Train Your AI Receptionist on Your Website Content
How Captivar reads a client's website to feed the AI receptionist, what gets captured, and how to edit or supplement the knowledge base by hand.
A receptionist with a great persona but no knowledge about the business is a confident liar. This article covers how Captivar teaches the receptionist what it needs to know — by reading the client’s own website — and how you take over that knowledge base when the website is missing details or saying the wrong things.
The knowledge base concept
Every site in Captivar has a knowledge base: a single block of text that contains everything the AI receptionist knows about the business. When a visitor sends a message, the knowledge base goes into the AI’s working memory along with the conversation history. The receptionist’s answer is grounded in that text.
If something isn’t in the knowledge base, the receptionist either declines to answer or escalates. This is the design — a receptionist that makes things up is worse than one that admits it doesn’t know.
The knowledge base is one editable text field. You can read it, edit it, replace it entirely. There is no version history or undo, so for large changes, copy the existing content somewhere safe before overwriting.
How Captivar fills the knowledge base initially
When you add a site to Captivar, the platform attempts to scrape the client’s website automatically. The goal is to give you a draft knowledge base on day one without manual effort.
Captivar tries three approaches, in order:
- WordPress REST API. If the site is WordPress and exposes the standard REST endpoints (most do), Captivar pulls every page and post via the API. Clean text, structured, ignores theme bloat.
- HTML crawler. If the WordPress route fails or the site isn’t WordPress, Captivar crawls the site like a search engine — starting from the homepage, following internal links, and extracting visible text from each page.
- External reader service. If the HTML crawler chokes (heavy JavaScript site, behind a paywall, weird redirects), Captivar falls back to an external reader that returns clean markdown.
After scraping, the result is written to the knowledge base. The first scrape typically takes 30 seconds to a few minutes depending on site size.
What good content looks like (and what doesn’t)
The receptionist will be only as good as the content it reads. Some sites are dream training data; others need significant manual curation.
Sites that scrape well:
- Pages that name services clearly and describe them in plain English
- A pricing or fees page (even with ranges, not exact numbers)
- An FAQ or “Common questions” page
- A team or about page with names and titles
- A contact page with hours and locations
Sites that scrape poorly:
- Pages that lean on imagery instead of text (a heroic photo and a one-line tagline)
- Industry jargon without explanation
- Generic copy that could describe any business in the category
- Stale information (last year’s hours, pricing that’s been raised, services that have been dropped)
- Pages locked behind logins or contact forms
If the client’s site falls into the second column, the scrape will produce a thin knowledge base. You’ll need to add content manually.
Reviewing what was captured
After the scrape completes, find the knowledge base in the dashboard:
Open the Receptionist page. Sidebar → Receptionist. Make sure the right site is selected in the site switcher.
Scroll to the Knowledge section. Below the configuration sections, you’ll find a panel titled Knowledge base with the captured content displayed in an editable text area.
Read it. Top to bottom. This is exactly what the AI sees on every conversation. If anything is wrong, missing, or outdated, that’s what you’ll fix.
What to look for:
- Inaccuracies. Prices that aren’t current. Services that have been added or removed. Staff names that have changed. Wrong hours.
- Missing content. Common questions the receptionist would need to answer that don’t appear anywhere. Insurance acceptance, parking, languages spoken, financing options.
- Confusing wording. Marketing copy that sounds great in a hero section but doesn’t read like an answer to a real question.
- Internal jargon. Acronyms or program names that mean something to staff but nothing to a prospect.
Editing the knowledge base by hand
The knowledge base is a plain-text field. You can rewrite, append, or restructure it however helps the receptionist answer better. Some approaches that work:
Add a “Common questions” section
Even if the website doesn’t have an FAQ, you can write one directly into the knowledge base. Format it as Q-and-A pairs the way you’d want the receptionist to respond.
## Common questions
Q: Do you accept new patients?
A: Yes, we accept new patients. Our next new-patient slot is typically
within two weeks. The first visit takes about 90 minutes and includes
a comprehensive exam, X-rays, and a cleaning.
Q: What insurance do you accept?
A: We accept Delta Dental PPO and Premier, Cigna, Aetna, and MetLife.
For other plans, we can usually file as an out-of-network provider —
call us with your plan name and we'll confirm.
Q: Do you have weekend or evening appointments?
A: We're open Saturday mornings 8 AM–12 PM. Evening appointments are
limited but available — ask when booking.
The receptionist will use these answers directly when a visitor asks something similar.
Add a “House rules” block at the top
A short block at the top of the knowledge base, telling the receptionist what to do in specific situations, often improves the quality of every answer.
## House rules
- If someone asks for medical advice, decline and offer to book a consultation.
- If someone asks for a specific price for a procedure, say it depends on
the individual case and offer a free 15-minute phone consultation.
- If someone mentions an emergency (active pain, swelling, trauma), tell them
to call the office immediately at (555) 123-4567 rather than book online.
- If someone wants to cancel an appointment, ask for their name and the
appointment date, then offer to reschedule rather than just cancel.
These rules act as overrides. The AI will follow them even if the rest of the knowledge base doesn’t address the exact situation.
Add staff and service detail the website hides
Many business websites are deliberately vague. The marketing copy says “We treat the whole family with personalized care” without ever listing what that means. Add the missing detail:
## Services we provide
- General dentistry: cleanings, exams, X-rays, cavity fillings
- Cosmetic: teeth whitening, veneers, bonding
- Restorative: crowns, bridges, dentures
- Implants: single and full-arch
- Orthodontics: Invisalign (referrals for braces)
## Services we don't provide
- Wisdom tooth extraction (we refer to Dr. Smith at Westside Oral Surgery)
- Pediatric specialty cases (referrals to Brightside Kids)
- TMJ surgery (referrals)
Knowing what the business doesn’t do is often as valuable as knowing what it does. It prevents the receptionist from accidentally over-promising.
Re-scraping when the website changes
If the client updates their site significantly — new services, redesign, fresh pricing — you can trigger a re-scrape:
Open the knowledge base panel. Same place: Receptionist page → Knowledge base section.
Click “Re-scrape from website.” A button at the top of the panel. Confirms before running.
Wait for completion. Re-scraping takes the same time as the initial scrape — usually under a minute. The page will refresh when done.
A common workflow: keep your manual additions in a separate document (Notion, Google Doc, plain text file). When you re-scrape, paste them back in at the top of the new knowledge base. This way you always have the latest website content plus your curation on top.
Testing what the receptionist actually knows
The best test is to chat with the receptionist as if you were a prospect.
Open the client’s website in a private browser window, click the chat widget, and ask the kinds of questions a real prospect would ask. Some you should always test:
- “What services do you offer?”
- “Do you accept [common insurance / payment method]?”
- “What are your hours?”
- “Can I book an appointment for [next Tuesday at 2 PM]?”
- “How much does [common service] cost?”
- “Do you handle [edge case the business actually does]?”
If the receptionist answers any of these incorrectly or escalates when it shouldn’t have, that’s a knowledge gap. Find the topic in the knowledge base, fix or add the information, save, and re-test.
The fix-test loop usually takes three or four cycles per site to get the receptionist into reliable shape. After that, it stays reliable until the business’s offering changes.
When to escalate vs answer
The receptionist’s default is to answer if it knows, escalate if it doesn’t. You can shift that balance by writing house rules.
Make it answer more freely:
You can confidently answer questions about general topics like dentistry,
oral hygiene, and what to expect at a typical first visit, even if the
specific answer isn't in this knowledge base. Use the knowledge base for
practice-specific details (hours, prices, staff, insurance) and general
knowledge for educational questions.
Make it escalate more often:
For any question involving medical advice, specific clinical situations,
or anything not directly covered in this knowledge base, do not answer.
Tell the visitor: "That's something Dr. Patel will want to answer
personally — can I have her call you back?" Then capture the lead.
There’s no universally right balance. Aggressive answer-mode improves visitor satisfaction but increases the risk of a wrong answer. Aggressive escalate-mode is safer but converts fewer prospects directly in chat.
What happens next
The receptionist now has a curated, accurate knowledge base. It can answer real questions, book appointments, and capture leads with confidence.
The next layer of conversion is the form. Where the receptionist works for visitors who want to chat, forms work for visitors who’d rather fill out a request and move on. The next article walks through building one.
Configure Your AI Receptionist — Persona, Greeting, and Tools
Walk through every setting on the AI receptionist: greeting, persona, tone, business hours, escalation, and which tools the receptionist can use.
Install the Captivar Tracker on Your Client's Website
One script tag installs Captivar's visitor tracker on any website — WordPress, Webflow, Squarespace, Shopify, custom HTML. The full step-by-step install.