entities or global themes. RAG × SEO The Evolution of RAG 2 6
27. Multimodal RAG broadens 'evidence'     Text Tables Images Layout Unstructured prose, articles & documentation Structured relational data & spreadsheets Visual content, diagrams & infographics Document hierarchy & spatial arrangement SEO expands from 'indexable text' to 'machine-readable evidence across modalities'. RAG × SEO The Evolution of RAG 2 7
28. Long context did not kill retrieval  Long-Context Advantage  Retrieval Advantage Deep Reasoning Window Precision & Efficiency Synthesizes full-document nuance and overarching intent without fragmentation. Pinpoints specific factual evidence at sub-second speeds with reduced compute overhead. Global Coherence Dynamic Scale Maintains comprehensive context memory across extensive conversation threads. Seamlessly queries billions of fresh, external data points on demand. The emerging pattern is hybrid: retrieve intelligently, then reason over a larger selected context. RAG × SEO The Evolution of RAG 2 8
29. Stage 4: Agentic RAG The retriever becomes a tool used iteratively: plan, search, inspect, reformulate, search again, verify.
30. Agentic retrieval is a loop Plan Verify Search Goal Reformulate Inspect Retrieval can now adapt based on what is missing, contradictory or uncertain. RAG × SEO The Evolution of RAG 3 0
31. What agentic RAG changes for SEO 01 02 03 04 One question can trigger many hidden searches RAG searches for 'missing' evidence Beyond single-pass answers Must be useful for sub-steps THE OVERARCHING PARADIGM SHIFT From 'keyword → page' to 'task → evidence graph' RAG × SEO The Evolution of RAG 3 1
32. Quirkiness - The geometry underneath RAG Dense retrieval works by turning language into coordinates. That geometry is powerful and counterintuitive.
33. Embeddings: meaning as position Cluster: Data Science jaguar Python arrays NumPy dataframes Cluster: Footwear trainers running shoes marathon footwear Similarity is inferred from neighbourhood structure And not from a human-readable taxonomy. RAG × SEO The Evolution of RAG 3 3
34. The curse of dimensionality As dimensionality grows, intuitive notions of 'near' and 'far' can degrade. 01 • EXPANSION 02 • CONVERGENCE 03 • DEGRADATION    Space explodes Distances concentrate Noise dimensions hurt Pairwise distances converge, making the relative contrast between nearest and farthest neighbors disappear. Irrelevant or noisy attributes accumulate variance, completely masking true semantic similarity. Volume grows exponentially with each added dimension, causing data points to become extremely sparse. Background: nearest-neighbour literature on the curse of dimensionality; see Cornell CS4780 notes and standard IR/ML treatments. RAG × SEO The Evolution of RAG 3 4
35. Hubness: some vectors become 'popular neighbours' In high-dimensional k-nearest-neighbour graphs, a small number of points appear in many neighbour lists while 'antihubs' appear rarely. 01 • HUBS ("POPULAR NEIGHBOURS") Frequent Retrieval A small subset of vectors appear in an unexpectedly high number of k-NN neighbour lists across varied queries. HUB 02 • ANTIHUBS ("ISOLATED POINTS") Rare / Zero Retrieval High In-Degree (Hub) Zero In-Degree (Antihub) ANTI-HUB Regular Vectors Vectors located in sparse regions almost never get retrieved as nearest neighbours, making niche information hard to reach. Source: Radovanović, Nanopoulos & Ivanović, JMLR 2010, “Hubs in Space”. RAG × SEO The Evolution of RAG 3 5
36. Why hubs matter in RAG Key retrieval vulnerabilities introduced by high-dimensional vector hubness. 01 • RECURRENCE 02 • ISOLATION 03 • PROPAGATION    Generic chunks recur Niche evidence disappears Errors propagate High-degree hub vectors act as topological defaults, appearing in neighbor lists across completely unrelated queries. Valuable specific content residing in sparse regions (antihubs) rarely gets retrieved, hiding critical nuance. When hub chunks contain inaccurate or generic context, downstream LLMs continuously amplify those false assumptions. Hubness has been shown to distort nearest-neighbour relations and retrieval in high-dimensional spaces (Radovanović et al.; Feldbauer & Flexer). RAG × SEO The Evolution of RAG 3 6
37. Hubness ≠ authority A retrieval hub is a geometric property, not an editorial endorsement. 01 • VECTOR SPACE PROPERTY 02 • EDITORIAL ENDORSEMENT   Geometric hub True authority High in-degree vector that appears in many nearest-neighbour lists due to mathematical spatial artifacts, not content quality. Genuine expertise, editorial trust, and accurate domain knowledge validated by human consensus and rigorous verification. SEO RISK Confusing repeated retrieval with genuine authority can reward bland, generic content until reranking or verification corrects it. RAG × SEO The Evolution of RAG 3 7
38. Semantic crossroads Content positioned between multiple semantic neighbourhoods. SEO Information Retrieval LLMs crossroa d Crossroads can be powerful bridges or ambiguous pages with no dominant retrieval identity. RAG × SEO The Evolution of RAG 3 8
39. When semantic crossroads help  Comparison pages  Glossaries / explainers Explicit side-by-side evaluations connecting intent across product or concept boundaries. Structured definitions linking domain-specific terminology into broader contextual networks.  Category hubs  Interdisciplinary content Curated thematic topic collections (distinct from mathematical vector hubness). Synthesized insights uniting multiple separate domains into cohesive frameworks. THE MANDATORY CONDITION The bridge must be explicit. State the relationship, entities, scope and intended question clearly. RAG × SEO The Evolution of RAG 3 9
40. When semantic crossroads hurt      Multi-intent pages Ambiguous entity names Boilerplate dominance Disconnected topics Conflicting internal linking THE RESULT Broad retrievability, weak rerankability or retrieval for the wrong intent. RAG × SEO The Evolution of RAG 4 0
41. What this means for SEO The optimisation target is no longer only a document. It is a portfolio of retrievable, attributable evidence units.
42. SEO reframed as evidence optimisation Intent Entity clarity Passage design Traditional SEO asks whether a URL deserves a rank. Retrieval Rerank Citation RAG-aware SEO asks whether a passage deserves to be evidence. Both matter. The second is increasingly important wherever answers are synthesised from retrieved sources. RAG × SEO The Evolution of RAG 4 2
43. RAG × SEO FRAMEWORK Action 1: Make passages self-identifying  01 Name the entity Explicitly identify primary subjects within every passage segment.  02  03 State the relationship Keep qualifiers local Clearly define how entities relate without relying on global context. Embed conditions and scope directly inside the chunk itself. GOAL A chunk should remain understandable after being detached from the page. RAG × SEO The Evolution of RAG 4 3
44. Lead with facts, not prose Action 2: engineer semantic distinctness Use preciseness Reduce boilerplate Different evidence – different page / section Comparisons OK but no blurred lines RAG × SEO The Evolution of RAG 4 4
45. RAG × SEO FRAMEWORK Action 3: Design for hybrid retrieval  01  02 Lexical anchors Semantic coverage Exact keyword matches, precise terminology, and specific entity names to ensure immediate lookup precision. Vector context, conceptual depth, and rich topical relevance to satisfy nuanced search intent and embeddings. STRATEGY Do both. Hybrid retrieval rewards pages that are exact enough to identify and rich enough to understand. RAG × SEO The Evolution of RAG 4 5
46. Action 4: Audit generic 'hubness' content Blah Blah blah Blah, blah, blah, blah RAG × SEO The Evolution of RAG 4 6
47. Action 5: publish evidence shapes models can use Definition Comparison Procedure Specification Claim Best shape Key property RAG value Short passage Explicit scope High Table + prose Same dimensions High Numbered steps Order + conditions High Table Units + version High Claim + source Provenance High Structure is not merely presentation. It changes how evidence survives extraction and retrieval. RAG × SEO The Evolution of RAG 4 7
48. Action 6: measure retrieval, not just traffic Recall@k MRR / rank Does your evidence appear anywhere in the candidate set for target questions? How early does the first relevant passage appear? Citation share Answer coverage How often is your domain/passages selected as attributable evidence? Which user questions can your corpus answer with sufficient evidence? Build a question set from real tasks, not only keywords. RAG × SEO The Evolution of RAG 4 8
49. 12 takeaways for SEO practitioners 1 2 3 4 5 6 RAG × SEO RAG is an IR system before it is a generation system. Retrieval failure caps answer quality. Chunk boundaries are ranking boundaries. Dense vectors trade lexical exactness for semantic reach. Hybrid retrieval is often stronger than dense-only. Reranking creates a second competition after retrieval. 7 8 9 10 11 12 High-dimensional geometry is unintuitive. Hubness can over-promote generic vectors. Antihubs can hide valuable niche evidence. Semantic crossroads need explicit relationships. Optimise passages for detachment and attribution. Measure retrieval and citation behaviour directly. The Evolution of RAG 4 9
50. Th RAG × SEO The Evolution of RAG
51. Sources & further reading Foundational • Lewis et al. (2020), Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, NeurIPS. • Khattab & Zaharia (2020), ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction, SIGIR. • Radovanović, Nanopoulos & Ivanović (2010), Hubs in Space: Popular Nearest Neighbors in High-Dimensional Data, JMLR. Geometry & hubness • Feldbauer & Flexer (2019), A comprehensive empirical comparison of hubness reduction in high-dimensional spaces. • Cornell CS4780 lecture notes: k-nearest neighbours / curse of dimensionality. RAG × SEO The Evolution of RAG 5 1