A small business shopping for a customer support chatbot sees twenty products that all promise the same three things: instant answers, happier customers, less time in the inbox. The sales pages are interchangeable. The products are not, and the differences that matter are rarely on the pricing page.
This guide is the set of questions to put to any vendor, the red flags that should end a conversation, and a one-page checklist to take into a trial. One disclosure: SpideyChat is our product. It appears below as one option, and the questions apply to it as much as to anyone else.
Before you talk to anyone: your own numbers
Four figures first.
- Enquiries per week, across all channels.
- The share that arrive outside your hours.
- The share that are the same ten or fifteen questions.
- What a missed enquiry costs you, roughly. A lost booking, a lost basket, a lost job.
The chatbot ROI calculator turns those into a monthly figure, and that figure is the ceiling on what any tool is worth to you.
Question 1: Where does it get its answers?
A support chatbot is only as good as what it has been given to read. Ask where the answers come from.
The answer you want is "from your content": your website pages, documents you upload, text you paste in. Ask how many pages it reads, whether it re-reads them when you change something, and whether you can add a correction without touching the website.
The answer you do not want is a general model answering from what it knows about your industry. That produces plausible, confident and occasionally wrong replies about your returns policy, which is worse than no reply.
Question 2: What happens when it does not know?
This is the most revealing question on the list, and the only way to answer it honestly is to test it. Ask five questions your website does not cover: a service you do not offer, a policy you have never written, a price that does not exist.
The right behaviour is to say it has no answer, offer to take the visitor's details, and log the question somewhere you can see it. The wrong behaviour is to produce something that sounds right. A vendor who cannot show you the list of unanswered questions from a live account has not thought about this.
Question 3: Can it capture leads and open tickets?
Answering is half the job; the other half is producing something your team can act on.
Ask whether it collects a name, email and phone number in the conversation, and where those land. Ask whether a customer reporting a problem gets a numbered ticket and whether your team sees the full exchange attached to it. Ask how you are notified, and whether the notification arrives at 11pm or waits for the morning.
If the answer to any of these is "it emails you a transcript", you are buying a chat window, not a support tool.
Question 4: How does hand-off to a human work?
Ask three things. When does the bot step aside, who decides, and what does the person receive.
A good product lets you set the triggers (a complaint, a request for a refund, a visitor asking for a person) and passes the chat to someone on your team who is on duty in the inbox, with the history intact. Ask what happens when nobody is on duty. The honest answer is that it takes details and opens a ticket, and anyone who claims otherwise is describing a staffed service rather than software.
Question 5: How is it priced?
Three pricing models are common, and they produce very different bills.
Flat monthly. A fixed fee for an allowance of replies or conversations. Your busiest month costs the same as your quietest, as long as you stay inside the allowance. Ask what happens when you exceed it: a hard stop, a slowdown, or an upgrade prompt.
Per conversation. A fee each time a visitor opens the chat, including the ones who type "hi" and leave. Costs rise exactly when the chatbot is doing its job.
Per resolution. A fee for each conversation the vendor counts as resolved. Ask who defines "resolved" and how you dispute it. A visitor typing "thanks" and leaving is often counted as a resolution.
Here is a worked example with invented figures. Say a small online shop gets 400 chat conversations a month, 300 of them answered without a person. On a flat plan the cost is the plan. On a per-conversation model you pay for 400. On a per-resolution model you pay for 300, at a rate set by the vendor, and you have 300 small arguments available about what counted. AI chatbot pricing models explained works through each model in more detail.
For reference, SpideyChat is flat: the Free plan is $0 for one chatbot on one website with about 375 replies a month, Starter is $9.99 a month for three chatbots, and Pro is $29.99 a month for ten chatbots, the ticket inbox and branding removal.
Question 6: Branding, install effort and data handling
Three smaller questions that cause disproportionate annoyance later.
Branding. Most products show a "powered by" badge on lower plans. Find out which plan removes it.
Install. Ask what you have to do to put it on the site. One line of code in the template, or a plugin for your platform, is the right answer. A developer project is not.
Data. Ask where conversations are stored, how long they are kept, who at the vendor can read them, how you delete them, and whether your content is used to train anything beyond your own bot. Ask how to keep sensitive details (medical, financial, legal) out of chat and routed to a person. Be wary of certification badges that the vendor cannot back with a document. Data questions to ask before you install a chatbot is the longer list.
Red flags
Any one of these should make you slow down. Two should end the conversation.
- No way to see the questions it could not answer.
- A demo that only runs on the vendor's own website, never on yours.
- Pricing that requires a sales call to learn.
- A chatbot that is set up to pass as a human member of staff.
- Vague answers about where data lives and how it is deleted.
- "It never makes mistakes." Everything makes mistakes. The question is what happens next.
- No way to export your content, leads and conversations if you leave.
The checklist
Tick what you can verify yourself, not what the salesperson says.
- Answers from my own pages and documents, and re-reads them when I change something.
- Says it does not know rather than inventing, and logs the gap.
- Collects name, email and phone in the conversation.
- Opens a numbered ticket with the full exchange attached.
- Hands over to a person on duty, and takes details when nobody is.
- Flat, predictable pricing I can map to my conversation count.
- Branding removal available on a plan I would pay for.
- Installs with one line of code or a plugin.
- Clear written answers on storage, retention, deletion and who can read transcripts.
- Export of my data if I leave.
Then run a trial
Pick two products that pass the checklist on paper and run them side by side for a fortnight, one week with your own questions and one week in front of real visitors. A two-week trial plan for choosing a website chatbot sets out the day-by-day plan and the pass marks. The pricing page has our plans, and the same questions should go to everyone else on your shortlist.