The untapped power of vector embeddings

On brightonSEO, Frank tells how to work with vector embeddings as an SEO specialist. He tells you what they are, how you can use them to automate your work and lays out some cases with vector databases.

Transcript

2. You’ll automate redirects, handle AI hallucinations more effectively, and understand what embeddings are
3. We don’t speak the same language Humans think in concepts, computers process data.
5. Source: Google Trends (query: vector embeddings)
6. 6 @frankvndijk What are embeddings? “Embeddings are numerical representations of data (like words, images, or audio) in a multi- dimensional space” Images Audio Text Embedding model 0.9 0.7 0.2 0.6
7. 7 @frankvndijk Images Audio Text Embedding model 0.9 0.7 0.2 0.6
8. 8 @frankvndijk cat Embedding model 0.9 0.7 -0.3 0.6
9. 9 @frankvndijk -0.2 cat Embedding model 0.9 0.7 -0.3 0.6 dog Embedding model 0.9 0.6 -0.2 0.8 pet Embedding model 0.9 0.7 -0.2 0.9 lion Embedding model 0.9 0.2 0.8
10. 10 @frankvndijk cat pet lion dog
11. 11 @frankvndijk cat dog 0.9 0.7 -0.3 0.6 0.9 0.6 -0.2 0.8
12. 12 @frankvndijk What is cosine similarity? “Cosine similarity measures the angle between two embeddings in a multi-dimensional space to determine how similar they are” cat dog
13. 13 @frankvndijk 0 1 Identical No similarity Similarities really arise Cosine similarity Always a score between 0 (or -1) and 1
14. 15 @frankvndijk dog cat lion shark fox meerkat
15. 16 @frankvndijk Practical examples Chatbots Recommendation systems Search engines
16. 17 @frankvndijk Search query ‘spring’
17. 18 @frankvndijk Search query ‘spring season’
18. 19 @frankvndijk “We no longer live in a keyword era, but in an era of search intent” since 2013…
19. 20 @frankvndijk What is a vector database? "A vector database stores data as high- dimensional vectors for efficient similarity searches and AI applications." 0.4 0.8 0.3 0.7 0.3 0.7 0.4 0.6 0.5 0.6 0.5 0.5
20. 21 @frankvndijk Dataset Embedding model LLM Relevant data Answer Embedding model User question
21. 22 @frankvndijk 128.000 tokens context window +/- 96.000 words 4.096 token limit +/- 3.000 words Limits of ChatGPT
22. 23 @frankvndijk 0.4 0.8 0.3 0.7 Embedding model User question Connect ChatGPT to our database
23. 24 @frankvndijk Connect ChatGPT to our database
24. 25 @frankvndijk This has major advantages It helps to prevent hallucinations Have control over what data you use Use real time or new data
25. 26 @frankvndijk Let’s get practical!
26. 27 @frankvndijk Easy way of starting with embeddings Python Screaming Frog
27. 28 @frankvndijk
28. 29 @frankvndijk Different models for creating embeddings
29. 30 @frankvndijk Embedding models from OpenAI text-embedding-ada-002 text-embedding-3-small text-embedding-3-large Released December 2022 1536 dimensions Released January 2024 1024 dimensions Released January 2024 3072 dimensions *Source: benchmark from datacamp.com
30. 31 @frankvndijk Screaming Frog OpenAI API Embedding model Request Response
31. 32 @frankvndijk Correct settings in SF Make sure that your crawl configurator is set properly: • Extraction => Store Rendered HTML • Rendering => JavaScript
32. 33 @frankvndijk Connect with OpenAI Make the connection with OpenAI in your crawl: • Add your API from OpenAI • Choose the “Extract embeddings form page content” template
33. 34 @frankvndijk Visible in your crawl Next, the embeddings will be visible in your crawl: • Go to the AI tab • Scroll to the embeddings
34. 35 @frankvndijk Gemini or Ollama Ollama Gemini
35. 36 @frankvndijk 0.4 0.8 0.3 0.7 0.8 0.4 0.3 0.2 0.2 0.7 0.9 0.1 0.1 0.5 0.3 0.9
36. 37 @frankvndijk Three scripts for embeddings Internal link opportunities Redirect mapping Duplicate content analyses I will give them away
37. 38 @frankvndijk SF crawl with embeddings Checking cosine similarity Checking existing link in HTML Internal link recommendations Webpage embedding Rest of the embeddings Internal link opportunities Checking similarity Checking relevancy Gathering pages
38. 39 @frankvndijk Checking similarity Checking relevancy Gathering pages Gather the pages we want to optimise Gather the pages we want to optimize
39. 40 @frankvndijk Check cosine similarity Checking if the similarity is at least 0.85 Checking similarity Checking relevancy Gathering pages
40. 41 @frankvndijk Checking similarity Checking relevancy Gathering pages Check for in content link Checking for potential link in HTML of page
41. 42 @frankvndijk Checking similarity Checking relevancy Gathering pages Looping through all pages Looping through this steps to find all recommendations
42. 45 @frankvndijk Use my Google Colabs to run them Give it your input Run the script Download the results
43. 46 @frankvndijk What about databases?
44. 47 @frankvndijk Case A client was not present in the informational and orientation phase of the customer journey Solution Creating content based on a semantic search in a vector database
45. 48 @frankvndijk Blog subject Writing process Extract data Blog content DB search Finding the right blog subjects Searching for our products in combination with ‘best’
46. 49 @frankvndijk Blog subject Writing process Extract data Blog content DB search Semantic searches for products that match Getting the products that matches the blog subject
47. 50 @frankvndijk Blog subject Writing process Extract data Blog content DB search Getting the product information Getting the products that matches the blog subject
48. 51 @frankvndijk Blog subject Writing process Extract data Blog content DB search Write content With AI we managed to make a concept of the blog content
49. 52 @frankvndijk Case Another clients website didn’t appear in the AI overviews for key keywords, while competitors did Solution Analyze and predict with embeddings and other things when content could be displayed
50. 53 @frankvndijk Keywords & URLs Write Rewrite Update content Validation Optimized content Prediction Scrape the AI overviews Scraping content shown in the AI overview
51. 54 @frankvndijk Keywords & URLs Write Rewrite Update content Validation Optimized content Prediction Scraping competitors Scraping the content of the competitor shown
52. 55 @frankvndijk Keywords & URLs Write Rewrite Update content Validation Optimized content Prediction More data for the prediction Gathering other relevant data that is important
53. 56 @frankvndijk Keywords & URLs Write Rewrite Update content Validation Optimized content Prediction Prediction partially based on embeddings We predict if the content is capable to be shown
54. 57 @frankvndijk Keywords & URLs Write Rewrite Update content Validation Optimized content Prediction Optimization advice We (re)write content so we are capable to be shown
55. 58 @frankvndijk Results After implementing: • Before: 32.63% of AI Overviews contained a link • After: 54.48% of AI Overviews contained a link +67% +49% Increase in clicks Increase in display of links
56. 59 @frankvndijk Join the embeddings movement Start experimenting with embeddings and discover what’s possible. This is where the future begins
57. 60 @frankvndijk Key takeaways What embeddings are and how they help us as SEO specialists How to automate an internal link audit and redirect mapping The handles for building a vector database and linking it to an LLM 01. 02. 03.
58. THANK YOU! LET’S CONNECT. linkedin.com/in/frankvndijk x.com/frankvndijk