How to Measure Whether Product Recommendations Are Working
Track who could see the recommendation, what they selected and what they actually bought. Compare revenue and contribution across the eligible audience.
Measure whether product recommendations are working by connecting eligible exposure, product selection, cart additions and completed purchases, then comparing the result with an experience without those recommendations. Track contribution and returns alongside revenue. Clicks and higher AOV alone cannot show that the recommendations caused an improvement.
For Niwa recommendations, test the complete shopping sequence: the customer asks a question, sees an actual catalog item, chooses a variant and continues through a supported cart action. The measurement system must distinguish that product interaction from a chat message. Niwa's storefront workflow does not make every conversation a sale.
Define the audience before the result
Decide which visitors are eligible for the recommendation. For a camera accessory, that might be sessions containing a qualifying camera in the cart. For a product comparison, it may be visitors to a specific product group.
Keep the eligibility rule the same for the original and test experiences. Choose a stable counting unit, such as a session or a visitor, and use it consistently. Do not compare accepted recommendation orders with all store visits and describe that as a conversion rate for the offer.
Record the placement and product relationship being tested. A recommendation on a product page and one inside an AI conversation may reach different people. Separate them before combining the numbers.
Follow the product, not just the widget
For each placement, record the relevant product ID, variant, quantity, price and a list identifier. Establish whether the item was already in the cart so an existing purchase is not miscounted as a new addition.
Track the sequence from display to selection, addition and purchase. Include removal or refund where your measurement setup supports it. A shopper can click a suggested accessory, inspect it and decide it is unsuitable. That click is useful interaction data, not revenue.
Treat recommendation exposure carefully. A panel present in the page's HTML may never be seen. If you use a viewability rule, define it and apply it consistently. Keep that measured exposure separate from experiment assignment, which should include eligible people whether they engage or not.
Use ecommerce events where GA4 is installed
Google's ecommerce measurement documentation defines events such as view_item_list, select_item, add_to_cart, purchase and refund. List identifiers and item details help distinguish one recommendation placement from another.
Preserve consistent item IDs and variant information through the journey. Supply the correct currency with value data and use a stable transaction ID for the completed purchase. Verify that a thank-you-page reload does not become another recorded sale.
For an AI product card, the tracking integration needs to observe the actual card and shopping action. A message saying an item was added does not establish an add_to_cart event, and a click on checkout does not establish a purchase event.
These are implementation requirements for your analytics setup, not a claim that every Niwa installation already sends these events. Validate them with a safe test purchase and reconcile it with the store's order record. Keep message contents and personal contact details out of event labels; use product and placement information.
Read the numbers as a connected set
Use selection rate to understand interest: eligible recommendation displays with a selection divided by eligible displays, using your defined counting unit. Use attachment rate to understand purchases: completed qualifying orders containing the suggested item divided by completed qualifying orders.
For commercial evaluation, compare completed-order conversion, net sales per eligible session or visitor, AOV, contribution and returns. Declare the denominators in the report so another person can reproduce the result.
Here is a hypothetical example:
- Original experience: 1,000 eligible sessions, 60 orders and $3,000 net sales. AOV is $50; revenue per eligible session is $3.
- Recommendation experience: 1,000 eligible sessions, 50 orders and $2,750 net sales. AOV is $55; revenue per eligible session is $2.75.
The average order rose, but total revenue from the same eligible traffic fell. You would still need product and fulfillment costs to compare contribution. This is why optimizing AOV alone can lead to the wrong decision.
Distinguish attribution from added value
An order containing a recommended item is attributed activity, not automatic proof of an incremental sale. Some customers would have bought that item without the suggestion. Others may replace a full-price item with a discounted alternative.
A controlled experiment with stable assignment gives you a stronger basis for comparison. Set the decision metrics and guardrails before looking for a winner. Account for sufficient observations, changing traffic and returns; one large order can distort a small sample.
If a controlled test is unavailable, compare similar periods and segments and label the result observational. Customers who choose to talk to an assistant may already have different needs or purchase intent from customers who do not. Do not call that self-selected comparison proof that AI caused the difference.
Investigate failure at the point it occurs
Few selections can mean poor placement or an irrelevant product. Additions without purchases can point to price, shipping or checkout friction. Increased returns can indicate unsuitable compatibility advice. Check the product and order evidence before changing several things at once.
Use WooCommerce Analytics for order and revenue reporting, then combine it with the recommendation events your installation actually records. Niwa can help examine available business and conversation information, but confirm every conclusion against its source data. See how Niwa uses the site's business context.
Contact Niwa with your store and the recommendation journey you want to verify. A concrete product, placement and completed-order check gives the setup something measurable to support.
