Written with AI assistance and reviewed by the NorwegianSpark SA editorial team.
Last updated: September 2026
"AI shopping assistant" is one label sold over at least three different products, and most of the frustration in this category comes from buying one of them while expecting another. This article separates them, sets out how to work out which one your shop actually needs, and — the part usually missing — how to tell afterwards whether it did anything, because the default way these things are measured guarantees a flattering answer.
For the wider stack around running a shop, our guide to AI tools for ecommerce covers operations, merchandising and back-office work. This page stays on the thing the shopper actually talks to.
The Three Products Sold Under One Name
Conversational search. Replaces or supplements the search box. The shopper types something in ordinary language — "a warm coat for a toddler that isn't bulky" — and gets products rather than a zero-results page. The job is understanding intent that keyword search cannot parse.
Recommendation and merchandising. Decides what to show on the homepage, the product page and the basket. Mostly invisible to the shopper. The job is ordering a catalogue that is too large to browse.
Support deflection. A chat widget answering "where is my order", "can I return this", "does it ship to Ireland". The job is reducing the number of tickets that reach a human.
These solve different problems, are bought by different departments and are measured with different numbers. A shop with excellent search and a support inbox full of delivery questions does not need conversational search, no matter how good the demo was. Naming which of the three you are buying is the first and most useful decision in this whole process.
Working Out Which One You Actually Need
The evidence is already in your own systems, and it is more reliable than any vendor's diagnosis.
- Read your site search logs. Specifically the queries that returned nothing. If people are searching in sentences, or for attributes your catalogue does not have as filters, conversational search has a real job. If they are searching for product names and finding them, it does not.
- Read a hundred support tickets. Sort them into "answerable from data we already hold" and "needs a human". The first pile is the honest size of the deflection opportunity, and it is usually smaller than the vendor's estimate and larger than yours.
- Look at where people leave. Exits from category pages point at a discovery problem. Exits from the basket point at pricing, shipping or trust, and no assistant fixes those.
- Count your catalogue. Recommendation engines need a catalogue big enough that browsing fails. Below a few hundred products, good navigation usually beats a model.
The Measurement Problem, Which Is the Whole Problem
This is the part that decides whether you can ever tell if the thing worked, and it is worth being blunt about.
The standard vendor dashboard reports "revenue influenced by the assistant" — the value of orders where the shopper interacted with it at some point. That number will always look excellent, and it is close to meaningless, because engaged shoppers are the ones who buy. People who were already going to purchase are more likely to use every feature on the page, including the assistant. The metric measures purchase intent and reports it as impact.
The only honest measurement is a holdout: some proportion of visitors do not see the assistant at all, and you compare conversion between the two groups over a period long enough to be meaningful. That is a harder thing to set up and a much less flattering number, which is why it is rarely the default. Ask whether the platform supports it before you buy, because retrofitting a holdout after the assistant is embedded everywhere is considerably more work.
Two secondary measures worth watching, both of which can move in the wrong direction:
- Return rate. An assistant that is persuasive but inaccurate raises sales and raises returns. Net of returns and return shipping, that can be a loss reported as a win.
- Contacts per order, not deflection rate. A widget that answers badly and pushes the shopper to email you afterwards records a deflection and creates a ticket.
Where These Fail, and the One Failure That Is Serious
Most failure modes here are cosmetic. One is not.
A support assistant that states a policy you do not have — a returns window, a delivery guarantee, a price match, an eligibility rule — has made a statement to a customer on your behalf, in writing, on your own site. Whether that binds you is a question for your own legal advice and varies by jurisdiction, and we are not going to guess at it. What is not in doubt is that it is a customer-service problem you will have to resolve one way or the other, and it scales with traffic.
The mitigation is structural rather than clever: the assistant should answer from your actual policy documents and order data, and it should say "I don't know, here is a human" rather than improvise. When you evaluate one, spend most of your time trying to make it say something untrue. Ask about a policy you do not have. Ask about stock for a discontinued item. Ask in a language you do not support. The demo will show you the happy path; the failure path is what you are buying.
The quieter failures are worth listing too. Conversational search can be slower than typing two words into a normal search box, and slower is worse. A widget that covers the buy button on a phone costs you money directly. And a chat interface that is not keyboard-navigable or screen-reader-compatible excludes customers, which is both a commercial loss and, in many jurisdictions, a compliance question for your own advisers.
Your Product Data Is the Ceiling
The most consistent finding in this category is unglamorous: these systems are limited by the structured data behind them, not by the model in front of them.
An assistant cannot filter on a material, a fit, a compatibility or a use case that is not recorded anywhere in your catalogue. If the information lives only in a paragraph of marketing prose, or only in an image, it may be partially extractable and it will be partially wrong. If a third of your products are missing the attribute a shopper asked about, a third of the answers are guesses.
Which means the highest-return work is frequently not buying an assistant at all — it is completing your attributes, standardising your category names, and fixing the products with a two-line description. That work also improves your normal search, your filters and your product pages, all of which keep paying whether or not you ever add a conversational layer. Our piece on writing product descriptions at scale covers the copy half of the same job.
What to Check Before You Buy
- Can you run a genuine holdout test, and does the platform support it natively?
- Does it answer from your live catalogue and order data, or from a snapshot that goes stale?
- What does it do when it does not know — does it escalate, or does it improvise?
- How much does it add to page weight, and what does that do to load time on a mid-range phone on mobile data?
- Is the widget keyboard-accessible and screen-reader-compatible?
- Where does shopper conversation data go, how long is it kept, and is it used to train anyone's models?
- What is the pricing unit — conversations, resolutions, sessions, revenue share — and what happens on your busiest day of the year?
- If you remove it, what breaks? An assistant that has quietly become your only search is a dependency.
Platform First, Assistant Second
Most of these are apps on top of a commerce platform rather than standalone products, so the platform decision constrains the assistant decision. Shopify is the widest ecosystem here, which matters practically: it means more options, more competition on price, and a documented way to get your product and order data out again. Zebao works the ecommerce AI side, and for shops where the real gap is following up with the customers you already have rather than converting new ones, ClearCRM is aimed at small teams with follow-up automation built in.
The Honest Counter-Argument
For a large number of shops, the correct answer is no assistant at all, and it is worth stating that clearly on a page that carries affiliate links.
Fast pages, clear photography, honest delivery estimates, a search box that handles plurals and typos, and filters that match how people actually shop will move conversion further than a conversational layer on top of a catalogue that has none of those things. A shopping assistant is an amplifier: it makes a well-structured catalogue easier to navigate, and it makes a badly structured one confidently wrong. The order of operations matters, and the boring work comes first.
Disclosure: this article contains affiliate links. If you sign up through them we may earn a commission at no extra cost to you. It does not change what we recommend, and no vendor has paid for a mention.
Related Articles
