What a recommendation actually is
A recommendation answers one question: given what this customer has bought and looked at, what else are they likely to want?
It is not clairvoyance. It is a pattern — customers who bought this also bought that, or customers like this one tend to buy this next.
Done well it is genuinely useful to the customer. Done badly it is the shop suggesting a second identical item to someone who just bought one, which makes you look like you are not paying attention.
Where they appear
Four places, with different economics.
On a product page. “Customers also bought” or “goes well with this”. The most familiar and usually the safest.
At checkout. Suggesting an add-on before the order completes. Effective, and the place where getting it wrong costs most — a badly timed suggestion adds friction at the moment you least want it.
In email. Based on purchase history. Works well for consumables and reorders.
To a salesperson. Suggesting what to discuss on the next call. Underused in B2B and often the highest-value version, because a person filters it before the customer hears it.
FIGURE 1: WHERE RECOMMENDATIONS EARN THEIR PLACE
On the product page
- Complementary items. Familiar, safe, and the customer is already browsing.
In email
- Reorder reminders and consumables. Works because the timing is predictable.
To a salesperson
- A prompt before a call. A person filters it, so a poor suggestion costs nothing.
The kinds
Three approaches, and knowing which you are using explains its weaknesses.
Bought-together. Customers who bought A also bought B. Simple, effective, and needs volume — a few hundred orders will surface coincidences rather than patterns.
Similar customer. Customers resembling this one bought B. Better for larger catalogues, and it needs enough customers to find resemblance.
Complementary by attribute. Products that go with this one, defined by their characteristics or by rules you set.
That last one deserves attention. For many businesses, a well-maintained list of “these go together”, set by someone who knows the products, outperforms a model trained on thin data.
If you have fewer than a few thousand orders, do that instead. It is not less sophisticated in any way that matters. It is more accurate.
Where they work best
Consumables and repeat purchases. Someone who buys filters every three months is highly predictable, and a well-timed reminder is a service rather than a sale.
Genuine complements. A printer and its cartridges. A machine and its spare parts. These relationships are real and stable.
Large catalogues. Where customers cannot reasonably browse everything, recommendation is navigation.
Where they do not
Considered, infrequent purchases. Somebody buying one significant item every three years has no pattern to learn from.
Small catalogues. If you sell forty products, customers can see them all. Recommendation adds nothing.
Thin data. A few hundred orders produces suggestions based on noise, presented with the same confidence as good ones.
FIGURE 2: WHEN TO USE A MODEL AND WHEN TO USE RULES
A model works when
- Thousands of orders in the history
- Repeat and consumable purchases
- A catalogue too large to browse
- Stable product relationships
Rules work better when
- A few hundred orders
- Infrequent, considered purchases
- A small catalogue
- You know the pairings better than the data does
The mistakes that annoy customers
Five, and all of them are avoidable.
Suggesting what they just bought. A second washing machine to someone who bought one yesterday. It is the most common failure and it makes you look inattentive.
Suggesting out-of-stock items. In Odoo this is easily avoided — the recommendation and the stock are the same database. There is no excuse for it.
Suggesting something more expensive with no reason given. If you are recommending an upgrade, say what it does differently.
Too many suggestions. Three relevant items beat twelve. A long list signals that you do not know which one fits.
Ignoring the obvious. If a customer bought a device that needs a specific consumable, suggest the consumable. That does not need a model.
The advantage in Odoo
Because it is one database, a recommendation can use what a standalone shop engine cannot.
Actual purchase history, not just website behaviour.
Real stock, so nothing out of stock is suggested.
Full customer context — their tier, their pricing, their credit position, whether they have an open complaint.
That last point matters more than it sounds. Recommending an upsell to a customer with an unresolved support ticket is a bad look, and a system that can see both avoids it.
Measuring it
Four numbers, and one caution.
Attach rate. How often a suggestion is added to an order.
Average order value. Should rise if it is working.
Return rate on recommended items. Watch this. If it rises, you are successfully selling people things they did not want, which is worse than selling nothing.
Unsubscribes on recommendation emails. Rising means the suggestions are wrong or too frequent.
The caution: attach rate alone is a bad measure. A high attach rate with a rising return rate is a loss dressed up as a win.
FIGURE 3: A SENSIBLE WAY TO START
Start with rules you know
- Obvious pairings, set by someone who knows the products. Often better than a thin model.
Exclude what is out of stock
- One database means there is no excuse for suggesting the unavailable.
Show three, not twelve
- A long list says you do not know which one fits.
Watch returns, not just attach rate
- Selling people what they did not want is a loss with good statistics.
Where to begin
1. Fix your product data. Categories, attributes and relationships. Recommendations are only as good as the structure underneath them.
2. Start with rules. Set obvious pairings manually. Measure the effect. This is often most of the available value.
3. Add automatic recommendations once you have enough order history to support them.
4. Exclude out-of-stock items always.
5. Limit to three suggestions.
6. Watch returns as closely as attach rate.
The short version
Recommendations work where purchases repeat, complements are real, and you have enough history for patterns to be genuine rather than coincidental.
For many businesses, hand-set rules outperform a model. That is not a lesser option — with a few hundred orders, it is the more accurate one.
Never suggest what is out of stock, or what they just bought. And measure returns alongside attach rate, because selling people things they send back is not a win.
Wondering whether recommendations would work for your catalogue?
Get in touch. We will look at your order volume and product relationships, and tell you honestly whether a model or a set of rules will serve you better.
