Ecommerce

AI product recommendations that don't annoy the customer

"Customers also bought" has sat under every product page for years, and most shoppers have learned to look straight past it. The difference between a recommendation that works and one that gets ignored is not the algorithm. It is whether anyone asked what the customer is actually doing.

By Henri SaloPublished Jul 13, 20267 min read

A shopkeeper's hand placing one item forward on a shelf

In a physical shop, a good assistant asks one question before recommending anything: what are you going to use it for? Everything useful follows from the answer.

Online, nobody asks. AI product recommendations infer from behavior (what you viewed, what shoppers like you bought) and print a row of products under the fold. It is a reasonable guess made without the one piece of information that would make it a good one.

Why most AI product recommendations get ignored

Three structural reasons, none of which better math fixes.

It guesses intent from clicks

Viewing three drills tells the system you are interested in drills. It cannot separate the builder replacing a worn tool, the tenant hanging one shelf and the person buying a gift. Three budgets, three products, one identical click stream.

It appears where attention has already moved on

Recommendation rows sit below the product description, at the point where the visitor has already decided or already left. The moment of real uncertainty came earlier, halfway down the specification, while the visitor was still working out whether the product fits.

It never explains itself

"You might also like" gives no reason, and shoppers filter reasonless suggestions without noticing they did it. Compare that with one sentence: "this one has the higher torque you want for masonry". Same product, entirely different response, because there is a because.

The pattern: the widget infers intent, arrives late, and cannot justify itself. A conversation asks intent, arrives when the question forms, and always has a reason.

What conversational recommendation does differently

It asks. That is nearly the whole distinction, and it changes three things.

  1. Intent is stated, not guessed. "I need something for a small bathroom, and I'm renting" narrows a catalog faster than any behavioral model. Within its first week of conversations, Ebike Center heard it stated: a large share of shoppers were looking for a bike under 2,000 euros. One shop, one first week, and the picture varies with the catalog.
  2. The recommendation carries a reason. A short list of two to five options, one sentence each on why. The reason is what converts, and it is also what makes a wrong recommendation forgivable: the customer sees the logic and corrects it.
  3. It can say no. "We don't have anything suitable for that" is a genuinely valuable answer. It prevents a return and buys credibility for every other thing the bot says.

Three rules that keep it from becoming annoying

Recommend a short list, never a wall

A few options with reasons is advice. A wall of twelve tiles is a search result with extra steps, and it hands the work back to the customer you were trying to help. Our own shopping assistant recommends two to five products, and each one arrives with the reason it was picked.

Do not always steer upward

A bot that recommends the expensive option every time gets identified within two questions, and everything it says afterwards is discounted. Recommending the cheaper product when it genuinely fits is what makes the expensive recommendation believable when it comes.

Refuse when the data is thin

Compatibility is the clearest case. If the fitment data is incomplete, the bot must say so rather than infer. A confident wrong answer about whether a part fits produces a return, a support ticket and a lost customer: three costs against one saved conversation.

Measuring whether it earned anything

Most chatbot reporting counts conversations and resolution rates. Neither answers the question finance will ask: did it bring in money?

MetricWhat it tells you
Assisted revenue Value of orders where the customer talked to the assistant before buying. The number a finance team asks for. It lives in your store analytics, next to the order data. Revdash does not report revenue: it shows resolution rate, leads, top questions and unanswered questions.
Recommended → purchased How often a suggested product reached the basket. Distinguishes recommending from guessing.
Average order value, with and without Whether guidance moved basket size at all.
Return rate on assisted orders The check most people skip. If assisted orders come back more often, the bot is selling the wrong thing well.
Questions with no answer A direct list of the gaps in your product data. Revdash collects these for you.

Read assisted revenue honestly. It means the customer talked to the assistant and then bought, not that the assistant caused the purchase. Some of those people would have bought anyway. Use it as an order-of-magnitude figure and, where you can, compare it against a period without the assistant.

What has to be true before it works

Conversational recommendation is only as good as the product data behind it. Three prerequisites:

  • A product feed with real attributes. Names and prices are not enough. The attributes that decide suitability (size, compatibility, material, use case) have to exist somewhere machine-readable.
  • Accurate stock data. Recommending what cannot be bought is worse than not recommending. We refresh product data once a day, so a stock error you fix on Monday afternoon reaches the bot on Tuesday.
  • Descriptions written for humans. If your product copy is a specification dump, the assistant has nothing to reason from. This is the most common blocker and the least technical one.

Frequently asked

How do AI product recommendations work in ecommerce?+

Behavioural engines infer interest from what a shopper viewed and what similar shoppers bought, then display a row of products. Conversational recommendation instead asks what the customer intends to do, reads the product feed, and proposes two to five options with a reason for each. The difference is stated intent versus inferred intent.

Do AI recommendations increase average order value?+

They can, but the honest measure is assisted revenue compared against a period without the assistant, alongside the return rate on assisted orders. A recommendation engine that raises basket size while also raising returns has not helped. Treat single-number uplift claims from vendors with caution, including ours.

What data does an AI shopping assistant need?+

A product feed with genuine attributes (size, compatibility, material, intended use) plus accurate stock data and product descriptions written for people rather than as specification dumps. Thin product data is the usual reason these projects disappoint, and no model compensates for it.

Does it work with Shopify and WooCommerce?+

Yes. Revdot's shopping assistant has ready-made integrations for Shopify and WooCommerce, and for the Nordic platforms MyCashflow and Shuriken, and on the Business tier it can also read products from your own database. On any platform, structured product data gives better answers on compatibility and sizing than product pages alone.

Will customers find it intrusive?+

They find it intrusive when it interrupts, offers a wall of options, or always pushes the expensive item. They do not find it intrusive when it waits to be asked, offers a short list with reasons, and is willing to say nothing suitable exists.

Revdot's shopping assistant does this job: it asks what the customer needs, reads your product data, and recommends two to five products with a reason for each. It works with Shopify, WooCommerce, MyCashflow and Shuriken, and it is included in Business (149 € a month) and Scale, on the same subscription as the chat widget, inline bot and lead form. In Revdash you see the resolution rate, the leads, the most asked questions and the ones the assistant could not answer.

Try it on your site → See pricing

Related reading: RAG explained without the jargon, on how an assistant reads your own data before it answers.

Try it on your site

Your own bot tells you more than any article.

Enter your site address, and Revdash builds a bot from your content and emails you the link in 10 minutes. Prices are public, from 39 € a month.

What is your main goal?

Revdash reads your site · knowledge base · bot in your inbox

The Revdot team answers in person:WhatsAppBook a call+358 45 231 1420eelis@revdot.ai