Chatbots fail to sell because they are configured to answer, not to help a customer decide. The six mistakes below appeared in almost every one of 238 conversations we exported from a paid AI chatbot before replacing it, and each has a fix that has nothing to do with the model. The model was fine. The sales method was missing.
Key facts
- Source: 238 conversations, 379 customer inputs, 384 chatbot replies, one online store, four months
- Result before the fix: 108 clicks, 2 carts, 0 attributed purchases
- The six mistakes: lists instead of questions, closest-match substitution, sending customers to the phone, long answers in the wrong language, confident small errors, measuring clicks
- Common cause: nobody wrote down how a salesperson in that store runs a conversation
- Who this is for: anyone who installed a chatbot, saw the usage numbers, and cannot find the sales
Mistake 1: it lists instead of asking
The customer types two words. The chatbot returns five to ten products with prices and a paragraph on how to choose. The customer clicks one. The conversation ends.
In our export the chat opened with a question 7 times in 379 inputs. It opened with a priced list 248 times. For "toilet seat", six different customers received the same list of ten seats. None was asked about bowl shape or hinge spacing, which is the only thing that decides whether a seat fits.
What it costs: the customer is left alone with a choice they came to the chat to avoid making alone. They click at random or leave.
The fix: one qualifying question before any product appears, for every product group where the right answer depends on something the customer has not said. Then two or three products, not ten. This is the first of the seven best practices for an ecommerce chatbot, and the one that matters most.
Mistake 2: it offers the closest match
The store had no 160 mm transition piece. The chat offered a 36,780-dinar backflow valve because its name contained "DN160". For a product code that did not exist, it offered a similar code of a different product.
Language models are built to return the nearest thing. In a shop, the nearest thing is a wrong part sold with confidence.
What it costs: trust. A customer who knows the part sees the substitution and discounts every other answer. A customer who does not know buys it and returns it.
The fix: "We don't have that" in one sentence, then a question about purpose. The rule has to be explicit in the chat's instructions, because the default is to substitute.
Mistake 3: it sends the customer to the phone
Of 384 replies, 123 ended with "call us" or "email us". Zero offered "leave your number and we'll call you". A quarter of the conversations happened between 10 pm and 7 am. Customers who said they wanted a specific cabinet, seat or cistern lid left with "I'll get back to you". Nobody knows whether they did.
What it costs: every lead the chat could not close is lost, and nobody counts it. The dashboard shows a resolved conversation.
The fix: when the chat cannot confirm a detail or the customer is ready to buy, ask for a number or email and promise a call within working hours. At night, always. A chat that captures a contact at midnight has done its job. A chat that says "call us" at midnight has shown the customer the door.
Mistake 4: long answers in the wrong language
Twenty-seven replies in English to Serbian questions, mostly triggered by product codes. Replies of 600 to 1,400 characters to customers who were, by the store's own analytics, half on phones. The same word typed twice, the identical list twice.
What it costs: on a phone, 1,200 characters is four screens before the price. Most customers do not scroll four screens.
The fix: reply in the customer's language regardless of what the input looks like, three or four sentences, and never repeat a list already given. These are configuration lines, not model upgrades.
Mistake 5: small errors in a confident tone
A valve type explained under the wrong name. One manufacturer described as Serbian, Slovak and Slovenian in three separate answers. A quiet fan model presented as the turbo version, then retracted when the customer pushed back.
What it costs: none of these is a large error. Each one is a small withdrawal from the only account a chat has, which is the customer's willingness to believe the next sentence.
The fix: a knowledge base with the facts customers actually ask about, corrected by the salesperson who knows the stock, and an explicit instruction that "I'm not sure, let me get that confirmed" is a good answer. The same chatbot already did this well for diameters, decibel ratings and drawings. It just had not been given the rest.
Mistake 6: measuring clicks
The vendor dashboard reported 108 product clicks and a healthy conversation count. The store's order data showed zero purchases attributed to the chat in four months.
What it costs: four months. Clicks look like work, so nobody asks the only question that matters.
The fix: before switching a chatbot on, define two numbers: orders where the chat was the last touch before checkout, and contacts captured when the chat could not close. Review both monthly. If both are zero, the chat is a search box with manners, and the vendor's conversion statistics describe someone else's store.
Why the model is not the problem
The same chatbot that made all six mistakes also produced two genuinely good consultations: a comparison of two cistern models and a correct explanation of which drain grating carries 12.5 tonnes versus 1.5. In both, the customer supplied all the context unprompted, and the chat did exactly what a salesperson would.
So the model could ask, compare and advise. Nobody told it to ask, what to ask, or when to stop listing. The experienced salesperson in that shop knew all of it. It lived in his head.
AI amplifies what it finds. In this store it found a catalogue and multiplied it 238 times. I wrote about the same effect at company scale in AI will speed up your chaos. A chatbot is that effect at the size of a single conversation.
What to do before you install the next one
- Sit with your best salesperson and write, for each product group, the one question that separates the models, what customers get wrong, and what comes back.
- Give the chat three rules: ask before listing, no closest match, take the contact.
- Set the two numbers you will measure, and read them, not the click count.
If you cannot do step one, the chatbot will fail the same six ways regardless of vendor. The full before-and-after with figures is in the case study. How a salesperson's knowledge moves into a system without leaving with the salesperson is on the AI integration page.