Every business has a season where the website suddenly gets busier than the team does. Retail has the run-up to the holidays. Accountants have year-end. Trades have the fortnight before everyone closes for the break. Whatever shape yours takes, the pattern is the same: more visitors, more repeat questions, and less slack in the day to answer them.
Website chat helps with that, but only if it was set up for the volume rather than the quiet months. Here is the check we run before a busy period, in the order that catches the most problems.
Start with the knowledge, not the settings
The single biggest cause of a bad chat answer in a busy season is knowledge that went stale in a quiet one. Your prices moved. A service was retired. The cut-off date for orders changed and only the person who wrote the email knows.
Open your knowledge sources and read them the way a customer would. Anything with a date in it deserves attention: opening hours over the break, delivery cut-offs, lead times, seasonal availability. Anything with a price in it deserves more.
Three practical notes while you are in there. You can edit source text directly, so a small correction does not require a re-upload. Reindex after an edit so retrieval picks up the change. And if you are adding new material, remember a URL import handles one page. There is no full-site crawl, so add the specific pages that matter rather than pointing at your sitemap and hoping.
A quick test that works
Write down the ten questions your team answered most last December. Ask the chat each one. Anything it gets wrong or refuses is a missing knowledge source, and you have just found it before a customer did.
Tighten grounding where the answers cost money
Grounding mode is the setting people configure once and never revisit. A busy season is a good moment to revisit it, because the cost of a wrong answer goes up when you are dealing with more people and have less time to fix things.
Strict mode checks for relevant knowledge and can return your configured refusal without calling the answer model when it finds nothing suitable. That is the right behaviour for anything priced, promised or time-bound. Balanced mode prefers your supplied knowledge and instructs the model not to invent business facts, which suits general support where tone matters. Open mode drops the retrieval gate entirely, which is rarely what you want when the phone is already ringing.
While you are there, reread your refusal message. Most of them were written in a hurry and read like an error. In a busy season it is the sentence that decides whether someone waits for a person or leaves, so make it useful: say what you can’t confirm, and offer the next step.
Decide what happens at 10pm
Traffic in a busy season does not respect your opening hours. Enquiries arrive in the evening, at the weekend, and on the days you closed for the break. The question is not whether you can answer them then (you can’t) but whether you capture them.
Set your business hours and timezone, then choose the after-hours behaviour deliberately. Pair it with a By availability branch in your flow so the offline path is written for someone who knows they are waiting: a short apology, a Form step that collects what your team will need in the morning, and a clear expectation of when they will hear back.
An offline enquiry with a name, a phone number and a description of the job is worth several times an offline enquiry that says “hi, is anyone there”.
Get the inbox ready for more people
Two conversations a day can be handled by whoever notices them. Twenty cannot. Before the volume arrives, agree who is picking things up and make it visible through assignment rather than assumption.
A few small things pay for themselves quickly:
| Prepare this | Why it matters in a busy week |
|---|---|
| Shared saved replies for the seasonal questions | Stops five people writing the same delivery-cut-off answer five different ways |
| Private notes on anything unusual | The next person picks it up without re-asking the customer |
| Agreed use of resolve and reopen | You can tell at a glance what is still live |
| A quick search habit | Contact details, messages, notes and tags all search, so returning customers do not start from zero |
Saved replies insert into your draft rather than sending themselves, so a shared library speeds people up without letting a stock answer go out unread.
Check the widget still fits the site
If your website changes for the season with a banner, a countdown or a campaign landing page, check the widget still sits somewhere sensible. Placement offsets exist for exactly this. It is also worth confirming the teaser text matches the season: “Ask about delivery before the 20th” earns more opens than a permanent “Hi there”.
Look at the numbers afterwards, and honestly
When the rush is over, the response-time reports tell you what actually happened. Use the 30-day window and read three things: coverage, including the conversations nobody answered; the 90th percentile, not the average; and resolution timing separately from first response.
Remember what those numbers are and are not. They use retained history and elapsed time in UTC, not business hours, and not a contractual SLA. An overnight enquiry answered first thing reads as a long response time even though your team did nothing wrong. Compare like with like, week to week, and the trend is still the most useful thing you will have.
A short version
Refresh the knowledge first, because everything downstream depends on it. Set grounding to Strict wherever a wrong answer costs money, and rewrite the refusal message so it helps. Configure business hours and an offline path that collects real detail. Prepare shared saved replies and agree who owns the inbox. Then check the report afterwards and fix the one thing that stands out.
None of it takes long. All of it is easier now than in the middle of your busiest week.
Ready to tune your setup? See how the grounded AI agent and knowledge base work
