Product
Solutions
Compare
Resources
Get early access Talk to us
Virtual Assistant

An AI chatbot for WordPress that only says what your site says

It reads the website you already have, answers visitors in plain English, and puts a link to the page every answer came from. When the site does not say, it says so rather than inventing something. Then it turns the conversation into an enquiry that lands in the same list as your contact form submissions.

Available now in early access. Included on every plan.

northgate-dental.co.uk/treatments
Northgate Dental
Ask Northgate DentalAnswers from this website

Do you take BUPA?

Yes. Northgate Dental accepts BUPA, Denplan and Bupa Dental Cover, and the practice can claim directly for most treatments. Read more: Payment and insuranceIs this for a new or existing patient?

New. How long does a filling take?

A single white filling is normally a 30 to 40 minute appointment. Read more: White fillings

Type your question
In short

What is a WordPress AI chatbot?

A WordPress AI chatbot is a chat panel on your site that reads your own pages and answers visitors' questions out of them, instead of walking people through a scripted menu. The ones worth having show you where each answer came from, and say they do not know when your website does not cover it.

That second sentence is the whole buying decision. Nearly every product in this category says it trains on your content; almost none of them will tell you what happens the moment a visitor asks something your content does not answer. Three things are worth checking before you put one on a client site:

  • Can you see where an answer came from? An answer you cannot trace is an answer you have to take on trust.
  • What does it do when your site is silent? Declining is a feature. Improvising is a liability.
  • Where does the enquiry end up? A lead in a separate chatbot inbox is a lead somebody forgets to check.
The problem

Most website chatbots either cannot answer or will not stop answering

There are two kinds on the market. A scripted decision tree, which cannot answer anything nobody thought of in advance, so the visitor taps through four menus and leaves. And a general AI bolted onto a website, which will answer anything, including things the business never said and does not offer.

The second one is the expensive mistake. In Moffatt v Air Canada (2024 BCCRT 149) the Civil Resolution Tribunal of British Columbia rejected the airline's argument that its chatbot was a separate legal entity, and held that a company is responsible for all of the information on its website, whether it appeared on a static page or came from a chatbot. Your client owns what the chatbot tells their customers.

So the useful question is not how clever the answers sound. It is where each answer came from, and what happens when the website does not contain one.

What this is for

  • Answering the questions a receptionist answers forty times a week
  • Showing a client exactly which page an answer came from
  • Catching the visitor who would have left without filling anything in
  • Qualifying an enquiry before anybody picks up the phone
  • Putting chat enquiries and form enquiries in one list instead of two
How it works

It reads the website, then answers out of it

There is no knowledge base to write and no intents to configure. Connect a website and the assistant reads what is already published on it.

01

It reads your website

Services, treatments, prices, opening hours, FAQs and policies, from the pages and posts already on the site. It keeps reading as the site changes, so a price edited on Tuesday is the price it quotes on Wednesday. Nothing has to be typed in by hand.

02

A visitor asks in their own words

Not a menu of buttons. Someone types "do you take BUPA?", "how long does a filling take?" or "can I bring my dog?" and gets about five lines back, written the way a person would say it.

03

It answers with a link to the page

Under most answers is a Read more link that opens the page the answer came from. That is what turns "the bot said X" into "your website says X, here".

04

You get the enquiry

Once it has helped, it asks one question of its own, and the answers become an enquiry in the same Lead Captures list as every contact form on the site.

Grounded answers

If your website does not say it, the assistant does not say it

Every answer is assembled out of passages found on the client's own site, and carries a link to the page it was taken from. That link is the single most useful thing on the panel when an agency is showing the product to a client, because it makes the answer checkable in one click rather than something to be trusted.

When the website does not state something, the assistant says it does not know and offers a real route instead: the business's own published phone number where the site has one, otherwise an offer to take the visitor's details. It does not fill the gap with something plausible, and it does not reach out to the open internet for a better-sounding answer.

We do not publish an accuracy percentage, because any number is one screenshot away from being disproved. Telling you when it does not know is the stronger claim, and it is the one the product can keep.

When your site is silent

A visitor asks something the website has never addressed. Three things happen, in this order:

  • It says it does not know, rather than producing an answer that sounds right
  • It gives the business's published phone number, or offers to take a message where there is no number on the site
  • It does not then ask the visitor for their details as if it had helped, because it did not

The client can also add their own FAQs and correct anything from the dashboard, which is how a gap gets closed permanently.

Qualifying questions

It asks the questions your receptionist would ask

Answering questions is the half that impresses. This is the half that pays for itself. After the assistant has given the visitor something useful, it adds one short question of its own to the end of the reply. Around twenty industries ship with a set already written, and these are three of them.

A solar installer asks

  • Home or business?
  • Roughly what is your monthly electricity bill?
  • Which area is the property in?
  • When are you hoping to have it installed?

A dental practice asks

  • Which treatment?
  • New or existing patient?
  • Which days or times suit you?

An estate agent asks

  • Buying, selling, renting or letting?
  • What type of property?
  • Which area?
  • What budget?

Four rules that keep it from being annoying

Every set ends the same way, with "what is the best email address or phone number for the team to use?". What stops that from feeling like an interrogation is the four rules underneath it, and every other chatbot on the market breaks all four.

  • It never asks until it has helped. A question only ever rides on a reply that actually answered something. If the assistant had to say it did not know, it does not then ask for a phone number.
  • One question at a time, always the last line of the reply, never a form dropped into the middle of a conversation.
  • It gives up on a question it cannot get. A question is asked at most twice, then dropped and the conversation moves on. An assistant that drops a question is mildly annoying; one that will not move past it is broken.
  • It listens. If the visitor already said which area they are in, it does not ask again.

The client can switch any question off, reword any of them and add up to four of their own from the dashboard. None of it needs a developer.

Pre Defined Step Form

Or ask the questions instead of answering them

Some businesses do not want a conversation at all. They want the visitor walked through a fixed set of questions, in the agency's own words, tapping answers rather than typing. That is the Pre Defined Step Form: a script the agency builds in the dashboard, laid out as a tree so the branching is visible on screen rather than buried in a dropdown.

The visitor gets full-width buttons to tap, a Back button if they answer wrong, and no typing box at all while buttons are the only way forward. Sell it to a client as the contact form people actually finish.

It runs inside the same panel as the assistant, and the two do not talk over each other: while a script is running, the assistant does not staple its own qualifying question underneath.

The client can also see where people drop out, which question lost them, which is the one thing a form sitting on a page can never tell anybody.

Get a quoteFour questions, about a minute

What are you looking for?A quote for solar panelsA battery for an existing systemSomething is not working

A quote for solar panels

Is the property a home or a business?HomeBusinessBack

One list

Chat enquiries land in the same list as your form submissions

This is the point worth making loudest, because it is the one competitors do not make. A chat enquiry is not kept in a separate chatbot inbox. It goes through exactly the same pipeline as every contact form on the website, Contact Form 7, Gravity Forms, WPForms, Elementor, whatever the client already runs, and lands in the same Lead Captures screen, under the same filters and the same exports. One place to look, one place to work from.

app.wpcentrify.com/lead-captures
Lead CapturesChat and form enquiries in one listLive
EverythingUnreadChatForms
TBTom BakerNew patient, white filling, mornings
Virtual AssistantNew8 min
SJSarah JonesQuote for a crown, and an appointment next week
Contact Form 7New20 min
PNPriya NairChatted, then filled in the form. Merged into one enquiry
Chat and formMerged2 h
A chat enquiry, a Contact Form 7 enquiry and a merged one, in the same list.

The same spam protection

The engine that judges form submissions judges these too, with reasons rather than a score, so adding a chatbot is not opening a hole in the wall for somebody to flood.

The same intent and priority

The same reading of what an enquiry is about and how urgent it sounds, so a chat enquiry sorts alongside the rest instead of sitting in a corner of its own.

The whole record

The answers the visitor gave, in their own words, plus the contact details, all on the enquiry and sealed, with a link back to the conversation it came from.

One conversation, one enquiry

The record is created the moment there is a way to reach the visitor, then updated as more is learned. It is never duplicated, so a long conversation does not produce five leads.

Chat then form is one enquiry, not two

If the same person submits a website form shortly after chatting, the two are recognised as one and merged. Nobody rings the same person twice.

A custom field name never loses an enquiry

Details are recognised by what they look like, not by what the field happens to be called, so a form whose field is best_number or correo still produces a lead.

Capabilities

What you get

It looks like the client's website

Colours, the launcher button, the heading, the welcome line, the consent sentence and every other phrase it says are set per website from the dashboard, under your brand rather than ours.

The panel speaks several languages

The interface wording ships in several languages, so the panel does not arrive in English on a site that is not.

Your own FAQs and corrections

Nothing has to be typed in for the assistant to work, but the client can add their own FAQs and correct anything, which is how a gap in the website gets closed without publishing a new page.

Hand over to a person

Where the client has live agents, asking for a human puts the conversation in a queue and a person picks it up. Where they do not, or it is 11pm, it says so honestly and takes a message rather than promising an agent who is never coming.

No second plugin to install

It rides the connection WPCentrify already has to the site. Nothing extra to install on the client site, nothing for them to license, and nothing on the visitor's path waits on a network request to us.

Enquiries are their own permission

Chat enquiries inherit the same roles and per-client scoping as form enquiries, so an account manager sees their clients' conversations and nobody else's.

Every conversation, until you delete it

Conversations belong to the workspace and stay until somebody removes them. The Conversations screen deletes one, or everything matching a filter, behind a typed confirmation, and the enquiry survives either way.

Included, not an add-on

The Virtual Assistant is part of the platform rather than a separate subscription, on the same one price per site with everything included as the rest of it. It is free while WPCentrify is in early access.

It reads what is published

Product pages, service pages and posts are read like any other page, so a WooCommerce shop's published descriptions and prices are answerable. It does not see a visitor's cart, their order status or anything behind a login.

Privacy

A visitor's message is checked before anything leaves our servers

Answers are written on WPCentrify's own machines wherever they can be. When a hosted AI model is used to phrase a reply more naturally, the visitor's actual message is checked first, every time, to see whether it names a real person: a name, a home address, a phone number, an email, or anything about a particular person's health, money or circumstances. It makes no difference whose. The visitor's own, a relative's or a friend's.

If it does, the message does not go. The assistant answers from the website's own sentences instead, entirely on our servers. The reply is a little plainer; the visitor never learns anything was different.

It is a gate, not a filter. The message goes whole, or it does not go at all. Nothing is edited, blanked out and then sent anyway, which is how every redaction system eventually leaks the one thing it did not recognise. And it fails closed. If the check cannot run, the model is unreachable, it times out, it answers something we cannot read, the answer is personal and the message stays here.

Contact details never go in the transcript either. When a visitor gives a name, an email or a phone number, those values go into the enquiry record and the conversation records only that something was shared, so a transcript can be read, exported or shown to a colleague without their details riding along inside it. Enquiry records are encrypted where they are stored. Every agency's data sits behind its own boundary, and anything WPCentrify staff do inside a customer workspace is written to an audit log that cannot be edited or deleted. Our Data Processing Agreement covers the paperwork a client's compliance people will ask for.

What the assistant knows about a visitor

  • Only what the visitor types, plus the page they were on. No tracking pixel, no profile bought from anywhere, nobody followed around the internet.
  • The IP address is never stored. It is turned into a one-way fingerprint on arrival, enough to recognise abuse and block it, and useless for identifying anybody.
  • Conversations belong to the workspace and stay until somebody deletes them, one at a time or everything matching a filter, behind a typed confirmation.
  • Deleting a conversation never deletes the enquiry. The enquiry is a separate record with its own copy of the details, and it survives.
  • The plugin on the client's own site is somewhere data passes through, not somewhere it is stored. We hold almost nothing on your website, and erase it within a day of delivery.

The consent sentence the panel shows is yours to word, alongside the rest of the consent setup on that site.

Clinics and regulated businesses

On health and addiction websites it is stricter still

On those sites the assistant screens every message for somebody in crisis before it does anything else, answers with crisis contacts, and stops trying to sell anything for the rest of that conversation. It will not give clinical or medication advice.

Sending a visitor's message to any outside model is off by default there. It has to be switched on deliberately by the client, and most never will. Everything the assistant does on those sites happens on our own machines.

The client also cannot add their own qualifying questions at all on a health or addiction website, and clinical wording is refused, on the dashboard and again on the server. That is deliberate. The flow is allowed to be a reception desk and is never allowed to become a medical intake form.

Worth knowing

  • It answers from the website, so a site that does not say something is a site the assistant cannot answer about
  • It does not book appointments, take payments or look anything up in the client's booking system
  • It is not a CRM: no pipelines, no deal stages, no email sequences
  • Live handover needs the client to actually have agents; without them it takes a message and says so
  • We do not name the AI provider on this page, because that would tie the product to a vendor that may change
  • Anything a compliance team needs in writing belongs in the DPA, not in a badge on a marketing page
Switching it on

How to add the AI chatbot to a WordPress site

There is no chatbot plugin to install, no script to paste into the theme and no separate account for the client. If the site is already connected to WPCentrify, this is a switch and some wording.

01

Connect the website

The same connector plugin the rest of WPCentrify already uses, authenticating with a rotating scoped key rather than a WordPress administrator password. Most people connect their first site in under two minutes. A site you already manage needs nothing done to it at all.

02

Switch the assistant on for that site

Per website, never portfolio-wide by accident. It starts reading the published pages and posts straight away, so there is no knowledge base to write and no intents to configure before it can answer anything.

03

Set the wording, then watch it

Colours, launcher, heading, welcome line and consent sentence, plus which qualifying questions it may ask. Read the first conversations, add a client FAQ wherever it had to say it did not know, and the gap is closed.

Worth doing before you point a client at it

None of this is required for the assistant to work. It is the difference between a panel that answers and a panel that earns its place on the invoice.

  • Check the site actually publishes a phone number, because that is what it offers when it cannot answer.
  • Read the first day of conversations and add a client FAQ for anything it had to decline.
  • Pick the qualifying questions that suit the business, and switch off the ones that do not.
  • Decide the handover: live agents, or an honest message taken out of hours.
  • On a health or addiction site, confirm nothing is being sent to an outside model, which is how it ships.
Answers

Questions about the WordPress AI chatbot

A WordPress AI chatbot is a chat panel on your site that reads your own pages and answers visitors' questions out of them, instead of walking people through a scripted menu. The ones worth having show you where each answer came from, and say they do not know when your website does not cover it. WPCentrify's Virtual Assistant does both: most answers carry a Read more link to the page they were taken from, and it declines rather than improvises.

If the site is already connected to WPCentrify, you switch the Virtual Assistant on for that website in the dashboard and it appears. There is no second plugin to install, no script to paste into the theme, and no separate account for the client. Connecting a site to WPCentrify in the first place is one plugin and a key.

It answers from passages it finds on the client's own website, and most answers carry a Read more link to the page they came from, so any answer can be checked in one click. When the website does not state something, it says it does not know and gives the business's published phone number instead of guessing. We do not publish an accuracy percentage, because telling you when it does not know is a stronger and more honest claim than a number.

The published pages and posts of the website it is switched on for: services, treatments, prices, opening hours, FAQs and policies. It keeps reading as the site changes, so edits are reflected without anybody retraining anything. The client can also add their own FAQs and correct any answer from the dashboard.

It says it does not know, and offers a real route instead: the business's own published phone number where the website has one, otherwise an offer to take the visitor's details. It will not then ask for a phone number as though it had helped, because a qualifying question only ever rides on a reply that actually answered something.

Into Lead Captures, the same list as every contact form on the website, under the same filters, the same spam protection and the same exports. Contact Form 7, Gravity Forms, WPForms and Elementor submissions sit in that list beside chat enquiries rather than in a separate chatbot inbox.

That is one enquiry, not two. The two are recognised as the same person and merged, so nobody on the team rings them twice. Equally, one long conversation produces one enquiry that is updated as more is learned, never five separate leads.

The assistant knows only what the visitor types and the page they were on. The IP address is never stored, only a one-way fingerprint used to block abuse. Before any message is sent to a hosted AI model it is checked for anything naming a real person, and if it finds something the message is not sent at all and the answer is written from the website's own sentences on our servers. If that check cannot run, the message stays here.

Yes, and those sites run under stricter rules. Every message is screened for somebody in crisis before anything else happens, the assistant will not give clinical or medication advice, sending messages to any outside model is off unless the client deliberately switches it on, and the client cannot add their own qualifying questions at all. Clinical wording is refused on the dashboard and again on the server.

No. Nothing on the visitor's path waits on a network request to us, which is a design rule with a test enforcing it rather than an aspiration, and there is no second plugin loading on the site.

Where the client has live agents, asking for a human puts the conversation in a queue and somebody picks it up. Where they do not, or it is the middle of the night, the assistant says so honestly and takes a message rather than promising an agent who is never coming.

It is part of the platform rather than a separate subscription, so it is included on every plan on the same one price per site basis as backups, monitoring and client reports. It is free while WPCentrify is in early access, with no credit card required. There is no per-message or per-conversation charge to model separately from the rest of the plan.

No. Deleting a conversation never deletes the enquiry. The enquiry is a separate record with its own copy of the details and it survives, which the confirmation screen says before you delete anything.

Available now

Put it on a client site this week

An AI chatbot that answers from the pages you already wrote, links to every one of them, and drops the enquiry in the list you already check.

Free during early access. No credit card required.