Analytics
Every other eCommerce consultant sells AI personalisation as a cure-all. Dynamic product recommendations. Predictive pricing. AI-driven email flows. The promises are impressive. The results in practice usually are not.
Every other eCommerce consultant sells AI personalisation as a cure-all. Dynamic product recommendations. Predictive pricing. AI-driven email flows. The promises are impressive. The results in practice usually are not.
Not because AI is bad, but because most shops do not have enough data to use AI sensibly, and because simple rules often deliver the same result without the overhead.
Where AI personalisation has been compared with a rule-based system, the rule usually wins or the difference is not significant. Worked example: 10 shops run the same comparison and in 8 of them the rule wins or the difference never reaches significance.
The AI is more expensive, more complex and harder to debug, which is how a positive lift turns into a negative return.
AI needs three things: data, data and data. A machine learning model that generates product recommendations needs at least 50,000 sessions a month and 6 months of history to detect statistically reliable patterns. Most shops that buy AI tools have 10,000 sessions and 3 months of data. The model guesses, and often guesses wrong.
A fashion shop with 35,000 sessions a month implemented an AI recommendation tool for 2,400 euros a month. The promise: “The AI learns your customers’ buying behaviour and shows every visitor the perfect products.” After 3 months, the test ran: AI recommendations against the existing rule “Show products from the same category”.
Result: the AI was 3% behind the rule on revenue. It often recommended products that looked visually similar but sat in a different price bracket. A customer looking at an 89-euro dress was recommended a 340-euro dress. The AI had not learnt price sensitivity, because there was not enough purchase data for that specific association.
The “same category” rule was dumber, but it was predictable. Customers looking at trainers saw more trainers. Relevance was high, even if the “intelligence” was low.
The simplest rule for product recommendations is: “Show products from the same category the customer last viewed.” This rule needs no machine learning. It needs a database query. And it works well enough in 70% of cases.
An electronics shop tested three variants:
Rule A: Same category, best-selling products. No personalisation.
Rule B: Same category, the customer’s most recently viewed products. Simple personalisation.
AI: A collaborative filtering algorithm that learns patterns from the buying behaviour of all customers.
The only case where the AI won significantly: a shop with 180,000 sessions a month and 12 months of data history. Here the AI was 12% ahead of rule B. Break-even was at around 80,000 sessions a month. Below that, the AI did not pay off.
AI-driven price optimisation sounds tempting: every customer sees the price they are willing to pay. In practice it is ethically questionable and technically risky. One shop tested dynamic prices based on device (mobile vs. desktop), time of day and traffic source.
Result: the conversion rate rose by 4%. But customer complaints rose by 340%. Customers who saw the price on their phone and then found the same item for 15% more on desktop felt cheated. The social media attention was devastating. The test was stopped after 10 days.
The lesson: AI can work technically, but the collateral damage outweighs it. Price transparency is trust. Trust converts better in the long run than dynamic price optimisation.
Most “AI personalisation” in emails stops at “Hello [first name]”. That is not AI. It is a placeholder. And it does not work. Test personalised emails (name, recently viewed products) against plain ones before you pay for the feature. Plenty of shops find no significant difference in conversion rate.
What worked was not AI but segmentation. A sporting goods shop divided its email list into 3 groups: new customers, regular customers, abandoners. Each group got a different email flow. No machine learning, no AI. Just clear rules based on buying behaviour.
Result: the segmented emails converted 23% better than the “personalised” ones. The personalisation was superficial (name + product). The segmentation was deep (customer lifecycle + need). Deep segmentation beats superficial personalisation, every time.
Most cost calculations for AI tools only include the licence fee. But there is more:
Setup costs: Integrating an AI tool into an existing shop takes 2–4 weeks. Developer costs: 8,000–15,000 euros. Most shops underestimate this effort by a factor of 3.
Training time: AI models need 3–6 months to deliver data-driven recommendations. During that time they make decisions based on assumptions, so no better than rules. The cost of this learning phase is rarely factored in.
Debugging effort: When a rule recommends the wrong product, you see immediately why: the category was mapped wrongly. When an AI recommends the wrong product, you do not know why. The model has learnt a weighting a human cannot follow. Debugging takes 10× more effort.
A cosmetics shop had an AI tool that suddenly started recommending men’s grooming products to female customers. The AI had learnt a pattern: customers who bought certain serums also bought men’s grooming products.
The explanation: these customers were also buying for their partners. The AI had not distinguished between “buying for yourself” and “buying for others”. The fix took 3 weeks, because nobody knew where in the model the misassignment lay.
Our workflow for personalisation looks like this:
Divide customers into 3–5 segments (new customer, regular, VIP, abandoner, win-back). Each segment gets different content. No machine learning. Just clear assignment rules.
If you have over 80,000 sessions a month and have exhausted the rules, AI can capture the last 10–15% of optimisation. But test it: AI against the best rule. Measure the ROI over 3 months. If the AI is not significantly better, stick with the rules.
A home & living shop followed this workflow. After 8 months it had implemented 12 rules, defined 4 segments and run an A/B test with an AI tool.
Result: the AI was 4% ahead of the rules, not statistically significant. The rules stayed. The AI was not bought. The 2,400 euros a month in licence fees went into A/B tests instead, with a measurably better ROI.
No. Below 50,000 sessions a month, rule-based systems are more efficient, cheaper and easier to maintain. The AI does not have enough data to learn patterns. The results are no better than rules, but much more expensive.
At least 3 months for the first usable results. 6–12 months for optimal performance. In the meantime, most AI tools fall back on rules, which is exactly what you would have without AI.
Yes, and that is the best approach. Use rules for the basic logic (categories, segments, bestsellers). Use AI only for complex patterns rules cannot capture (e.g. “Customers who bought A and B often also buy C, but only in the Christmas season”). AI should complement the rules, not replace them.
Licence: 1,000–5,000 euros a month for mid-market tools. Setup: 8,000–20,000 euros one-off. Maintenance: 20–40% of setup costs a year. Training phase: 3–6 months with no measurable ROI. Total costs over 12 months are typically 30,000–80,000 euros. Compare that with the potential revenue gain, and test rules first.
Not by the AI tool’s internal metrics (“recommendation click rate”, “engagement score”). Measure revenue per session, conversion rate and average order value, with and without AI. The only ROI that counts is revenue. Everything else is a vanity metric.
You will find more in-depth material in the Academy track Data-Driven Decision Making ↗.
Analytics
Google Analytics, Facebook Pixel, Hotjar, Klaviyo, a CRM system, an email tool: each one collects data, but hardly anyone knows which data matters for decisions. The result: dashboards full of numbers nobody reads.
Analytics
Web analytics delivers important information about a website’s performance. It lets you understand visitor behaviour and judge how well your marketing strategies work.
Analytics
Whether you have to switch to Google Analytics 4 is no longer a question. Universal Analytics was switched off in July 2023.
In the Dropoff Analysis you see where your shop loses buyers. Afterwards you know where the first step pays off.