brightonSEO US 2024: A Measured Approach: Testing UX with LLMs

Slide 1

Slide 1 text

A Measured Approach: Testing UX with LLMs JR Oakes Locomotive Agency Speakerdeck.com/jroakes

Slide 2

Slide 2 text

JR Oakes JR Oakes is the VP of Strategy for LOCOMOTIVE Agency. He has been an SEO since 2011 and was formerly an architectural glass artist. His focus areas are in SEO, machine learning, language, and user experience. jroakes locomotive.agency

Slide 3

Slide 3 text

My Interests Have Broadened a Bit Generative AI Personalization Engines Causal ML Technical SEO

Slide 4

Slide 4 text

We OPENSOURCE Link Link Link

Slide 5

Slide 5 text

If you used an AI tool or technology for web design related tasks, what did you use it for… https://www.mindinventory.com/blog/ai-in-ui-ux-design/

Slide 6

Slide 6 text

71 of consumers expect companies to deliver personalized interactions… 76 percent get frustrated when this doesn’t happen. - McKinsey & Company %

Slide 7

Slide 7 text

A Measured Approach: Types: In-Browser, Proprietary, Inference APIs Examples User State Implementation Testing UX with LLMs

Slide 8

Slide 8 text

CONFIDENTIAL Section 1a In Browser LLMs

Slide 9

Slide 9 text

AI in Chrome

Slide 10

Slide 10 text

AI in Chrome

Slide 11

Slide 11 text

Hacking the AI Example Code

Slide 12

Slide 12 text

Web LLM: High-Performance In-Browser LLM Inference Engine

Slide 13

Slide 13 text

Slide 14

Slide 14 text

No content

Slide 15

Slide 15 text

No content

Slide 16

Slide 16 text

No content

Slide 17

Slide 17 text

No content

Slide 18

Slide 18 text

No content

Slide 19

Slide 19 text

No content

Slide 20

Slide 20 text

Privacy Preserving Increasing Size / Performance ratio Pros Can be tuned to a particular task Reduces need for network requests

Slide 21

Slide 21 text

Cons Takes a long time to load Temporarily bricks your laptop Much less accurate than full-size models

Slide 22

Slide 22 text

Try it out now! Try it out now! Everything Calculator On Desktop

Slide 23

Slide 23 text

CONFIDENTIAL Section 1b Proprietary Systems

Slide 24

Slide 24 text

Vertex AI Search Search

Slide 25

Slide 25 text

Search Search Ask a new question

Slide 26

Slide 26 text

Vertex AI Agents Agents

Slide 27

Slide 27 text

Agents: Data Stores Agents

Slide 28

Slide 28 text

Agents: Tuning Agents

Slide 29

Slide 29 text

Agents: Integration Agents

Slide 30

Slide 30 text

Recommender Vertex AI Recommender

Slide 31

Slide 31 text

Recommender Recommender: Objectives

Slide 32

Slide 32 text

Recommender Recommender: Documents Google Analytics 4 View

Slide 33

Slide 33 text

Recommender Recommender: Events Google Analytics 4 View

Slide 34

Slide 34 text

Recommender Recommender: Results

Slide 35

Slide 35 text

CONFIDENTIAL Section 1c Inference APIs

Slide 36

Slide 36 text

LLM Inference is getting FAST and CHEAP 0

Slide 37

Slide 37 text

LLMs are becoming very dependable

Slide 38

Slide 38 text

Inference in ~1 second Based on the collected information, here’s a JSON object representing the core high-level objectives of the user: Groq API

Slide 39

Slide 39 text

Very fast Increasing Size / Performance ratio Pros Can be tuned to a particular task Cost/1M tokens dropping drastically

Slide 40

Slide 40 text

Cons Price prohibitive for very high-volume sites Dependant on potentially bubble companies Data shared with third-parties

Slide 41

Slide 41 text

CONFIDENTIAL Section 2 User State as Context

Slide 42

Slide 42 text

Think of User State as Context

Slide 43

Slide 43 text

Think of User State as Context

Slide 44

Slide 44 text

This is Great Context

Slide 45

Slide 45 text

What else can we add?

Slide 46

Slide 46 text

What else can we add?

Slide 47

Slide 47 text

Code for Collecting User State Example Code

Slide 48

Slide 48 text

CONFIDENTIAL Section 3 Example Workflows

Slide 49

Slide 49 text

Framework

Slide 50

Slide 50 text

Example 1: Reducing Cognitive Load

Slide 51

Slide 51 text

Example 1: Reducing Cognitive Load We can infer some preference from the context

Slide 52

Slide 52 text

Example 1: Reducing Cognitive Load Filter to relevant pages or sets

Slide 53

Slide 53 text

Example 2: Reduce Friction by Providing Answers Addition of RAG

Slide 54

Slide 54 text

Example 2: Reduce Friction by Providing Answers

Slide 55

Slide 55 text

Example 2: Reduce Friction by Providing Answers Can we find that important buried answer?

Slide 56

Slide 56 text

Example 2: Reduce Friction by Providing Answers To unlock/enable the sale?

Slide 57

Slide 57 text

Example 3: Data Labeling/Cleaning

Slide 58

Slide 58 text

Example 3: Data Labeling/Cleaning ➔ Infer location Other Examples ➔ Assign to user cohorts ➔ Assign client vs prospect behaviour ➔ Infer product/service interest category ➔ Include summary of browsing behaviour in contact forms

Slide 59

Slide 59 text

Example 4: Guiding the User Code

Slide 60

Slide 60 text

Example 5: Highlight Relevant Features

Slide 61

Slide 61 text

Example 6: Dynamic Testimonial Insertion

Slide 62

Slide 62 text

Example 7: Dynamic content reordering 1 1 2 2 3 3 4 4 5 5 6 6

Slide 63

Slide 63 text

Example 8: Dynamic Help Context shows frustrating activity

Slide 64

Slide 64 text

CONFIDENTIAL Section 4 Implementation

Slide 65

Slide 65 text

Code Size and Complexity is an Issue Source

Slide 66

Slide 66 text

Use JSON

Slide 67

Slide 67 text

Creepy Factor To minimize, focus on prioritization and curation Creepy Sensitivity Index Low High Medium

Slide 68

Slide 68 text

Cost

Slide 69

Slide 69 text

Controlling Costs ➔ Use heuristics to know when it is helpful to utilize models; ➔ Control usage to particular site sections; ➔ Select models carefully. Many times a smaller size model is adequate to the task; ➔ Minimize output (e.g. Select the number rather than print the name)