Debunking and Demystifying Generative Information Retrieval

The birth of Generative Information Retrieval was 2023, so we're only a couple of years into the biggest paradigm shift the search world has experienced. Inevitably there's much confusion, mysticism and mythology filling the void of accurate understanding. In this talk you'll learn some of the common jargon - what really matters and what is simply noise (and even perhaps spam which might lead you into hot water). From chunking to vector databases and llms.txt.

Transcript

1. #WTSFest @dawnieando
2. Sigh... New SEO AI Myths We Need To Question #WTSFest @dawnieando
3. SEO Misinformation Is NOT New #WTSFest @dawnieando
4. The arrival of 'AI Search' has exacerbated further with 'SEO AI' myths we now need to question too #WTSFest @dawnieando
5. Some SEO AI Myths We Need to Employ Critical Thinking About Chunking llms.txt Information Gain AI Replaces SEO and more #WTSFest @dawnieando
6. Why SEO Myths Spread RAPID EVOLUTION SOCIAL ECHO CHAMBERS PATENTS ≠ PRODUCTION DESIRE FOR SIMPLE LEVERS AMBIGUOUS TERMINOLOGY #WTSFest @dawnieando
7. Myth 1: 'Chunking' #WTSFest @dawnieando
8. Chunking Has Several Meanings #WTSFest @dawnieando
9. The Magic Number 7 'Chunking' and Information Theory George Miller (1956) Magic Number 7 - +/- 2 Chunks help users commit information to working memory #WTSFest @dawnieando
10. The 'Segmenting' Principle in Education #WTSFest @dawnieando
11. Myth: Google Ranks Chunks • SEOs believe sections rank independently (Passage indexing) • Reality: chunking = ML preprocessing • Google evaluates pages holistically & by intent nowadays #WTSFest @dawnieando
12. Chunking in Machine Learning Chunking is a preprocessing technique in machine learning and NLP Large documents are split into smaller, manageable segments ('chunks') Helps models process context within token limits Common in LLM-based retrieval, embeddings, and summarisation systems #WTSFest @dawnieando
13. Why Chunking Is Used in ML? Language models have token limits Chunking ensures no loss of information Enables vector search and retrieval- augmented generation (RAG) Supports knowledge retrieval rather than ranking or quality evaluation #WTSFest @dawnieando
14. Several Types of ML Chunking Too Method Semantic quality Control Speed Best for Fixed-size Low–medium High Very fast Embeddings, RAG Sentence/paragraph High Medium Fast Summaries, QA Semantic Very high Medium-low Medium Knowledge retrieval Sliding window Medium High Medium Long-context tasks Hierarchical High Medium Medium Structured documents Model-specific Varies High Varies LLMs with long context Adaptive token-aware High High Medium Production RAG #WTSFest @dawnieando
15. Why Chunking Is Misunderstood LLMs use chunks ML terms misread as actionable SEO tactics Narrative simplicity #WTSFest @dawnieando
16. Don't Confuse Semantic Structuring and Focus for 'Chunking' #WTSFest @dawnieando
17. Wikipedia Have Been 'Chunking' For Decades If That's The Case #WTSFest @dawnieando
18. Also... Surprise #WTSFest @dawnieando
19. Chunking Summary • Chunking is an ML processing technique—not an SEO tactic • It has no direct or indirect ranking impact • SEO success comes from content quality, relevance, and site performance • Ignore the hype; stay focused on proven ranking factors #WTSFest @dawnieando
20. Spoiler Write naturally (with semantic structure and logical ordering) for your users #WTSFest @dawnieando
21. What to Focus On Instead Clear, readable, user-first content Strong topical authority and internal linking Satisfying user intent Technical performance and UX Structured data, accessibility, crawlability #WTSFest @dawnieando
22. #WTSFest @dawnieando
23. Next Myth: LLMS.TXT #WTSFest @dawnieando
24. This One is Tricky #WTSFest @dawnieando
25. LLMS.TXT Proposal #WTSFest @dawnieando
26. What LLMs.txt Is LLM friendly markdown file It's nothing like robots.txt It provides easily consumable excerpts of the most important pages on a website Lots of confusion over its purpose It's not for crawling or LLM controlling #WTSFest @dawnieando
27. LLMs.txt Has Competition – Cats.txt #WTSFest @dawnieando
28. I Vote For Pommies.txt Credit: Marco Giordano https://www.linkedin.com/in/marco-giordano96/ #WTSFest @dawnieando
29. Maybe I'm Biased? #WTSFest @dawnieando
30. Uptake is VERY Low #WTSFest @dawnieando
31. No AI System is Using LLMs.txt #WTSFest @dawnieando
32. But Then... #WTSFest @dawnieando
33. Google Dev Added LLMS.txt #WTSFest @dawnieando
34. But Then... #WTSFest @dawnieando
35. …They Removed It #WTSFest @dawnieando
36. Meep #WTSFest @dawnieando
37. LLMs.txt Argument For Agentic Search Efficiency? #WTSFest @dawnieando
38. Agentic Search Argument #WTSFest @dawnieando
39. Credit Andre Volpini – Wordlift – Google Dev Approach? #WTSFest @dawnieando
40. Perhaps It's Being Used In Other Ways? e.g. Agentic Search Clues like... #WTSFest @dawnieando
41. Myth: llms.txt Influences Rankings Controls training? No Signals quality? No Ranking impact: zero #WTSFest @dawnieando
42. Why llms.txt Myth Spreads? AI fear Blog exaggeration Confusion: LLM training vs crawling #WTSFest @dawnieando
43. Next Myth: Information Gain #WTSFest @dawnieando
44. It All Centres Around This - Information Gain Patent #WTSFest @dawnieando
45. Myth: Information Gain Ranking Factor Belief: unique information = ranking gains Reality: patent concept, not necessarily deployed Truth - Google prefers helpfulness #WTSFest @dawnieando
46. Isn't That Just...? #WTSFest @dawnieando
47. #WTSFest @dawnieando
48. Information Gain – A Mathematical Theory of Communications (Claude Shannon, 1948) #WTSFest @dawnieando
49. Information Gain in Machine Learning #WTSFest @dawnieando
50. Classification Modelling Predicting how 'pure' a classifier split in data is Purer data to a 'class' gives higher information gain #WTSFest @dawnieando
51. Information Gain Quantifies the reduction in entropy after splitting the data on a particular feature. Higher gain means a more useful split. #WTSFest @dawnieando
52. Claude Shannon #WTSFest @dawnieando
53. Decision Tree Classifier 'Purity' #WTSFest @dawnieando
54. Did You Spot The Nano Banana Decision Tree Hallucinations? #WTSFest @dawnieando
55. Thanks Geppetto #WTSFest @dawnieando
56. Why Could Information Gain Be Misread Patent worship Desire for measurable factors #WTSFest @dawnieando
57. Why Research ≠ Search RESEARCH PROTOTYPES SCALING ISSUES MOST PAPERS NEVER SHIP #WTSFest @dawnieando
58. And There's More #WTSFest @dawnieando
59. Myth: AI Content is Penalised #WTSFest @dawnieando
60. Lots of Evidence to The Contrary #WTSFest @dawnieando
61. AI Content is Automatically Low Quality
62. Google on AI Generated Content "Automation has long been used to generate helpful content, such as sports scores, weather forecasts, and transcripts" #WTSFest @dawnieando
63. Spoiler: Every Ecommerce Site is 'Automated' Content #WTSFest @dawnieando
64. AI Takes Over SEO #WTSFest @dawnieando
65. Is AI Overview 'Settling Down'? #WTSFest @dawnieando
66. And... Of Course... The Classic #WTSFest @dawnieando
67. But... SEO is in a Hybrid State #WTSFest @dawnieando
68. We Can't Even Agree on Our Own Name #WTSFest @dawnieando
69. The Current Learning Curve is Steep
70. SEO Focus Areas Helpfulness Originality Entity Clarity Crawl Efficiency Freshness #WTSFest @dawnieando
71. Advice from Bing's Fabrice
72. Finally... Search is complex Critical thinking is increasingly important Disambiguation of terminology is critical #WTSFest @dawnieando
73. Thank you X - @dawnieando Bluesky - @dawnieando Linkedin – Ms Dawn Anderson Bertey.com #WTSFest @dawnieando