.jpg)
Back in 1998, Greg Linden was an engineer at Amazon. At one point, he had an idea. What if, he wondered, the shopping cart page showed recommendations for other items the customer might want to add to their purchase? Linden built something that did exactly that, then began showing it to his colleagues. They were impressed. Nearly everyone was convinced it would improve sales. Except one. One senior marketing vice president pushed back loudly. If we show recommendations, he argued, people will buy fewer things! Amazon loses customers every day who abandon their shopping carts without purchasing; what if these recommendations just distracted shoppers right at the moment of checkout? Enough people rallied behind that argument, and the colleague reportedly told Linden to scrap his project, per Linden's retelling of the story later published in Microsoft and Amazon presentations on experimentation as Exhibit A of what they call the HIPPO problem.
The highest-paid person's opinion is overpowering real data.
No matter. Linden automated it anyway. His experiment was launched to a tiny fraction of Amazon’s real customers. Linden measured what happened to sales. That test won. By a lot. Amazon added recommendations to everyone’s shopping cart page. That small experiment grew into one of the most copied e-commerce studies of all time. Refining it over the next couple of years into their “item-to-item collaborative filtering,” today, what Lindens first shopping cart rode onto sits behind Amazon’s recommendation engine, which you may have heard is responsible for close to 35% of Amazon’s total revenue.
Here’s what you should remember from that story:
Automation isn’t valuable because it helps you do more marketing. It’s valuable because it lets you prove something actually works once, with real data, and then give that proven thing to every future customer automatically without requiring a human to manually make that same decision thousands or millions of times.
Most people approach automation backwards. Instead of setting aside their opinion to let the data decide what actually works, they automate their opinion. Instead of testing a small bit of traffic to see what messaging, emails, or products convert first, they build an entire system on what they assume will work. Linden’s story worked because he proved his idea manually first. Only then did he automate.
#1: Automate the repeatable, not the decision point. Is a task done exactly the same way, word-for-word, every single time it occurs? If yes, that’s a great candidate for automation. Whether that’s sending a welcome email when someone joins your list or sending a reminder email the day before your event. Is that task more nuanced than that? Will sending require reading the individual and adjusting your approach based on who they are? If yes, that task does not belong in automation (at least not yet). The point of automation is to reach thousands or millions of people efficiently. You cannot efficiently send a genuinely human message to that many people. Automating a judgment call produces emails that sound robotic because they are.
#2: Test the manual process first. Before Amazon’s recommendation engine hit millions of customers automatically, it hit a small percentage of customers manually. Don’t build an automated email sequence and turn it on. Send version 1.0 manually to 10 people. Whatever the lowest-effort version of your marketing is, run it by hand to a small group of people first and see what actually converts. Then automate the winning version.
#3: Trigger your email instead of scheduling it. The real brilliance behind Linden's Amazon story wasn’t just that he offered product recommendations. It was that he showed those recommendations only when someone had already demonstrated interest by adding an item to their cart. Automation should always try to be reactive to what someone just did. That signup, that click, that abandoned form. It’s rare that sending an email exactly one week after someone signs up is what’s best for your audience. Email triggers exist for a reason. Use them.
#4: Use tools that make this free. MailerLite and Beehiiv, both tools I recommended last episode for list-building, both offer automation triggers in their free tiers. Welcome sequences, tag-based follow-ups, simple if-this-then-that rules—you don’t need to code to create automated email sequences that respond to how your subscribers interact with you.
#5: Audit your automation regularly. Just because something is automated doesn’t make the HiPPO irrelevant. Set aside time every few weeks to look under the hood at what your automation is actually sending. Like Linden continued to do after his shopping cart launched—instead of setting it and forgetting it—and optimize it as needed. Maybe your welcome series still performs well. But that social follow-up you put in six months ago doesn’t. You’ll never know if you don’t look.
Amazon’s shopping cart doesn’t teach us that machines make better decisions than people. It teaches us that machines will apply a single decision to everyone equally. So if you have a hunch about what works, prove it first. Scale your success. Let the data decide.
