Should Shopify product recommendations optimize for conversion rate or average order value?
A recommendation can raise conversion rate and still be bad for the store.
Say a shopper is looking at a $120 jacket. Showing a cheaper $35 accessory may get an extra click or purchase. Showing a matching $60 scarf may make the order larger. Showing another jacket may help the shopper choose when they were undecided.
Those are different jobs. A single target cannot describe all of them.
For Affina, our pre-launch product recommendation app, this is a central design question. It will place AI recommendations on product and cart pages, learn from each store's own catalogue and customer behaviour, and measure its choices against a real 10% holdout. That means a portion of eligible traffic will not see Affina's recommendations. It gives you a baseline from the same period, on the same store.
The short answer is: don't optimize recommendations for conversion rate or average order value alone. Optimize for the value created per visit, then inspect conversion rate and AOV to understand why it moved.
Conversion rate can favour the smaller order
Conversion rate answers a useful question: of the people who saw a recommendation, how many went on to buy?
It matters. If recommendations make shoppers hesitate, distract them, or send them to poor substitutes, conversion can fall. That is a real cost.
But conversion rate has a blind spot. It does not tell you how much was spent.
Imagine two groups of 100 visitors:
- Group A has 4 orders at $50 each.
- Group B has 3 orders at $90 each.
Group A converts at 4%. Group B converts at 3%.
If you only optimize for conversion rate, Group A wins. But Group B produces $270 in revenue, compared with $200. Revenue per visit is $2.70 for Group B and $2.00 for Group A.
That does not mean lower conversion is always acceptable. It means a higher conversion rate is not automatically better.
Shopify makes the same practical point in its conversion-rate guidance: conversion varies a great deal by price and category, and AOV should be reviewed beside conversion rate rather than treated as an afterthought. Shopify's recommendation guidance also recommends looking at click-through rate, conversion rate, and AOV together.
AOV can favour the order that never happens
Average order value has the opposite blind spot.
A recommendation system can raise AOV by putting expensive products in front of shoppers. But if those products are a poor fit, more people may leave without buying anything.
Take another 100 visitors:
- Group A has 5 orders at $60 each.
- Group B has 2 orders at $120 each.
Group B has the higher AOV: $120 instead of $60. Yet it creates less revenue overall: $240 instead of $300.
This is why “show the most expensive item” is not an AOV strategy. It is just a rule. It may be useful in a narrow situation. It may also turn a straightforward add-on into a harder decision.
Before making AOV a target, check how your own reports define it. Shopify has published a page that calls AOV a median in one place, while also showing the standard total revenue divided by total orders formula. That formula produces a mean average, not a median. The label in your reporting matters when you compare tests.
You can check this now:
- Open your Shopify analytics reports.
- Find the AOV metric you use for store decisions.
- Read the metric description or formula.
- Make sure everyone on your team means the same thing by “average.”
Revenue per visit joins the two numbers
Revenue per visit, often shortened to RPV, puts conversion rate and AOV in one measure.
The basic calculation is:
`revenue ÷ visits`
It can also be thought of as:
`conversion rate × average order value`
A peer-reviewed framework for recommendation testing uses revenue per visit as its main outcome for this reason. A system can improve RPV by raising conversion rate, raising AOV, or improving both.
That is a more honest goal for recommendations. A shopper needs to buy, and the order needs to have value.
It also gives you a clearer way to read a test:
- RPV rises, conversion rises, AOV is flat: recommendations likely helped shoppers find something worth buying.
- RPV rises, conversion is flat, AOV rises: recommendations likely added useful items or guided shoppers to a better-fit product.
- AOV rises but RPV falls: bigger orders did not make up for lost orders.
- Conversion rises but RPV falls: recommendations may be steering shoppers into lower-value purchases.
- Click rate rises but purchases do not: the recommendation was interesting, but not useful enough to finish the order.
Clicks are a diagnostic, not the finish line. Shopify's native recommendation reporting includes click rate and purchase rate, and its standard reports use the last 30 days of data. Those numbers can help you spot what changed. They do not define a universal AOV optimization setting.
The right objective changes by recommendation type
Product recommendations are not all trying to do the same thing.
Shopify separates them into two broad types:
- Related products are similar or substitutable items. A shopper viewing one black dress may be shown other dresses.
- Complementary products are add-ons. A shopper viewing a camera may be shown a memory card or case.
Shopify automatically generates related recommendations. Complementary ones need to be configured manually.
That distinction matters for the metric you expect to move.
A related recommendation can protect conversion when the product a shopper first chose is not quite right. Perhaps the size is unavailable, the colour is wrong, or the shopper wants a cheaper option. In that case, conversion rate deserves close attention.
A complementary recommendation has a different job. It asks whether something makes the original purchase more complete. In that case, AOV may be more likely to move.
But the scorecard should still return to RPV. A complementary item that lifts AOV but causes fewer completed orders has not necessarily helped.
On a cart page, this becomes especially important. A shopper has already built an order. The recommendation needs to earn its space. A compatible refill, a gift box, or a product needed to use the item can be helpful. A random premium product is more likely to interrupt the path to checkout.
Why a real holdout matters
You cannot answer this question by comparing this month with last month.
Traffic sources change. Stock changes. Promotions end. A holiday arrives. Your best-selling product goes out of stock.
A holdout gives you a comparison group at the same time. With Affina, the planned 10% holdout means that eligible shoppers in that group provide a check against the experience without Affina recommendations.
The useful comparison is not just “did recommendation clicks increase?” It is:
- Did revenue per visit differ between recommendation traffic and holdout traffic?
- Did conversion rate change?
- Did AOV change?
- Did one metric improve by sacrificing another?
The holdout does not make a small change meaningful by itself. It does stop you from giving all credit to the recommender when the whole store had a stronger week.
Choose a business outcome, then keep the guardrails visible
For most stores, sales value per visit is the sensible main outcome. Conversion rate and AOV should stay beside it as guardrails and explanations.
There is one further question only you can answer: are all orders equally valuable?
If one category has low margins, costly returns, or tight stock, revenue alone may not tell the whole story. That does not make RPV wrong. It means you should review recommendation choices against the realities of your catalogue before declaring a winner.
Start with a simple scorecard:
- Revenue per visit: the primary result.
- Conversion rate: did recommendations help or hinder purchase completion?
- AOV: did they add useful value to orders?
- Click rate: did shoppers engage with the placement?
- Product-level review: are the recommended items sensible for the store?
Affina is not live yet, so there is nothing to install today. If you are considering it for launch, start by writing down the baseline RPV, conversion rate, and AOV you would want a holdout test to compare.