How the Future of Search Works

Transcript

1. 1 How the Future of Search Works Michael King iPullRank Speakerdeck.com/ipullrank @iPullRank
2. 2 2 Download this deck: https://speakerdeck.com/ipullrank
3. 3 Salutations! I’m Mike King (@iPullRank)
7. 7 7
8. 8 Our understanding of how Google Works is Out of Date
9. 9 9 In SEO we still think Google is Here
10. 10 10 Google Shifted from Lexical to Semantic a Decade Ago
11. 11 11 Google Has Been More Like This Since Hummingbird
12. 12 Semantic Search is Fueled by High Density Embeddings …just like large language models.
13. 13 13 Since the Introduction of BERT, Google Has Looked More Like This
14. 14 14 Under the SGE Model, Google is Structured Liked This
15. 15 15 This is a huge problem because SEO software still operates on the lexical model.
16. 16 The threat of Google’s Search Generative Experience (SGE)
17. 17 At I/O Google Announced a Dramatic Change to Search The experimental “Search Generative Experience” brings generative AI to the SERPs and significantly changes Google’s UX.
18. 18 18 Queries are Longer and the Featured Snippet is Bigger 1. The query is more natural language and no longer Orwellian Newspeak. It can be much longer than the 32 words that is has been historically in order 2. The Featured Snippet has become the “AI snapshot” which takes 3 results and builds a summary. 3. Users can also ask follow up questions in conversational mode.
19. 19 19 Sundar is All In. In Sundar’s recent press run he keeps saying how Google will be doubling down on SGE. So it’s going to be a thing moving forward.
20. 20 20 The Search Demand Curve will Shift With the change in the level of natural language query that Google can support, we’re going to see a lot less head terms and a lot more long tail term.
21. 21 21 The CTR Model Will Change With the search results being pushed down by the AI snapshot experience, what is considered #1 will change. We should also expect that any organic result will be clicked less and the standard organic will drop dramatically. However, this will likely yield query displacement.
22. 22 Rank Tracking Will Be More Complex As an industry, we’ll need to decide what is considered the #1 result. Based on this screenshot positions 1- 3 are now the citations for the AI snapshot and #4 is below it. However, the AI snapshot loads on the client side, so rank tracking tools will need to change their approach.
23. 23 23 Context Windows Will Yield More Personalized Results SGE maintains the context window of the previous search in the journey as the user goes through predefined follow questions. This will need to drive the composition of pages to ensure they remain in the consideration set for subsequent results.
24. 24 The State of Google’s Search Generative Experience (SGE)
25. 25 25 Ask to Trigger.
26. 26 26 Auto-Trigger.
27. 27 27 HELLO EMPTY REAL ESTATE!
28. 28 28
29. 29 29 Now we’re seeing autoTriggered results with Show more
30. 30 30
31. 31 31
32. 32 32 I scraped 91k keywords to find out what’s going on…
33. 33 33 It’s an “experiment” so we don’t know much, but here’s what we can infer.
34. 34 34
35. 35 35 Autoloading AI Snapshots initially took 11- 30 seconds to load.
36. 36 36 Average AI Snapshot Load Time 6.08 seconds
37. 37 37 We Know that Featured Snippets Take 35.1% of Clicks
38. 38 38 Snapshot Type Distribution
39. 39 39 39.66% Results Had AI Snapshots
40. 40 40 AI Snapshot Speeds Vary by Category
41. 41 41 Positions 1,2 & 9 are Most Often Present in the AI Snapshot
42. 42 42 The AI Snapshots Most Often Use Six of the Top Ten Results
43. 43 43 Using All that Information We’re Modeling the Threat of SGE
44. 44 44 Get your threat report: https://ipullrank.com/sge-report
45. 45 What is Retrieval Augmented Generation (RAG)?
46. 46 46 This is Called “Retrieval Augmented Generation” Neeva (RIP), Bing, and now Google’s Search Generative Experience all use pull documents based on search queries and feed them to a language model to generate a response.
47. 47 47 Google’s Initial Version of this is called Retrieval-Augmented Language Model Pre-Training (REALM) from 2021 REALM identifies full documents, finds the most relevant passages in each, and returns the single most relevant one for information extraction.
48. 48 48 DeepMind followed up with Retrieval-Enhanced Transformer (RETRO) DeepMind's RETRO (Retrieval-Enhanced Transformer) is a language model that combines a large text database with a transformer architecture to improve performance and reduce the number of parameters required. RETRO is able to achieve comparable performance to state-of-the-art language models such as GPT-3 and Jurassic-1, while using 25x fewer parameters.
49. 49 Google’s Later Innovation Retrofit Attribution using Research and Revision (RARR) RARR does not generate text from scratch. Instead, it retrieves a set of candidate passages from a corpus and then reranks them to select the best passage for the given task.
50. 50 50 SGE is built from REALM/RETRO/RARR + PaLM 2 and MUM MUM is the Multitask Unified Model that Google announced in 2021 as way to do retrieval augmented generation. PaLM 2 is their latest (released) state of the art large language model. The functionality from REALM, RETRO, and RARR is also rolled into this.
51. 51 51 If You Want More Technical Detail Check Out This Paper https://arxiv.org/pdf/2002.08909.pdf
52. 52 Retrieval-Enhanced Transformer (RETRO) https://arxiv.org/pdf/2112.04426.pdf
53. 53 53 RARR Paper https://arxiv.org/pdf/2210.08726.pdf
54. 54 54 Sounds cool, but how does it work?
55. 55 55
56. 56 56 It’s so easy that I built a proof of concept
57. 57 57
58. 58 58
59. 59 59
60. 60 60
61. 61 61 AvesAPI + Llama Index + ChatGPT = Raggle Rankings data Vector index & operations Clearly you know what this does.
62. 62 62 It’s pretty simple # Make an index from your documents index = VectorStoreIndex.from_documents(documents) # Setup your index for citations query_engine = CitationQueryEngine.from_args( index, # indicate how many document chunks it should return similarity_top_k=5, # here we can control how granular citation sources are, the default is 512 citation_chunk_size=155, ) response = query_engine.query("Answer the following query in 150 words: " + query)
63. 63 63 Limitations of my POC It doesn’t do follow up questions It’s not responsive It only does the informational snippet
64. 64 64 You can play with it at raggle.net
65. 65 Optimizing for SGE?
66. 66 66 Search Works Based on the Vector Space Model Let’s go back to the vector space model again. This model is a lot stronger in the neural network environment because Google can capture more meaning in the vector representations.
67. 67 67
68. 68 Dense Retrieval You remember “passage ranking?” This is built on the concept of dense retrieval wherein there are more embeddings representing more of the query and the document to uncover deeper meaning.
69. 69 69 Dense Retrieval is Scoring down to the Sentence Level
70. 70 70 It’s all about the chunks. So use Llama Index to determine the your chunks and improve the similarity to the query.
71. 71 71
72. 72 72 The Fraggles Show What SGE Used for the AI Snapshot
73. 73 73 Fraggles Relevance Relevance against the chunks to keyword: Relevance against AI Snapshot:
74. 74 74 What is Mitigation for SGE? 1. Manage expectations on the impact 2. Understand the keywords under threat 1. Re-prioritize your focus to keywords that are not under threat 1. Optimize the passages for the keywords you want to save
75. 75 The Content Opportunity of RAG
76. There’s a Lot of Synergy Between KGs and LLMs There are three models gaining popularity: 1. KG-enhanced LLMs - Language Model uses KG during pre-training and inference 2. LLM-augmented KGs - LLMs do reasoning and completion on KG data 3. Synergized LLMs + KGs - Multilayer system using both at the same time https://arxiv.org/pdf/2306.08302.pdf Source: Unifying Large Language Models and Knowledge Graphs: A Roadmap
77. Organizations are doing RAG with Knowledge Graphs • Anyone can feed their data into an LLM as a fine-tuning measure to improve the output. • People are currently using their knowledge graphs to support this.
78. 78 78 The code is not much different sitemap_url = "[SITEMAP URL]" sitemap = adv.sitemap_to_df(sitemap_url) urls_to_crawl = sitemap['loc'].tolist() ... # Make an index from your documents index = VectorStoreIndex.from_documents(documents) # Setup your index for citations query_engine = CitationQueryEngine.from_args( index, # indicate how many document chunks it should return similarity_top_k=5, # here we can control how granular citation sources are, the default is 512 citation_chunk_size=155, ) response = query_engine.query("YOUR PROMPT HERE")
79. 79 79
80. 80 80 Custom Indexes for RAG Are Coming Soon to AIPRM
81. 81 ChatGPT Responses without RAG vs with RAG RAG yields content that is more likely to be factually correct. Combined with AIPRM’s prompts, you’re able to better counteract the more bland content that is flooding the web. RAG
82. Fact Verification • Although Google has historically said they do not verification of facts. • LLM + KG integrations make this a possibility and Google needs to combat the wealth of content being produced with LLMs. So, it’s likely they will use this functionality. Source: Fact Checking in Knowledge Graphs by Logical Consistency Source: FactKG: Fact Verification via Reasoning on Knowledge Graphs
83. 83 Roll the Credits
84. 84 84 Get your threat report: https://ipullrank.com/sge-report
85. Thank You | Q&A mike@ipullrank.com Award Winning, #GirlDad Featured by Get Your SGE Threat Report: https://ipullrank.com/sge- report Play with Raggle: https://www.raggle.net Download the Slides: https://speakerdeck.com/ipullrank Mike King Chief Content Goblin @iPullRank