Your Website Chatbot's First 30 Days: A Week-by-Week Plan
A four-week plan for a new website chatbot: install and test, read every conversation, tune the greeting and rules, then check the numbers and pick a plan.
Published Updated 8 min read
In its first 30 days, a new website chatbot needs about one focused task a week: install and test it in week 1, read every conversation and fill the gaps in week 2, tune the greeting, suggested questions and instructions in week 3, and look at a few honest numbers and choose a plan in week 4. Plan on two to three hours in the first week and under an hour a week after that. Most of the work is adding missing answers, not adjusting settings.
I build ChatBento, an AI support chatbot for small-business websites. This is the plan I'd hand anyone who just added one, whichever tool they use.
Before you start: what you need
- A simple sheet. Columns for date, question (in the visitor's words), what the bot said, and the fix.
- Your ten most common questions, from your inbox, phone notes or memory. These become your first Q&As and your test.
- One person who owns the bot. Someone who reads the conversations and answers the handoffs.
- A weekly calendar slot for four weeks.
Week 1: install it and test it properly (2 to 3 hours)
The goal this week is a bot you'd trust to answer a customer at 11 p.m.
- Train it on the right pages (45 minutes). Start with the pages customers actually need: services or products, pricing, shipping or service area, policies, contact and FAQ. Skip blog archives and old news. Then add Q&As for your ten common questions, written the way a customer would ask them. How to train an AI chatbot on your website covers what to include and leave out.
- Turn on human handoff (5 minutes). In ChatBento, when a visitor asks for a person, wants to be contacted, or asks something your content doesn't answer, the chat shows a contact form (name, email, message) that's emailed to you. Make sure it's on before anyone sees the bot.
- Run the test (45 to 60 minutes). Ask at least 20 questions: your core facts, prices and policies, a few things your site genuinely doesn't answer, an off-topic request, an attempt to trick it, and a request for a person. My 30 questions to ask your chatbot before it goes live has the full script with what good and bad answers look like. The two answers that matter most: correct facts, and a plain "I don't have that information" when your site has no answer.
- Fix what failed, then retest (30 minutes). Wrong or missing facts get fixed in your content (a corrected page or a new Q&A), not with instructions.
- Install it and do one real handoff (15 minutes). ChatBento installs with one script tag, and there are guides for each website builder. Add your domain to the allowed websites list. Then open your live site on your phone, ask for a person, submit the form with your own email, and check it arrives.
A note on message limits. On ChatBento, one AI reply counts as one message, and test chats in the dashboard count because they're real replies. The Free plan has 50 messages a month, so a 20-question test plus retests can use most of it before a single visitor arrives. If you want room to test, a paid plan can be tried free for 7 days: it needs a card, nothing is charged that day, and it bills after 7 days unless you cancel.
Week 2: read every conversation (about 20 minutes, three times)
This is the week most owners skip, and the one that matters most. Real visitors ask things no test script predicts, in words you'd never use.
Three times this week, read every conversation. In ChatBento they're on the Inbox tab, labelled Lead (they left their details), Wants a human (the form was offered but no email left) or Test (you, from the dashboard). Skip the tests.
For each real conversation, put it in one of three groups:
| What happened | What it means | Fix |
|---|---|---|
| The bot said something untrue | A wrong or outdated page, or two pages that disagree | Correct the page, retrain it. Do these first, the same day. |
| The bot didn't know | The answer isn't in your content, or it's worded differently | Add a Q&A in the visitor's own words |
| The bot answered correctly | Nothing broken | Note it if it's asked often (useful in week 3) |
A Q&A written from a real conversation looks like this:
Q: Do you do same-day repairs?
A: Yes, for phones and tablets booked before 12 p.m., Monday to Friday. Laptops usually take 2 to 3 working days.
Write the "no" answers too. If three people asked about a student discount you don't offer, a Q&A that says so beats "I don't know" every time.
Your chatbot's conversations are a content plan goes deeper on reading them, including which fixes belong on your website rather than only in the bot.
Week 3: tune the greeting, suggested questions and instructions (about 1 hour)
Now you have two weeks of real questions, so you can tune the bot to them instead of guessing.
The greeting (15 minutes)
The greeting is the first message visitors see. Swap a generic "How can I help?" for one that names what you can actually help with, using the topics from your sheet. "Hi, ask me about prices, delivery or returns" tells people the bot is worth trying. 30 chatbot greeting examples has versions by industry and page type.
Suggested questions (15 minutes)
ChatBento shows up to four suggested questions as buttons before the visitor types. Use your most-asked questions from week 2, in your visitors' words, and only ones the bot answers well. A button that leads to "I don't have that information" is worse than no button.
Instructions (30 minutes)
Instructions are short rules about behavior: tone, what to ask for first, when to offer a person. In ChatBento they go in the Settings tab. Add a line only when week 2 showed you a problem the content can't fix. For example:
- Order problems needed a second round trip: "For order problems, ask for the order number first."
- Replies sounded stiff for a casual brand: "Reply in a friendly, casual tone."
- Unhappy visitors got the policy recited at them: "If someone is unhappy, apologize once and offer to pass it to the team."
Keep prices and policies out of instructions; they belong in your content. How to write instructions for your support chatbot has 25 examples, and the free instructions generator drafts a starting block.
After each change, ask five questions it should affect and two it shouldn't.
Week 4: look at the numbers that matter (about 45 minutes)
At small-business volume you need four counts, not a dashboard of charts:
- Real conversations. Not counting tests. This tells you whether visitors use the bot at all.
- Leads. Visitors who left their details through the contact form. Check that every one got a reply in the time you promise. A handoff nobody answers is worse than no chatbot.
- Unanswered questions. How many distinct questions the bot couldn't answer this week, compared with week 2. This list should be getting shorter.
- Wrong answers. The one number that should be zero by now.
If you want the stricter version, resolution, not deflection explains how to work out a resolution rate by hand in 20 minutes a week. Compare against your own week 2, not against vendor benchmarks.
Decide on a plan
Then check how many AI messages you've used this month (the ChatBento dashboard shows it) and compare it with the plan limits:
| Plan | Price | AI messages a month | Chatbots | Pages |
|---|---|---|---|---|
| Free | $0 | 50 | 1 | 50 |
| Hobby | $8/mo | 2,000 | 1 | 200 |
| Standard | $30/mo | 10,000 | 3 | 2,000 |
| Pro | $100/mo | 40,000 | 10 | 10,000 |
Yearly billing gives two months free. The Free plan always shows a small "Powered by ChatBento" link.
The monthly cap is hard: once it's reached, the widget shows the contact form instead of AI answers until the next month, so the bill never grows with traffic. But a cap hit mid-month means visitors spend the rest of it filling in forms. If your first month, minus your own testing, came close to the limit, move up a plan.
Decide whether to keep it
Be honest here too. If after four weeks there are almost no real conversations, the bot may be hard to find, your visitors may prefer the phone, or your site may already answer everything clearly. A bot that gets a handful of after-hours questions and a couple of leads a month is often worth keeping. One that nobody opens is not, whatever it costs.
After day 30
Switch to a monthly 30-minute review of conversations, gaps and leads. Go back to weekly for a few weeks whenever prices, products or policies change: that's when a bot is most likely to quote something out of date.
The 30-day checklist
Print this and tick it off.
Week 1
- Trained on your key pages, with Q&As for your ten common questions
- Human handoff switched on
- At least 20 test questions asked and scored
- Failures fixed in content and retested
- Installed, domain on the allowed websites list
- One real handoff sent from your phone and received
Week 2
- Every real conversation read, three times this week
- Wrong answers fixed the same day
- A Q&A added for every question the bot didn't know, "no" answers included
- Changed pages retrained
Week 3
- Greeting names what the bot can help with
- Up to four suggested questions taken from real conversations
- Instructions added only for problems content can't fix
- Each change tested
Week 4
- Real conversations, leads, unanswered and wrong answers counted
- Every lead answered on time
- Messages used compared with the plan limit
- Plan chosen, and a monthly review in the calendar
If your sheet keeps filling with the same basic questions, the FAQ automation use case shows how a chatbot turns your FAQ and Q&As into answers worded however visitors ask. And if week 3 is the step you're dreading, start with the chatbot instructions generator and edit from there.