An AI chatbot for ecommerce sells only when it behaves like a good shop assistant: it asks one question before it shows anything, it says "we don't have that" instead of offering the closest match, and it takes the customer's contact when it cannot close. Everything else, speed, tone, product knowledge, is table stakes. Most ecommerce chatbots have the table stakes and none of the three behaviours, which is why they generate clicks and no orders.
Key facts
- Where this comes from: 238 exported conversations from a paid AI chatbot in an online bathroom-equipment store, July to October 2026, then a replacement built on a written sales method
- The baseline: 108 product clicks, 2 carts, 0 purchases attributed to the chat
- The core defect: the chat asked a question first in 7 of 379 customer inputs
- What a selling chat does: one question before the list, no closest match, take the contact, short answers in the customer's language
- Who this is for: store owners and ecommerce managers choosing or configuring an AI chatbot
What an ecommerce chatbot is supposed to do
The vendor pages list the benefits of a chatbot in ecommerce: instant answers, 24/7 availability, lower support cost, product recommendations, cart recovery. All of that is true and none of it is the job.
The job is the same as a shop assistant's: help a customer who does not know enough to choose, so that they decide and buy the right thing. A customer who already knows exactly what they want does not need a chat. They need a search box. The chat exists for everyone else.
Measured against that job, most ecommerce chatbots fail quietly. They answer every question, correctly and fast, and the customer leaves without buying because nobody helped them choose. I have the figures for one such chatbot in a case study of 238 conversations. The seven practices below are what we changed, and why.
1. Ask one question before you show products
The customer types "trap". The typical chatbot returns nine traps with prices. A shop assistant asks: "For a washbasin, kitchen sink, shower tray or washing machine?" Then shows two.
That single question is the difference between a search engine and a salesperson. It is also where almost every chatbot fails, because the default behaviour of a language model with a product catalogue is to retrieve and list. In the store we measured, the chat opened with a question 7 times in 379 inputs. It opened with a priced list 248 times.
The rule we gave the replacement: for any product group where the choice depends on something the customer has not said, ask that one thing first. Not three things. One. Then recommend two or three products. If the customer says "just something that works", skip the question and show the two cheapest.
This is the same principle I describe in the book Diagnostic Sales™ as the difference between a price dispenser and a salesperson. The dispenser never asks why.
2. Never offer the closest match
A customer asked for a 160 mm transition piece. The store did not have one. The chat offered a backflow valve for 36,780 dinars, because the valve's name contained "DN160". Three times, to the same customer.
Language models are built to find the nearest thing. In sales, the nearest thing is a lie. A customer who knows what they need spots the substitution instantly and stops trusting every other answer. A customer who does not know buys the wrong part and returns it.
The rule: if the code, size or brand is not in the catalogue, say so in one sentence and ask what the part is for. "We don't have a 160 mm transition piece. What are you connecting, and to what?" That is a diagnosis. Offering the valve is a guess with a price tag.
3. Take the contact instead of sending the customer away
Of 384 chatbot messages in our export, 123 ended with "call us" or "email us". Not once did the chat say "leave your number and we will call you". A quarter of the conversations happened between 10 pm and 7 am, when there was nobody to call.
A customer who is told to call will mostly not call. A customer who leaves a number can be called at 8 am by someone who knows the stock. The second customer buys. The first one is a lost lead nobody ever counted.
The rule: whenever the chat cannot confirm a detail, or the customer signals they are ready to buy, ask for a phone number or email and promise a call within working hours. At night, always.
4. Answer in the customer's language, in three sentences
Twenty-seven answers in English to questions asked in Serbian, usually when the input was a product code. Replies of 600 to 1,400 characters to customers on a phone. The same word typed twice, the identical list returned twice.
Short is not a style choice. Half of ecommerce chat traffic is mobile, and a 1,200-character reply is four screens of scrolling before the customer sees a price. The rule: reply in the language the customer wrote in, keep it to three or four sentences, and never repeat a list you already gave.
5. Be honest about what you do not know, and stop there
The chatbot we studied did one thing well: when it did not know an outlet diameter, a decibel rating or whether a drawing existed, it said so and referred the customer onward. That is correct behaviour. It then undermined itself in other answers by explaining a valve type under the wrong name, describing one manufacturer as Serbian, Slovak and Slovenian in three conversations, and presenting a quiet fan model as the turbo version before retracting.
None of the errors was large. Together they cost trust, which is the only thing a chat has. The rule: the knowledge base must carry the facts customers ask about most, with the salesperson's corrections, and the chat must be told that "I don't know, let me get someone to confirm" is a good answer.
6. Put the salesperson's knowledge in, group by group
The experienced salesperson in that shop knows that "I need a bath tap, just something that works" gets the two cheapest models immediately, while "trap" gets a question first. He knows which toilet seats come back because of hinge spacing and which fans are bought for the wrong room. None of this was in the chatbot, because nobody wrote it down.
An AI chatbot for ecommerce amplifies what it is given. Give it a catalogue and it multiplies the catalogue. Give it a sales method and it multiplies the method. The work is not in the tool. It is in sitting with your best salesperson and writing, for each product group: what question separates the models, what customers get wrong, what comes back, and when to stop selling. We cover how that knowledge moves from a person into a system on the AI integration page.
7. Measure purchases and captured contacts, not clicks
The dashboard showed 108 product clicks. It looked like work. Purchases attributed to the chat were zero. Clicks measure curiosity. Purchases and captured contacts measure whether the chat helped anyone decide.
Set two numbers before you switch a chatbot on: orders where the chat was the last touch before checkout, and contacts captured when the chat could not close. If both are zero after a month, the chat is a search box with good manners, whatever the vendor's conversion-rate statistics say.
Ecommerce chatbot examples: before and after
The same catalogue, the same site, two different chats:
- Customer types: "trap". Chatbot without a method: Nine traps, prices, 900 characters. Chatbot with a sales method: "For a washbasin, sink, shower tray or washing machine?"
- Customer types: "transition piece 160". Chatbot without a method: A backflow valve at 36,780 dinars. Chatbot with a sales method: "We don't have 160 mm. What are you connecting?"
- Customer types: "I'll take the 60 cm cabinet" at 11 pm. Chatbot without a method: "Call us during working hours". Chatbot with a sales method: "Leave your number, we'll confirm stock and call you at 8."
- Customer types: "HL134" (a product code). Chatbot without a method: Answer in English, 1,100 characters. Chatbot with a sales method: Two sentences in Serbian, one link
The second column is not a better model. It is the same class of model with three rules and a salesperson's notes.
Should you buy an AI chatbot for your ecommerce store?
Yes, if you can first write down how your best salesperson runs a conversation from the first word to the customer's decision. The chat will follow that. If you cannot write it down, buy nothing yet. The chat will multiply your catalogue, you will get clicks, and you will read a case study like ours about your own store.
For the mistakes in detail, see why chatbots fail to sell. If your store runs on Shopify, there is a checklist of what to verify before installing a chatbot app.