It was almost midnight when a seller I was helping scrolled back through her store's chat log from that evening. One message, 22:07: "do you have this in 42?" No reply. The visitor left ninety seconds later. The shoe was in stock. In 42. Sitting in her own warehouse.
That one line is the entire case for putting a chatbot on an online store. Not engagement. Not efficiency. Somebody with a card in their hand asked a small question and got silence.
I've read a lot of these logs by now, and the pattern never changes: the questions are small, the answers already exist on the site somewhere, and the seller is asleep.
Live chat was a promise most stores can't keep
Live chat sold everyone on a fantasy: a real human, right there, instantly. In practice that human has a body, a dinner and a Thursday. Most stores I look at answer their site chat in twenty to forty minutes during working hours, and not at all after nine at night, which happens to be exactly when a large share of shopping traffic shows up.
The widget stays on the page anyway. So the visitor types, waits, reads the little "we usually reply within a few hours" line, and quietly closes the tab. You paid for the ad. You paid for the visit. Then you turned an interested buyer into an unread message.
An AI chat widget is not better than a good human. It is better than nobody, and nobody is what you actually have at 11pm. We've written separately about how sharply conversion falls as reply time stretches, and the shape of that curve in the first five minutes is brutal.
The three questions that decide the sale
Strip a hundred pre-purchase chats down to their bones and you get roughly the same three things, over and over:
- Do you have it? Size, colour, variant, quantity. The answer lives in your stock table and changes hour by hour.
- Will it work for me? Measurements, materials, compatibility. "I'm 178cm and 80kg, which size?"
- What happens after I pay? Shipping time, cost to their city, returns, warranty, whether it arrives before Thursday.
None of these are hard. All of them are urgent. An assistant wired into your product data and your shipping rules answers all three in a couple of seconds, at three in the morning, and doesn't get bored on the four hundredth repetition.
The condition is that it answers from your data rather than from guesswork. A chatbot that improvises stock levels is worse than no chatbot at all, because now you've promised a size 42 you don't have and next week you're refunding it with an apology attached.
Cart rescue is a timing problem, not a discount problem
Most abandoned-cart tooling fires an email three hours later offering 10% off. That isn't rescue, it's a bribe paid to people who were probably coming back anyway.
The interesting moment is earlier. Someone is sitting on the checkout page. They've been there ninety seconds and haven't touched a single field. They are not confused about the button. They're stuck on something they couldn't confirm: shipping cost to their city, whether the return window is real, whether the payment page is safe.
A message at that exact moment that says something specific, like "shipping to your city is 2 to 3 working days, returns stay open for 7 days, anything you want to check before you finish?", recovers more carts than a discount does. It also doesn't train your customers to abandon on purpose so the coupon shows up.
The rule I keep coming back to: never trigger on entry, never trigger twice, and always open with information instead of a greeting. "Hi 👋 how can I help?" is wallpaper. Nobody answers wallpaper.
Your site chat and your social inbox are the same inbox
Here's what actually happens. A customer sees a reel, asks in DMs, gets an answer, comes to the site two days later and asks the same thing again, because the site widget has no idea who they are. Then they ask a third time on WhatsApp before ordering.
Three conversations. One customer. Three chances to contradict yourself about the return policy.
If your website chatbot and your Instagram, WhatsApp and Telegram replies run off two separate brains with two versions of the truth, you'll eventually say the wrong thing and the customer will catch it. One knowledge base feeding every channel is the single thing that separates a store that looks professional from one that looks improvised. Our piece on what actually works in AI customer service for online stores goes deeper on keeping those answers consistent.
Teach it your catalogue, not a script
The old chatbot model was a decision tree. Press 1 for orders, press 2 for returns. Customers hated it and they were right, because nobody's question fits neatly into your menu.
The version worth installing reads your product pages, your FAQ, your shipping table and your stock feed, then answers in ordinary sentences. Setup looks less like building a flow and less like programming, and more like handing over what you already know. In practice that means:
- Product data with real attributes, not just a title and a price.
- Your shipping and returns policy in plain language, including the exceptions you currently keep in your head.
- The five things you always say on the phone that never made it onto the site.
That last one is where most of the value hides. If your shop runs on WordPress, the step-by-step guide to adding an AI chatbot to WordPress covers the install side. The content side is yours, and it's the part that decides whether the thing is useful or embarrassing.
Know when the bot should step aside
An assistant that never hands off will eventually try to answer a complaint about a broken order, and that is how you lose a customer permanently.
Write the handover rules before launch, not after the first disaster. Mine are simple: anything about money that has already been paid, anything where the customer is angry, any legal or medical claim, and any question the assistant has failed twice. Those go to a human with the full conversation attached, so the customer never has to start the story over.
The rest, which in a typical store is somewhere around 70 to 80% of everything that lands in chat, is pure repetition and should sit entirely on the bot. Worth reading alongside this: the three unanswered questions that quietly kill sales.
A first week that proves it, or doesn't
Don't judge a chatbot by how clever it sounds. Judge it by a number you can pull on Friday.
- Before you install anything, export last month's chats and count how many got no reply at all. That's your baseline, and it is usually uglier than you expected.
- Launch the widget on product and checkout pages only. Chat on the homepage is noise.
- Read every conversation for seven days. All of them. You'll find three wrong answers and fix them in ten minutes.
- Track two numbers: the share of chats answered under a minute, and the orders where the buyer chatted first.
- At the end of the week, compare that second number against the same stretch before launch. If people who chat aren't converting better than silent visitors, your answers are wrong, not the idea.
Sellers who run that loop for a month end up with an assistant that sounds like them. Sellers who install it and walk away end up switching it off and blaming the technology.
If you want your site widget and your Instagram, WhatsApp and Telegram conversations answered by one assistant that knows your catalogue and your rules, that's what we built Vardast for. Connect your store and your channels, hand it your product knowledge, and let it answer in your voice around the clock while passing the hard ones to you. Start with the questions you're already tired of typing.