Launch day feels like the finish line. The widget is on the site, the greeting looks right, and the test questions all came back fine. Then real visitors arrive and ask things nobody on your team would have thought to test.
That is normal, and it is the reason the first month matters more than the setup afternoon. An assistant on day one knows only what you gave it. By day thirty, if someone has been paying attention, it knows what your customers actually ask.
This plan splits the AI receptionist's first month into four weeks with one job each. It is written for a small business where the person doing this also has a day job, so every week has a time budget.
Before day one: set up the notebook
You need somewhere to write things down. A shared document or a spreadsheet with four columns is enough: date, what happened, what you changed, and whether it worked.
Also write down three starting numbers from launch week, even rough ones: conversations, leads captured and unanswered questions. You will compare against them on day thirty. Without a baseline, every change feels like an improvement.
Week 1: read everything, change almost nothing
Time budget: 20 to 30 minutes a day.
The temptation in week one is to fix each problem the moment you see it. Resist, mostly. You are looking for patterns, and one odd conversation is not a pattern.
- Read every conversation from start to finish, not just the first and last message. The interesting part is usually the middle, where a visitor rephrased a question twice and then left.
- Tag each conversation in your notebook: answered well, answered badly, no answer, handed over, or visitor left.
- Note the exact words visitors used for your services. If you say "domestic cleaning" and three people asked about "house cleaning", write that down.
- Fix only genuine errors straight away: a wrong price, wrong opening hours, an old phone number.
- On day seven, count the tags.
Worked example (swap in your own): 70 conversations in week one. Tagged as 41 answered well, 6 answered badly, 13 no answer, 4 handed over, 6 visitor left. That makes 25 conversations out of 70 worth a closer look, and 13 of those are straightforward content gaps.
Week 2: close the gaps that matter most
Time budget: two sessions of about 45 minutes.
Open the unanswered questions log. SpideyChat records these as knowledge gaps, so you are not relying on memory. Group them by topic, and you will almost always find that a few topics account for most of the misses.
Work on those first. For each, choose the right format:
- A Q&A pair for a specific question with a short, stable answer, like "Do you take card payments on site?"
- A document upload for something detailed that already exists, such as an aftercare sheet or a terms document.
- A website page when the question is one your site should have answered anyway. Add the page, then re-crawl.
- Nothing when the question is off-topic or something you do not want answered in chat. Make sure the assistant's fallback is polite and captures details.
Then fix the six conversations tagged "answered badly". These are usually caused by unclear or contradictory content, like two pages giving different prices. Remove the old one rather than adding a third. Writing Q&A pairs that sound like you helps keep the new answers from sounding like a policy manual.
Week 3: tighten the screening questions
Time budget: one hour, plus a conversation with whoever follows up leads.
By now you have two weeks of captured leads. Sit down with the person who calls them back and go through ten together. Ask them one question for each lead: what did you need that was missing, and what did we ask for that you never used?
The answers are nearly always surprising. Maybe the postcode is essential and the "how did you hear about us" question is ignored. Maybe the team always has to ring back to ask whether the job is residential or commercial.
| Screening question | Used by team? | Action |
|---|---|---|
| Postcode or town | Every time | Keep, move earlier |
| Type of property | Often had to ask again | Add |
| Preferred date | Sometimes | Keep, make optional |
| How did you hear about us | Never | Remove |
| Budget | Visitors often skipped | Replace with a range to choose from |
Change the questions, then watch the "visitor left" tag for the rest of the week. If fewer people drop out and your team has what it needs, you are done. The screening questions guide has more examples by trade.
Week 4: review handovers and write the rules down
Time budget: 45 minutes, then set the weekly routine.
Pull every conversation that involved a person: live takeovers, transfers to a human and support tickets. For each, check two things. Was handing over the right call? And did a human actually respond, and how quickly?
You will usually find two kinds of problem. Some handovers should not have happened, because the assistant could have answered if the content existed, which sends you back to week two. Others were right but sat untouched overnight, which is a people and process problem, not a software one.
Write the handover rules as a short list your team can see: which topics always go to a person, who is on duty when, and the reply time you promise customers. Check that the actions you rely on, such as ticket creation or transfer to a human, are actually switched on for this assistant.
Day 30: set the routine that replaces this plan
Compare your notebook with the starting numbers. Then replace the four-week plan with one weekly twenty-minute slot: skim ten conversations, clear the unanswered questions, check that every lead got a reply.
Put that slot in the diary now, on a day that is not Monday. Choose the person who owns it, and give them the notebook. If you are still at the stage before launch, start with setting up in an afternoon, and when you want to see how your analytics will look, the features page shows what gets recorded.