The problem: AI chatbots that don't know your business
A generic AI chatbot answers from what it learned on the internet, not from your pricing, your services, or your policies. Building one that answers from your own content usually means hiring AI engineers and running servers, which small and mid-sized businesses won't do.
The Chat Company needed to make that a quick self-serve setup: point it at a website, and get a chatbot that knows the business, looks like the brand, and installs anywhere.
Taking over an AI-generated demo: what I kept and what I rewrote
When I started, The Chat Company was a first version the founder and another developer had generated with AI tools as a demo. It showed the idea well enough to pitch, but it was buggy and wasn't built for real customers.
I reviewed all of it, kept about 20% (the general backend code that worked), and rewrote the other 80%. Then I built the rest of the product on that foundation. Nine weeks later it was a working SaaS, and it now serves 30+ businesses.
How a business sets up an AI chatbot for its website
The whole onboarding runs as a step-by-step flow that remembers where each business left off, so a slow crawl or a closed tab never means starting again.
Enter the website URL
The crawler maps every page on the domain.
Pick the pages
The business chooses what the bot should learn.
The bot learns them
Pages are scraped, split into chunks, and indexed for search.
Style the chat window
Pick one of 56 templates and adjust it live.
Paste one script tag
Works on WordPress, Shopify, or plain HTML.
RAG: making the chatbot answer from the website, not guesswork
RAG (retrieval-augmented generation) means the bot looks up the relevant parts of your content before it answers. Each page is split into chunks of about 500 characters that overlap slightly, so no sentence loses its meaning at the edges. For every question, the five most relevant chunks are pulled and given to the model as context.
Two safeguards matter in production. Every search is filtered by business, so one customer's content can never show up in another customer's chatbot. And if the lookup is slow or fails, the chat still answers instead of breaking.
A visitor asks a question
The five most relevant chunks of that business's content are found
They're added to the prompt as context
The model streams an answer grounded in the website
One script tag that works on any website
The chat widget is a single JavaScript file with its styles built in. It loads the business's settings, streams answers, and tracks sessions, without touching the rest of the site. That's why it installs the same way on WordPress, Shopify, or a hand-coded HTML page.
Businesses style it in a live customizer: 56 templates, plus controls for colours, fonts, message bubbles, the launcher button, position, and pop-up behaviour. Every change shows instantly in a preview next to the controls.
Chatbot analytics: seeing what visitors actually ask
The dashboard shows sessions, messages, unique users, token usage, and a replay of every conversation. For a business, that's a list of the questions its website doesn't answer well. For the product, token tracking and capped conversation length keep AI costs predictable.
The knowledge base view shows exactly which pages and chunks the bot has learned. The screenshot below is from Authentic Urban, the D2C store I also built.
Inside The Chat Company
Results
Key decisions
Filter every search by customer
One business's content must never appear in another's chatbot. Filtering at the search layer makes that a hard guarantee, not a hope.
Ship the widget as one self-contained file
Installation had to be one step on any platform. A single file with its own styles doesn't depend on the host site's code.
Cap conversation history and answer length
Long conversations send more text to the model every turn. Capping them keeps each customer's AI bill predictable.
Onboarding as a resumable step-by-step flow
Crawling and indexing take time and can fail. Saving each step means a business always picks up where it left off.
Related reading