Podcast: Download (Duration: 43:15 — 49.8MB)
Over the past month I shipped four AI-powered upgrades to Bumblebee Linens and my courses, and three of them are already lifting sales in measurable double-digit ways. This episode is me walking my co-host Toni Herrbach through exactly what I built, why it works, and what any Shopify or WordPress seller could copy without a coding background.
The short version: an “also bought” recommendation engine that lifted product-page conversions 18% on day one, a vector-based onsite search that lifted search-driven sales 10% in two weeks, a private course chatbot (SteveBot) that answers student questions with links to the exact lessons that cover them, and a Discord-based course community that finally replaced Facebook groups.
Here is the full breakdown of each project, the specific tools and libraries involved, what surprised me, and what I would tell you to try first if you run a store or a course.
Get My Free Mini Course On How To Start A Successful Ecommerce Store
If you are interested in starting an ecommerce business, I put together a comprehensive package of resources that will help you launch your own online store from complete scratch. Be sure to grab it before you leave!
Table of Contents
Key takeaways
- AI-driven “frequently bought together” plus visual-similarity fallback lifted Bumblebee Linens product-page conversions ~18% in the first day.
- Vector search on ~1,000 SKUs lifted search-driven sales ~10% in two weeks after fixing a ~60% zero-results problem.
- SteveBot ingests every lesson, blog post, and Seller Summit video and answers questions with links to the source lesson (no hallucinations because it only pulls from your content).
- Discord replaced Facebook groups for the course community. Threads, custom bots, and voice rooms are built in.
- Underlying tech: FP-Growth for basket analysis, image embeddings for visual similarity, a vector database for search, and RAG for the chatbot.
- Warning: the initial lift from any site change tends to overstate the steady-state lift. Expect the 18% to settle closer to 10%.
AI recommendation engine: how “frequently bought together” lifted sales 18%
The biggest short-term win was building an Amazon-style “frequently bought together” section for every product page on Bumblebee Linens, and it lifted sales 18% on the first day it went live. That number will drift down as the novelty wears off, but even a 10% steady-state lift is worth the two days of build time.
The math is done by a Python library called FP-Growth, which takes every order in the database and tells you which products co-occur, at what confidence level, and at what lift percentage. Feed it your full order history and it produces a matrix of “if a customer bought A, they are N% more likely to buy B.”
The catch on a large catalog is coverage. Bumblebee Linens has close to 1,000 SKUs, and following the 80/20 rule, only the top 20% had enough co-purchase data to populate a real “bought together” module. The bottom 80% would have shown nothing.
Visual similarity fallback: how AI fills the gap for low-data products
For any product with too little co-purchase data, an image-embedding model looks at the product photo, turns it into a numeric vector, and pulls the visually closest products in the catalog as the recommendation. This means every product page shows a populated “bought with” section, even brand new SKUs with zero order history.
If a customer buys a Battenberg lace handkerchief, the model finds every visually similar Battenberg lace handkerchief in the catalog and shows those. The recommendation is grounded in real product features (lace pattern, color, form factor) rather than a random pull.
I applied the same combined engine (co-purchase data plus visual similarity fallback) to two other high-conversion surfaces on Bumblebee Linens: the add-to-cart popup, which roughly doubled its lift after the switch, and the “you might also be interested in” box under each product.
AI onsite search: how vector search fixed a 60% zero-results problem
Vector search replaced my broken Shopify default search and lifted search-driven sales about 10% in two weeks. The trigger was checking my search analytics for the first time in years and discovering that roughly 60% of searches on Bumblebee Linens were returning zero results.
Two root causes. Shoppers cannot spell (there are at least ten ways to write “hanky,” and “Battenberg” trips up everyone), and shoppers type conversational queries like “I’m looking for a wedding handkerchief” that classic keyword search cannot parse. There is a well-known industry stat that visitors who get a zero-result search are around 70% likely to bounce, and Amazon has trained shoppers (especially on mobile) to reach for search first.
The fix is a vector database. I had AI generate a detailed description of every product from its photo, including every occasion it could be used for and every type of buyer who might want it. That description gets embedded into a vector. When a shopper searches “hydrangeas” or “I need something for a bridal shower,” the query is embedded the same way and the database returns the nearest matching products even if the literal keywords never appear in the listing.
SteveBot: how a RAG chatbot answers student questions with source links
SteveBot is a retrieval-augmented chatbot that ingests every course lesson, blog post, and Seller Summit video I have ever recorded, and answers student questions with links back to the exact lessons that cover the topic. It does not hallucinate because it only pulls from my own content.
The problem it solves: my course library now has more than 450 videos, and even I cannot remember which lesson covers which topic. Students were emailing me questions I knew I had answered somewhere, and I was spending 15 minutes hunting through PowerPoint slides trying to find the right video. SteveBot cuts that lookup to seconds for both students and me.
Early feedback has been strong. I have received voicemails about it, students have told me it beats searching WordPress by a wide margin, and I use it myself when a student emails a question I know I have covered. The trickier version of this (Tonibot, the same idea trained on Toni’s course) launched during her month traveling and has been described as “Netflix binge” addictive because students can see anonymized queries.
Discord for course communities: why I moved off Facebook groups
Discord is Slack on steroids and has replaced Facebook groups as my course community platform, primarily because you can code anything into it. Threads, voice rooms that spin up and dissolve on demand, and custom bots (including SteveBot as a Discord bot) all live in one interface.
The interface has a learning curve in the first few days, but the ceiling is much higher than any dedicated course-community platform. I am adding an AI spam-monitor bot that reads every message, classifies whether it is promotional, and drops offenders into a penalty channel. That kind of custom moderation is not really possible in Facebook, Circle, or Skool.
The one interesting finding: the public SteveBot channel in Discord is less used than the standalone SteveBot on the website, because people can see what other members are querying and prefer private lookups. Small privacy affordances matter.
Comparing the four AI projects: what to build first
Here is how the four projects stack up on effort, lift, and how easy they are for a non-coder to copy:
| Project | Result on my site | Build effort | Non-coder path |
|---|---|---|---|
| “Frequently bought together” + visual similarity | +18% sales day 1, likely 10%+ steady state | ~2 days of coding | OneClickUpsell and similar Shopify apps do a lightweight version |
| Vector onsite search | +10% search-driven sales in 2 weeks | ~2 weeks of setup | Klevu, Searchspring, Algolia-style AI search apps |
| Course chatbot (SteveBot) | Massive time-save on student support | Multi-week build + ongoing lesson ingestion | Custom RAG builds only; no true off-the-shelf plugin yet |
| Discord community + bots | Better engagement than Facebook groups | 1 to 2 days to migrate | Native Discord, no coding required for the basic setup |
The first project I would ship on any Shopify store is a real “frequently bought together” module. The second is AI-powered search if your zero-results rate is above 20%. Check that number in your analytics before anything else.
Why real-time AI recommendations are computationally hard
FP-Growth basket analysis and image similarity indexing are computationally intensive and cannot run in real time on shopper visits. On my machine, FP-Growth takes a couple of minutes to run across the full order history, and the image similarity index takes a similar amount of time.
The workaround is to precompute. My similarity index regenerates automatically every time a new product is added, and the basket-analysis job runs once a week. The results are written into the database, and product pages just look them up on render. That is probably why most off-the-shelf Shopify apps in this category feel less accurate: they are shortcutting the problem in ways that trade quality for speed.
AI chatbots for customer support: should you add one to your store?
I am considering an AI live chat widget on Bumblebee Linens for basic questions (shipping status, size charts, order lookup), and I am cautious about it for a specific reason: a bad chatbot experience actively hurts your brand. If a customer types something and nothing responds within 30 seconds, they close the tab and often leave the site entirely.
The pattern that seems to work: be upfront that the visitor is talking to a bot, present a “bubble menu” of the top handful of questions (Where is my order? My order is damaged? Return a product?), and hand off to a human immediately when the query goes off-script. ElevenLabs recently released a voice-based customer service product that sounds human, and that is on my list to test.
The real point here is that customer service is a sales channel. When Toni and I compared her recent American Express call (10 out of 10) with her husband’s recent Ring Doorbell call (a two-hour loop of canned responses), the lesson is that a great customer service interaction can turn a problem into loyalty, and a bad one can end the customer relationship. About 80% of the calls that come into Bumblebee Linens convert to a sale, which is why we still staff human phone support.
The bigger AI shift for ecommerce and SaaS
AI is on track to compress the Shopify app store and a big chunk of the SaaS tools sellers pay monthly for. A lot of the top-grossing Shopify apps do one specific thing and charge $20 to $50 a month for it, and a growing share of those features are now easy to build directly with AI-generated code.
I see sellers stacking a $79 Shopify plan with $532 a month of add-on apps, and much of that stack is doing work AI can now do inside your store code. If you are early in your ecommerce journey, resist plugin bloat. You do not need most of those tools until you have real sales, and by the time you do, half of them will be replaceable.
The other side of this is your team. AI only pays off if the people around you actually use it, and a surprising number of contractors and employees still do not. Getting your team on board with AI for the parts of their work that are repeatable is one of the highest-leverage moves you can make right now.
Frequently asked questions
What AI tools did I use to build the ecommerce recommendation engine?
The core is FP-Growth, a Python library for market-basket analysis that identifies which products are frequently bought together with confidence and lift scores. For products without enough co-purchase data, I use image embeddings (turning each product photo into a vector) and a similarity index to surface visually similar products as the fallback recommendation.
How much did AI onsite search lift my sales?
AI-powered vector search lifted search-driven sales about 10% in the first two weeks after launch on Bumblebee Linens. The bigger prize is that it eliminated a roughly 60% zero-results rate, which was silently costing me a large share of my mobile traffic.
Can I build a course chatbot like SteveBot without coding?
There is no true plug-and-play WordPress plugin for a full custom RAG chatbot like SteveBot yet. You can build a lightweight version with tools like Custom GPTs, Chatbase, or a purpose-built RAG service, but ingesting a large video library requires transcribing every video and running the transcripts through the ingestion pipeline for each new lesson.
Why is Discord better than Facebook groups for course communities?
Discord supports threads, voice rooms that spin up on demand, and custom bots you can code, which Facebook groups do not. You can also integrate an AI chatbot directly into the server, run an AI-powered spam moderation bot, and control the interface much more finely than any of the dedicated course-community platforms.
Will AI kill Shopify apps and SaaS tools?
Many Shopify apps that charge $20 to $50 a month for a single feature are at real risk, because those features are increasingly easy to build directly with AI-generated code. Complex SaaS tools with hosting, integrations, and support requirements will hold on longer, but sellers should audit their monthly app spend now. Plugin bloat is a real drag on new stores that have not yet made their first meaningful sale.
What should I build first if I want to add AI to my ecommerce store?
Start with a real “frequently bought together” module on your product pages. It is the fastest path to a measurable sales lift and it works on almost any catalog. Second priority is AI-powered onsite search if your zero-results rate is above 20%. Check your search analytics before you buy anything.
How do I know if my Shopify search is broken?
Check your search analytics for the “no results” rate. If more than 20% of searches on your store return zero results, you almost certainly have a spelling, synonym, or intent problem that vector search would fix. Anything approaching my old 60% number is actively bleeding revenue.
I Need Your Help
If you enjoyed listening to this podcast, then please support me with a review on Apple Podcasts. It's easy and takes 1 minute! Just click here to head to Apple Podcasts and leave an honest rating and review of the podcast. Every review helps!
Ready To Get Serious About Starting An Online Business?
If you are really considering starting your own online business, then you have to check out my free mini course on How To Create A Niche Online Store In 5 Easy Steps.
In this 6 day mini course, I reveal the steps that my wife and I took to earn 100 thousand dollars in the span of just a year. Best of all, it's absolutely free!
