The usual way to launch a website assistant is to paste the script into the site template on a Tuesday afternoon and see what happens. Suddenly it is on every page: the homepage, the blog, the careers page, the privacy policy. If it gets something wrong, it gets it wrong in front of everyone at once, and you find out from a confused customer rather than from your own checks.
A soft launch avoids that. You put the assistant on one page, the one where visitors are most likely to have a real question, and you watch closely for a fortnight. You learn what people actually ask, fix the gaps while the volume is small, and only then roll it out.
It costs you two weeks. It saves you from learning in public.
Why one page tells you more than ten
Spreading a new assistant across a whole site gives you a thin layer of conversations from every kind of visitor. Someone on a blog post asks something general. Someone on the contact page wants a phone number. Someone on the homepage says "hi" and leaves. It is hard to see patterns in that.
Concentrating on one page gives you a consistent audience with a consistent reason for being there. Twenty conversations from people reading your pricing page will reveal whether the assistant handles price questions properly. Twenty scattered conversations reveal very little about anything.
It also keeps your workload sensible. In the first fortnight you should read every transcript, and that is only realistic when there are a manageable number of them.
Choosing the page
Pick the page where visitors are closest to deciding and most likely to have a question the page does not fully answer.
| Candidate page | Good for a soft launch when | Watch out for |
|---|---|---|
| Pricing | Visitors ask about extras, discounts or what is included | Prices in the knowledge base must be current before you start |
| Main service page | One service brings most of your enquiries | Questions about other services will still come |
| Booking or quote request page | People hesitate before filling in the form | The assistant captures requests, so make sure someone follows up |
| Areas covered | Location decides whether you can help at all | A thin areas page means thin answers |
| Homepage | Your site has very few pages | Low-intent visitors make results noisy |
For most service businesses the answer is the pricing page or the main service page. Our post on which pages need an assistant goes into this choice in more detail.
The two-week plan
- Before day one, test privately. Ask the assistant your ten most common questions and a few you know it cannot answer. Fix anything wrong. The approach in testing with real past enquiries is a good way to do this.
- Day one: add the script to the chosen page only. Not the template, not the footer.
- Days one to five: read every conversation the same day. Note wrong answers, vague answers and questions it could not answer.
- Each evening, fix at the source. A wrong answer usually means a page, document or Q&A pair needs editing. Change it, then ask the same question again to confirm.
- Day seven: review the week. Look at conversations, leads captured, ratings and the unanswered questions list together.
- Days eight to fourteen: repeat, reading more lightly. By now most issues should be ones you have seen before.
- Day fourteen: decide. Go wider, extend the trial, or change the setup.
A worked example with numbers
Here is an illustration. Replace every figure with your own.
Say your pricing page gets 600 visits a fortnight. You switch on a proactive greeting there and, over two weeks, 30 visitors start a conversation. Reading all 30 takes about ten minutes a day.
Of those 30, suppose 18 get a correct answer and move on, 7 leave their details for a quote, 3 ask something the knowledge base does not cover, and 2 get an answer that is wrong because an old discount page is still published.
That last pair is the whole point of the soft launch. Two visitors saw a wrong answer instead of two hundred. You unpublish the old page, re-crawl and re-test. The 3 unanswered questions become Q&A pairs. The 7 captured quote requests tell you the lead capture flow works.
Had you launched site-wide, the same problem would have appeared across every page for as long as it took you to notice.
What to watch, and what to ignore
In a soft launch, some numbers matter more than others.
- Wrong answers. The most important signal. Count them from the transcripts, not from ratings, because many visitors never rate.
- Unanswered questions. These show what your content is missing. A steady stream of the same question means one missing fact.
- Leads captured. Confirms the handover to your team works end to end. Check that each lead actually reached you.
- Conversations that stop after the greeting. A few are normal. A lot may mean the greeting feels pushy or irrelevant to that page.
Ignore the total conversation count for now. It depends mostly on traffic to that one page and says little about quality.
Deciding to go site-wide
You are ready to roll out when three things are true for a full week: no wrong answers, unanswered questions that are all new rather than repeats of ones you already fixed, and captured leads reaching the right person.
If you are not there, extend the trial by a week and keep fixing. If the page is simply too quiet to learn from, move the assistant to a busier page rather than waiting.
When you do go wider, add it to the template, but consider switching off the proactive greeting on pages where visitors are not shopping, such as blog posts and legal pages. Then keep reading a sample of conversations each week for the first month.
Start this week
Choose your page using the table above and write down why you picked it. Spend thirty minutes testing the assistant privately with your common questions. Then add the script to that one page and put a daily ten-minute slot in your calendar to read the conversations.
You can do all of this on the free plan to begin with, or try the assistant on your own site before adding any code. For what to track once you are live everywhere, see the metrics worth watching.