Long, complex Input post:

At Rocket Internet, we once wasted hundreds of hours of work on recommendation engines.
Here’s what we learned. 👇
Back in 2013, during my time at Rocket Internet, we wanted to increase Basket Sizes and Revenues for Lazada (now part of Alibaba) by implementing an AI-based Recommendation Engine. 🤖
Countless vendors and internal data scientists spent weeks gathering data, cleaning data, and testing various machine-learning approaches.
The results were very disappointing as none of the approaches seemed to have an impact.
Then, two changes made all the difference:
1. Changing the target metric
2. Eliminating AI and Machine Learning
Let’s have a quick look at different types of recommendation engines:
• Recommend Substitute Products: This is the “if you liked The Barbie Movie then you will also like Terminator 2” type of recommendation.
• Recommend Complementary Products: This is the “you are buying a Samsung Galaxy, why are you not also buying this matching phone case, charger, and screen protection” type of recommendation.
So, here’s what we did:
Changing the Target Metric:
We used to measure the success of the recommendations by comparing revenue per visit in a test vs a control group. Revenue per visit was not the right metric as recommendations have an impact across sessions, so we changed it to gross margin per visitor.
Eliminating AI and Machine Learning:
We realized two things:
1. Machines require relatively little data to come up with good recommendations for substitute products but they do require a lot of data to come up with good recommendations for complimentary products. In other words: A machine needs a lot of training data to know which charger will work with that Samsung Galaxy. We did not have enough data.
2. 50% of Revenues were generated by just 1000 Products.
The second point meant that it was totally feasible to take AI out of the equation and let a human find the perfect complementary products for those 1000 Bestsellers.
All we had to do was keep that Bestseller list updated and write a simple logic that sorted the products chosen by humans using simple business rules (e.g. prefer high-margin items, or items that needed to be cleared from the warehouse).
The result: a 12% increase in Gross Margin per Visitor!
Without any AI and machine learning.
Do you have examples where out-of-the-box thinking beats AI?

-----

Short and concise new post:

Hundreds of hours wasted on AI recommendations.

We found success without AI.

At Rocket Internet, we tried AI to boost revenue with a recommendation engine. 

Results were disappointing. 

Two changes made all the difference:

• Changed the target metric to gross margin per visitor.
• Eliminated AI for product recommendations.

Here's what worked:

• Humans selected complementary products for 1000 bestsellers.
• Simple logic sorted products based on business rules.

Result: 12% increase in gross margin per visitor.

Sometimes, human intuition beats AI.

P.S.: Join leading data experts and receive tips, strategies, and resources about how to make your CEO and stakeholders love you: https://dataactionmentor.com/signup-free-newsletter/?utm_source=linkedin&utm_medium=social&utm_campaign=post_human_intuition_beats_ai

Note how the agent also creates links including utm parameters independently.

The biggest mistake I made as {{Role}}: 

{{Mistake}}. 

{{Add Context}}. Instead, {{Action}}. 

• {{Action}} 

• {{Action}} 

• {{Action}} 

• {{Action}} 

• {{Action}} 

{{Drive Home Points / Summarize Points}}. 

{{Optional: Additional Contextual Line / Summary}}. 

{{Winning Outcome when following the advice}}.