I build these for clients and run one on this site. Below is how an API-free chatbot works, where it beats an AI one, where it falls short, and how to set one up so it actually produces leads.
How a chatbot works without an AI API
An AI chatbot sends every visitor message to a model like GPT or Claude and returns whatever it writes. An API-free chatbot does the matching itself. You give it a list of entries, each with a few ways a visitor might ask the question, some extra keywords and one approved answer:
Q: How much does it cost?
Q: What are your prices?
K: price, budget, quote
A: Every project is quoted after a short look at what you need.
Share a few details and you get a clear price before anything starts.When a message comes in, the plugin strips it down to meaningful words, maps synonyms together (“price”, “budget” and “how much” all become the same idea), scores every entry and replies with the best match if it clears a confidence threshold. If nothing matches well, it does the one thing that matters: asks for an email address so a person can answer.
Good matching needs three things most free plugins skip: stop-word removal so “what”, “do” and “you” don’t decide the answer, synonym groups, and the ability to answer two-part questions like “do you do SEO and Google Ads?” with both entries.
When an API-free chatbot is the better choice
- Your answers must be exact. Clinics, legal and financial services, event venues with fixed packages. An AI model will sometimes paraphrase a policy into something you never said. Approved answers can’t.
- You don’t want a monthly bill that scales with traffic. Every AI reply costs money. A spike from an ad campaign or a bot attack can turn into a surprise invoice.
- Most questions repeat. Look at your inbox and DMs. If fifteen questions cover most of what people ask, you don’t need a language model to answer them.
- You care about privacy. Visitor messages never leave your server, which makes the GDPR conversation much shorter.
Where it falls short
It won’t handle questions you never wrote an answer for, long rambling messages, or follow-ups that depend on context (“and what about the second option?”). It also can’t write fresh copy on the fly.
The fix is not pretending it can. When confidence is low, the bot should say so and ask for contact details. A visitor who gets “Good question, share your email and Rafi will reply personally” is a lead. A visitor who gets a confident wrong answer is gone.
If you later want smarter replies, the best setup is hybrid: AI answers first, your approved answers act as the backup when the API fails or the monthly cap is reached, and the AI is told to stay consistent with those answers.
Setting it up so it produces leads
1. Collect the real questions
Pull the last three months of enquiries from email, contact forms, WhatsApp and Instagram. Group them. You will usually find 10 to 20 clusters: price, availability, timing, location, process, “can I see examples”, and a handful specific to your business.
2. Write short answers that end with a next step
Two or three sentences, in your own voice, and every answer should point somewhere: a booking link, a page, or “share your details”. An answer that ends the conversation wastes the visitor’s attention.
3. Add the phrasings people actually use
Visitors don’t type “What are your prices?”. They type “how much”, “rates?”, “is it expensive”. Add two or three phrasings per entry and a few keywords. This is where most of the matching quality comes from.
4. Capture the lead properly
Ask for name and email (phone optional) and send the full conversation with it, so you never reply with “what were you looking for?”. Keep a copy inside WordPress as well as in your inbox, because email delivery fails more often than people think.
5. Make it work with page caching
This is the bug I see most on WordPress chatbots. Many plugins protect their form with a WordPress nonce, which expires. When a caching plugin like LiteSpeed or WP Rocket serves a page cached yesterday, the nonce is stale and the chat silently fails. Use rate limiting and a honeypot field instead of nonces on the public endpoints.
6. Read the conversations
After launch, read what people actually asked for the first two weeks. Every unanswered question is a new entry. This is the step that turns a decent bot into a good one.
What it looks like in practice
For 2 B Luxe Events, an event planner, I built a concierge that answers package and availability questions, switches to voice for visitors who prefer to talk, and emails every lead with the full transcript. On this site, the assistant in the corner runs in answers-only mode with no API at all. Try asking it about pricing or timelines.
If you want one on your own site, the AI chatbot service page explains how I set it up, and has a quick estimate of how many leads a chatbot could capture from your traffic.