When a quiet store chat is actually a problem
First check whether you have a problem at all. A few chats a week on low traffic means something different from a few chats on thousands of visits. Not everyone who did not write left with an unanswered question, because some people find the answer in the description and simply buy.
A better test sits in your inbox. If questions about size or delivery time reach you by email and phone while the chat stays silent, customers do have questions. They just do not ask them in the window you set up for them. In that case, look for the cause in the chat itself.
If you are still wondering how to add a chat to a website, installing it is the easy part. It is usually a plugin for your shop platform or a script snippet from the vendor pasted into your theme. The problems on this list start later, once the chat is already live.
The bubble nobody can see on a phone
The first mistake is the most down to earth. The bubble hides under the cookie banner or under a newsletter pop-up, and sometimes it is physically impossible to tap. On a phone it can also cover the "Add to cart" button, and then people close it by reflex before they read what is in it. If its color matches the page background, the eye just skips it.
On your own phone, open the shop in a fresh browser session, go to a product page and ask the chat the kind of question a customer would ask. See whether the message field disappears under the keyboard and whether, after closing the window, you land back on the same product. Then repeat the test with a different choice in the consent banner.
If the chat does not load under some setting, ask whoever handled the setup what triggers the script. Do not change consent settings just to get more chats. When you report the problem, include the page address and the steps that led to the error, because "doesn't work on my phone" gives nobody anything to reproduce.
The first message must not be hard
The second mistake is a greeting that says nothing. A generic "how can I help" leaves the customer with an empty field and a guessing game about what they are allowed to ask. On a product page, test an invitation with something concrete, such as "Ask about dimensions or what is in the set", provided the bot really knows that information. If your catalog does not have it, fix the description first, because new text in the bubble will not improve the answers.
A form at the door belongs to the same family. A chat that demands an email and a phone number before it will answer a question about fabric makes people suspect a newsletter is about to follow. Collecting a contact makes sense when the conversation needs it, for example when you have to come back with an answer or check an order, and then say plainly why you are asking. An email given for handling a request is not marketing consent either. Under Polish law you need prior consent to send commercial information, and starting a chat does not give it.
Who is really answering in the chat
The third mistake is a bot pretending to be a person. A first name and a smiling photo in the window do not replace saying that an automated system is answering, and a customer who works it out halfway through has every right to feel misled. Label the bot from the first message, and when a staff member takes over, show clearly that a human is now writing.
The fourth mistake is the reverse of the third: the chat promises a human who is not there. A status of "online" at 10:30 pm, long after you stopped working, makes a customer wait for an answer that comes the next morning. Set real staffed hours. Outside them, say honestly when you will get back to messages.
Check the route to a human as well. Ask the bot for a conversation with your staff without explaining why, and see whether it respects the request and whether your employee sees the full history so far. Then do the same after hours. The bot should collect a contact and leave the matter for the morning. In CartBot you take over a conversation in the panel together with its history, and when nobody is available the bot asks the customer for a contact.
An answer with nothing behind it and a chat with no cart
The fifth mistake sometimes comes out only when a complaint arrives. The bot states a parameter that is not in the description, or confirms a delivery date nobody checked. A customer asks whether the parcel will arrive by Saturday, and a general sentence about fast shipping gives no basis for promising a specific day.
You can test this with a single question about a feature the product description lacks. If the bot gives a specific answer, ask whoever handled the setup where it came from, and keep the conversation to repeat after the fix. A rule against guessing in the bot's instructions lowers the risk without removing it, so you need to read conversations either way.
The sixth mistake is a conversation that never leads to a product. A customer asks about a specific model and gets a paragraph of text, after which they have to search for the product page themselves. For a question like that the chat should show the product card or a link. If it has an add-to-cart button, click it and look at what landed in the cart, because a matching product name does not guarantee the right size. CartBot shows product cards with a photo and price in the conversation and adds products to the cart, although for some variants the customer finishes the choice on the product page.
A chat left to run on its own
The seventh mistake is assuming the chat runs itself after launch. You raise the free-shipping threshold, nobody updates the knowledge base, and for the next few weeks the bot promises free delivery to customers who did not reach the new threshold. Decide who updates the knowledge base and checks the bot after every change. The bot knows products and prices from the last catalog sync, so now and then glance at the state of that sync too.
Measurement can also be left unattended. Place a test order after a conversation and check that the report shows it once and with the right amount. Revenue attributed to chat is not profit. On the other hand, a conversation in which the bot explained a complaint to someone is worth something even if nobody bought anything afterwards.
The simplest habit is to read conversations once a week, starting with the ones handed over to a human. Write down each discrepancy together with the customer's question and the date of the fix, and after the change ask the bot the same question again.
Seven tests for one afternoon
Run each test in a new conversation and write down the result, so you can repeat it after you change the theme or a plugin.
- Open the chat on a phone with the cookie banner visible.
- See what a customer sees before the first message, meaning whether there is a concrete greeting or straight away a form asking for email and phone.
- Check that the bot introduces itself as a bot.
- Write after hours, ask for a human, then find that conversation in the panel and see whether the contact you left was saved.
- Ask about a parameter missing from the description and about delivery by Saturday.
- Ask for a specific model in a specific size, add it to the cart from the chat and check that the size is right.
- Change something in your delivery rules, ask the bot about it the next day, and finally place a test order and find it in the report.