Unbundling Amazon: Scaling with Machine Learning

Thursday night, the kids are asleep, and I just cooked up an Italian storm. After rinsing my dinner plate and washing my hands, I rip off the final paper towel from the roll. With no extras in the closet, what do I do next?

1. CVS
2. Walgreens
3. Amazon
4. Local Supermarket

Unless there’s an extenuating circumstance requiring paper towels immediately, 75% of Americans start their product search on Amazon. But this isn’t surprising, we’re all hooked on Amazon.

A better question, which I’m fairly confident Amazon won’t reveal, is what percent of singular product searches turn into a multiple product order? What else will I buy on top of paper towels?

This is where the story of machine learning (ML) begins. By unbundling the most successful eCommerce operation, we’ll dive deeper into how smaller eCommerce operations can immediately deploy affordable ML solutions to become their own Amazon.

Predicting Customer Intent

As a store owner, there’s endless headaches in dealing with inventory for thousands of items. Wouldn’t it be nice to know what goods those customers want to buy? That’s the holy grail.

In a perfect world, imagine the store owner could predict each customer’s intent. Before each customer gets to the store, the owner could line up unique products tailored to that specific person – based on what they are most likely to want, need, or buy.

In our not-so-perfect world, this is what Amazon actually does at scale – to the tune of 532MM square feet (9,236 NFL football fields) of warehousing filled to the brim with all types of products. The magic is that they connect the dots to what you want to buy, when you want to buy it, and most importantly: things “you may also like” (Amazon).

Which brings me back to the original point: what else will I add on top of the paper towels that I need? In other words, what else will Amazon help me to add to my cart?

What is Machine Learning (ML)?

For all 50MM daily Amazon users (1.5B+ monthly), Amazon personalizes the web journey in real time to every single customer’s affinities. Without an ML engine, this would be impossible.

All web surfers leave ‘digital footprints’ throughout the internet that can be harnessed by ML. These footprints may be certain web pages visited, purchases made, or details about their geography or browsing device. 

Within a split second: ML engines ingest unlimited data points, draw meaningful connections, and personalize recommendations to every single customer. This diagram paints a picture:

So when I search for “Paper Towels” on Amazon, their ML engine takes into account my digital footprints (millions of data points) such as:

• What’s my purchasing power & lifetime value?
• Do I compare products before buying?
• Will I leave Amazon to purchase directly (DTC)?
• Do I purchase for kids, pets, others?
• What related products have I purchased?

Those are only 5 of the millions of data points. With a bit of deductive reasoning, a non-technical human could guess correctly that I may also want baby wipes, hand soap, and Febreze.

The difference is that an ML engine personalizes recommendations at scale for the 50+ million daily Amazon users, simultaneously. And, notably, machine learning improves over time.

You can dig deeper into personalizing website shopper experiences in this guide.

How does Amazon use ML personalization?

To reach economies of scale, tech companies use ML to optimize human efficiency by eliminating manual data labor with automation. 

Amazon uses these 3 ML web components to capture over 43% of USA ecommerce business:

Personalized Web Journeys

No Amazon journey is the same. For all 50MM+ daily customers, ML engines personalize web experiences simultaneously. Here’s an experiment: find a friend, look up the same product on Amazon, and compare your search results. You’ll likely see different photos, pricing, discounts, product recommendations, and reviews. 

Personalized Emails

Abandoned browse & cart are common practices. If a web visitor abandons their search, Amazon’s ML engine will send an email at the optimal time with personalized recommendations to re-engage them. This is beyond recommending similar products; it’s recommending products based on customer preference.

Personalized Pricing / Promotions

It’s time to accept the fact that Amazon knows our buying power and threshold to convert. They personalize pricing and promotions to every single user throughout the sales funnel. A customer in Beverly Hills with high spending behaviors likely does not care about a few extra dollars like a college student with budget constraints. Bottom line is that this increases Amazon’s bottom line.

Why you need to control your data now

Every month, the ecommerce industry gets news of big tech withholding advertising data.

Be it Amazon, Apple, or Meta – it’s any marketplace’s incentive to withhold customer data from the companies. Why? Because it creates a reliance on their platforms, and keeps your shoppers as their customers.

As soon as you click “Buy Now,” it doesn’t matter who sold you the product. Amazon retains the sole right to retarget you through cross-channel follow ups. That’s because you’re Amazon’s customer, not the vendor’s. 

So what happens when Amazon Basics knocks off a company or turns the lights off because of a minor violation? Truth is that the company has no data to reach previous customers, can’t retarget, and possibly has no brand loyalty… Amazon claims it all. 

So as big tech gradually revokes rights to customer data, sellers will have no choice but to use marketplaces to cash flow. But even though they’re cash flowing, they are not acquiring long term customers. 

Now is the time to start controlling your data to create a long-term viable business… and fine tune your inventory forecasting.

Leveraging existing data to forecast inventory

After several months of an ML engine ingesting your business’s core competencies & customer behaviors, inventory can be forecasted with accuracy. Remember that ML engines are living organisms – they improve over time with the more data that’s fed to them. Given the current state of the global logistics crisis, demand and trend data are integral to forecasting. 

Example: A machine predicts you’ll sell 100,000 paper towels next month. In reality, you sell less and are left with excess inventory. The ML engine ingests the new sales data and retrains the mathematical models to improve for the next inventory prediction. It does this by defining the supply & demand… and then learning from the successes & shortcomings. Here’s the data you likely already have:

• Supply data: Sales data – or transaction data – has information about which product was sold, when it was sold, who it was sold to, billing information, shipping address, price, quantity, etc.
• Demand data: Website behavior data consists of page visits, physical location, time spent on page, and tons of technical information about the browser, cookie trails, device, IP, etc.
• Customer data: New and past shoppers have made purchases – or failed to convert. They’ve filled out product reviews or surveys, responded to your marketing and email campaigns, and have interacted with your brand in a myriad of ways.
• Trend data: Enrich the data sets with current events, shipping trends, supply chain delays, trending search topics, etc.

In this example, the machine is trying to predict how much inventory to order for next month. Now that the ML engine ingested these enriched data sets, you end up with very accurate and ever improving forecasts.

The key to this all is that the ML Engine is always awake, always learning, and always improving. No human can connect the billions of supply data points with the ever evolving demand and trend data. Only a quality ML engine that doesn’t complain about infinite workload could do this… and automate it so your team can focus on what humans do best: growing the business.

You can read more about how data and ML are used to determine demand and customer affinities in this guide.

What sized company can use ML?

A company needs to surpass a data threshold for ML to be better than a human. To consider utilizing a quality ML engine, a company should have at least one of the following:

• Several thousand monthly customers
• Lots of data (10k+ total customers)
• Web traffic above 10k/monthly visitors
• $5MM+ revenue (conditional on above)

If your company is just starting, there are many low cost solutions that are ‘ML-lite’: tools that focus on speed rather than predictive ML. Disclaimer: confirm that you get to keep the data so that you can eventually level up.

The beauty of DTC is that the company controls the journey, branding, communication, marketing, and beyond. Shopify’s future of commerce report expounds on ML-lite tools. Just remember – this type of applied technology is the starting piece of the puzzle to create brand loyalty with automated ML personalization. 

Bringing it all together

I’d be happy to dive deeper into ML and its use cases by sharing a variety of ML solutions from low cost widgets – such as basic Shopify apps – to powerful prediction engines. 

By implementing unique tools to optimize your DTC website, adding ML solutions will transform your website into its own Amazon… except you get to keep the data and customer relationships. 

As always, don’t be a stranger & email me at Evan@JarvisML.com or connect on LinkedIn to geek out on anything data & ML. Like a machine, I work 25 hours a day so reach me anytime. 

Evan Dolgow Bio (LinkedIn)

Evan is currently developing 4MM+ square feet of warehouse fulfillment centers to support the ecommerce industry. After partnering up his portfolio with a national developer, Evan joined Jarvis ML – the founders of Google Ads & Payments machine learning (ML) engine – to empower smaller ecommerce businesses with the same ML capabilities as Amazon.