Guides · 8 min read

Live chat best practices: how to run a chat people actually use

Most advice about live chat is about the chat. Almost none of it is about the shop. The habits below come from the other direction — what a small team can actually sustain on a Tuesday afternoon with eleven conversations open and one person on lunch.

Open the chat only where you can answer it

The most common mistake is putting the widget on every page of the site, including the ones nobody needs help with, and leaving it there overnight. A chat bubble is a promise. If it sits in the corner of your blog at 2am and nobody replies until 9am, you have made a promise you break every single night.

Two things fix this, and neither needs a rule engine:

• Put it where decisions happen. Product pages, checkout, pricing, the order tracking page, the returns form. That is where a question costs you money if it goes unanswered — and where an answer is worth the interruption.
• Say when you are there. Outside working hours the chat should say so and collect an email, not pretend someone is watching. “We answer 9–18, Monday to Saturday. Leave your email and we reply first thing” outperforms silence, because it converts a bounce into a conversation you can still win tomorrow.

The number that matters is first reply, not resolution

Teams obsess over how long a conversation takes to close. Customers do not experience that number. They experience the gap between sending a message and seeing a human reply — and after about two minutes of nothing, the tab closes and the sale goes somewhere else.

A useful target for a small team: under two minutes during working hours, and an honest automatic reply outside them. That is achievable without hiring, because most first replies are not answers. They are acknowledgements.

The information is identical. The experience is not: one customer waited nine minutes wondering whether anyone was there, the other waited nine minutes knowing someone was working on it.

Three conversations, not eight

Live chat vendors like to advertise how many chats one agent can hold at once. In practice, quality falls off a cliff after three. Beyond that you are not answering three people, you are keeping eight people mildly annoyed in parallel: replies get shorter, you stop reading the history, and you start answering the question you expected instead of the one that was asked.

If you are regularly past three, the fix is upstream, not in the inbox:

1. Find the three questions that generate the most conversations this month.
2. Answer them on the page where they are asked — delivery times on the product page, not in a policy nobody opens.
3. Whatever survives that, turn into a saved reply or let the AI answer from your help centre.

Deflection is not a trick to avoid customers. It is how you get the time to answer properly the ones that actually need you. See saved replies that don't sound saved for how to write those without ending up with a robot voice.

Proactive messages: rarely, and never on arrival

“Hi! 👋 How can I help?” firing three seconds after someone lands is the pop-up ad of support. It interrupts before the person knows whether they have a question, and it trains everyone to close the bubble reflexively — including the people who would have used it later.

If you use proactive messages at all, tie them to a signal rather than a timer:

• On checkout, after a payment error — “That card was declined by the bank. Want me to check it?”
• On an order tracking page where the status has not moved in five days.
• On a returns page, where people are already unhappy and often just need permission.

One good trigger beats ten timed greetings, and it does not cost you the people who would have asked on their own.

Know who you are talking to before you answer

Asking “can I have your order number?” is a tax you charge the customer for something your own system already knows. If they are logged in, pass their identity to the chat when the page loads — name, email, plan, last order — so the person answering starts with the context instead of collecting it.

This is the single change that most shortens conversations, because the whole opening exchange disappears. Four messages become one. It is also what makes an AI answer safe to send: an assistant with the actual order in front of it can say where the parcel is, while one without it can only quote the shipping policy — which is the reply everyone hates.

End conversations deliberately

A chat that is never closed is a chat nobody is responsible for. Resolve the conversation when it is done, and say so in one line: “I'll close this here — write again on this same chat if the courier doesn't turn up on Thursday and it comes straight back to me.”

Two things follow from that habit. Your open count becomes a real queue rather than a graveyard, so “what still needs me?” has an answer. And the customer learns that the chat is a thread they can return to, not a form they filled in — which is the difference between a support channel and a contact form with better styling.

Four settings worth checking today

• Working hours — set them honestly, including Saturday if you work Saturday.
• Who gets what — new conversations should land with someone, not in a pool everyone assumes someone else is watching.
• Availability — people should be able to mark themselves away, and routing should respect it. Nothing rots faster than a queue assigned to someone who went home.
• Notifications — one sound, on the inbox, that someone is responsible for hearing. Email notifications about chat messages are a way of answering slowly.