You Can't Generate What You Can't Retrieve

In this talk, we explore why Large Language Models depend entirely on retrieval systems to produce grounded answers, and how Retrieval-Augmented Generation (RAG) pipelines reintroduce classic search problems under a new name. From document structure and chunking to relevance, freshness, and evaluation, the same decisions SEOs have always influenced now directly affect what AI systems “know.”

Transcript

1. You Can’t Generate What You Can’t Retrieve The Hidden Architecture of AI Search Dawn Anderson
2. AI answers feel intelligent — until they’re confidently wrong.
3. Most AI Search Failures are Retrieval Failures
4. The Misconception • LLMs don’t search. • They generate from supplied context.
5. AI Search Pipeline
6. Why This Matters • Hallucinations • Bias • Outdated answers • Risk
7. Retrieval-Augmented Generation (RAG) • Retrieve first. • Generate second.
8. But RAG is a Process, NOT Magic
9. The Bottleneck Generation quality is capped by retrieval quality.
10. Retrieval is NOT Ranking
11. Classic IR Returns Relevance Structure Freshness Evaluation
12. Documents vs Chunks Retrieval units > pages
13. Chunking Trade Offs
14. Structure as Signal Headings, lists, tables
15. Unstructured or Unreachable Content Retrieval Failure
16. Freshness LLMs don’t know now
17. Stale Retrieval = Wrong Answers Wrong answers
18. Contextual Relevance Same query, different needs
19. Semantic blur - Embeddings Aren't Magic - Similarity ≠ relevance
20. Common RAG Failures Over-chunking Stale indexes Embedding-only retrieval
21. Hybrid Retrieval - Combined Strength Dense + Sparse + Rules
22. Bias Bias starts in retrieval – missing data
23. Hallucinations Fluent compensation - filling gaps
24. Evaluation Blind Spot – Observability Matters
25. Debugging Was it retrieved? Was it correct?
26. From Answers to Actions: Agentic Search – AI Agents Don't Just Answer – they Act on User Intent
27. What is Agentic Commerce? Agents discover, compare, and complete purchases.
28. Universal Commerce Protocol (UCP) Enables Agent-Merchant Interaction at Scale
29. Agentic Commerce Protocol
30. Why UCP & ACP Matter to SEO Visibility now includes agent accessibility, not just rankings
31. Context Graphs Enable Agentic Action Agents need structured understanding to act safely
32. SEO in An Agentic Commerce World is Now Also Agentic Commerce Readiness Optimisation • Optimising for agentic search introduces new priorities • Structured, agent- friendly product data • Capability metadata (e.g., available actions) • High-precision feed timeliness and accuracy • Protocol compliance (UCP, ACP, MCP)
33. What Are Context Graphs? Context Graphs: How AI Maintains Meaning • Context ≠ single document • Meaning emerges across entities, attributes, and relationships • Context graphs connect retrieved fragments into coherent understanding
34. Why Context Graphs Exist & Matter in AI Search? CHUNKING FRAGMENTS MEANING RETRIEVAL RETURNS PARTIAL VIEWS CONTEXT GRAPHS RESTORE COHERENCE
35. Context Graphs vs Knowledge Graphs Context graphs are NOT knowledge graphs Knowledge graphs = static, curated Context graphs = dynamic, query- specific Built at retrieval time, not indexing time
36. Structure Is the Input to Context Graphs Clear entities Explicit attributes Defined relationships Unambiguous boundaries
37. Context Graphs Power AI Commerce • Products are entities • Prices, variants, availability are attributes • Compatibility, bundles, substitutes are relationships
38. Structure Data is No Longer Optional • Structured Data is retrieval infrastructure • Schema ≠ rich snippets • Structured data feeds retrieval • Enables graph construction
39. Structured Data Meets Agentic Protocols • Agents need actionable understanding • Context graphs → decisions → actions • Retrieval must support execution, notjust answers • Protocols assume clean, machine-readable commerce data
40. SEO in a Context- Graph World Entity clarity > keywords. Graph consistency > pages
41. The New Definition of Visibility Being retrievable, understandable, and executable by machines
42. Why SEO Matters You influence what AI knows
43. Structure > prose Entities > keywords Retrieval > ranking SEO Implications
44. Takeaways • Be explicit (clarity) • Stay fresh (freshness) • Design for retrieval (structure)
45. The New Search Stack
46. Final Thought AI doesn’t fail loudly — it fails fluently.
47. You Can't ACT on What You Can't RETRIEVE
48. Thank you X - @dawnieando Bluesky - @dawnieando Linkedin – Ms Dawn Anderson Threads - @dawnieando Bertey.com