# Best AI chatbot for a website: how to actually choose one

> By Lawrence Dauchy, Founder, DFYe. Published 2026-09-09. 10 min read. Comparisons.
> Source: https://dfye.com/blog/best-ai-chatbot-website/
> Language: en

The most-installed chat plugin on WordPress is not AI at all, and the best free AI option is genuinely good. Both facts should change your shortlist.

**TL;DR.** There is no single best AI chatbot for a website, but there is a best free one and it is AI Engine: 100,000 installs, a 4.9 rating across 864 reviews, and only 1% of those at one or two stars. If you want a framework you finish yourself, start there and stop reading. The reasons to look further are finish, pricing shape and who holds the model key, not features.

The best AI chatbot for a website, for most people choosing one today, is **AI
Engine**: 100,000 active installs, a 4.9 rating across 864 reviews, and only 1% of
those reviews at one or two stars. It is free. If you are comfortable bringing your
own model key and building the knowledge base yourself, you can stop reading here and
go install it. Those figures come from the
[WordPress.org plugin API](https://api.wordpress.org/plugins/info/1.2/), queried on 8
September 2026, rather than from anybody's marketing page.

The reasons to keep reading are the reasons that answer stops being right, and none
of them is a feature list.

## What are you actually choosing between?

Five products cover most of the installed base, and only two of them are AI. Sorting
by install count makes the shape of this market obvious in a way that no feature
comparison does.

| Plugin | Installs | What it is | Best for |
|---|---:|---|---|
| [Chaty](https://wordpress.org/plugins/chaty/) | 300,000 | A button that opens WhatsApp | A shop where one person answers |
| [AI Engine](https://wordpress.org/plugins/ai-engine/) | 100,000 | An AI framework you finish | Developers who want the decisions |
| [Tawk.to](https://wordpress.org/plugins/tawkto-live-chat/) | 100,000 | Human live chat, free | Teams with support hours |
| [Tidio](https://wordpress.org/plugins/tidio-live-chat/) | 70,000 | Live chat with AI attached | Shops wanting both in one place |
| [Chatway](https://wordpress.org/plugins/chatway-live-chat/) | 30,000 | Human live chat | Small teams, simple setup |

The most-installed chat plugin on WordPress, by a factor of three, is a floating
button that opens WhatsApp. It answers nothing. That is not a criticism of it; it is
a reasonable product doing a clear job, and several hundred thousand site owners have
concluded that routing a question to a person beats answering it automatically.

Read that row before you compare anything else, because it reframes the question. You
are not choosing between AI chatbots. You are choosing whether to automate at all.

## Which is the best free option?

AI Engine, and not by a narrow margin. It ships updates constantly, it is liked by
the people using it, and a 1% bad-review rate across 864 reviews is not luck.

What free means here is specific and worth spelling out. You bring your own model
key, so the per-token cost is yours directly rather than marked up. You choose the
model. You build the knowledge base, write the prompt, and style the widget. The
software costs nothing and the finishing costs a day or two.

That is the right trade if those decisions are ones you want. It is the wrong trade
if you expected the thing to arrive already answering, because then you have acquired
a project rather than a tool, and the project is the part nobody budgets for.

Everything past this point is about that gap. Not about whether AI Engine is good.

## How do the ratings compare?

Better than the star averages suggest, and the useful number is not the average.

| Plugin | Rating | One or two star | What that suggests |
|---|---:|---:|---|
| AI Engine | 4.9 | 1% | Very few people had a bad time worth writing about |
| Tidio | 4.7 | 6% | Respectable, with a visible unhappy minority |
| Tawk.to | 4.6 | 9% | Widely used and widely complained about |

A star average compresses everything into one figure and hides the distribution. The
share of one and two star reviews is more useful, because that is the part of a
review page where people describe what actually broke rather than what they hoped
for. Nine percent is not a bad product; Tawk.to is used on a hundred thousand sites
and most of those people are fine. It does mean roughly one user in eleven had an
experience they wanted to warn strangers about.

Install counts deserve the opposite treatment. They measure distribution and age, not
quality, and the largest number in this category belongs to a product that answers
nothing.

## What counts as AI in these products?

Three different things, and the word covers all of them, which is the single most
common confusion when people compare this category.

The first is a **flow builder** with no model in it at all. You draw a decision tree:
if the visitor clicks Shipping, show this text; if they click Returns, show that one.
These are marketed as chatbots and they are genuinely useful for triage, because a
visitor who wants your returns window gets it in one click. They answer only what you
anticipated, and every question you did not anticipate falls through to a contact
form.

The second is a **model answering from general knowledge**. You put a language model
behind a chat window and it responds to anything, fluently, from what it was trained
on. It does not know your delivery times, your stock or your returns window, so when
somebody asks, it produces a plausible answer that is not yours. This is the category
of product that gives chatbots their reputation, and it is the one to check for
hardest.

The third is **retrieval over your own pages**. Your published content is split into
passages, the visitor's question is matched against those passages, and the model is
instructed to answer from what it was handed. This is the only one of the three that
answers questions about your business, and it is what most people mean when they say
they want an AI chatbot.

Products mix them. A flow builder with a model bolted on for the fall-through case is
a reasonable design and a common one. What matters is knowing which parts you are
buying, because the failure mode differs completely: a flow builder fails by saying
nothing useful, and a general model fails by saying something wrong with confidence.

There is a one-question test that works on any vendor. **Ask how the bot learns about
your site.** If the answer is that you upload your FAQs or train it on question and
answer pairs, it is a flow builder or a fine-tune, and it knows only what you
remembered to give it. If the answer is that it indexes your published pages and
re-reads a page when you edit it, that is retrieval. Both answers are legitimate. Only
one of them keeps up with your site.

## What should you actually compare?

Four things, none of which appear on a feature grid, because a feature grid is
written by the vendor.

**The pricing shape, before the number.** Per message means your bill tracks your
traffic rather than your revenue, so a product that goes briefly viral costs you
money. Per seat charges for human agents you may not want. A flat licence with your
own key moves the variable cost onto the model provider's published per-token price.
Ask what your worst month looks like, not your average one.

**Who holds the model key.** If the vendor holds it, every visitor question passes
through their account before reaching the model, and the list of parties with a copy
is longer than the pricing page implies. If you hold it, the vendor is not in that
path at all.

**What happens when it cannot answer.** Ask every vendor this in these words. A good
product declines below a confidence threshold and hands the question to a person. A
bad one answers anyway, fluently, from general knowledge, and you find out when a
customer quotes an invented delivery window back at you.

**Whether you can leave.** Can you export the transcripts and the knowledge base, and
in what format? A product you cannot get your own content out of is a product you have
not really bought, and an export that arrives as a wall of internal identifiers is not
an export you can use anywhere else. Ask for a sample file rather than a yes. This one
takes a single support email to test, and it is worth sending before you commit rather
than on the day you want to move.

Our own answers to all four are on [the features page](/features/), and the
side-by-side against the incumbent is on [the comparison pages](/compare/), which
name AI Engine as the better foundation for anybody who wants to build their own AI
features. Pretending otherwise would make the rest of that page less believable.

## When should you not install any of them?

Three situations, and they are not unusual.

If you get fewer than roughly ten questions a week, every product here costs you more
in setup time than it saves in answering time. An email address and a decent FAQ page
is the correct answer, and it stays correct for longer than most people expect.

If your site does not contain the answers, none of this helps. A retrieval system
reads what you published; if the delivery policy lives in your head and the returns
window lives in an email footer, the bot has nothing to find. Writing those pages
improves your site whether or not you ever install anything, and it is the
prerequisite rather than the alternative.

And if most of your questions are order status, a tracking link in the confirmation
email removes more of them than any chatbot will, at a fraction of the effort. Solve
that first and then measure what is left, because what is left is a much smaller and
much more interesting problem.

The market scan these figures come from, covering
[826,000 WordPress sites and almost no AI](/blog/wordpress-chat-plugin-market-2026/),
reached the same conclusion from the other direction. If you are still at the stage
of working out what this category even contains, start with
[what an AI chatbot for a website actually is](/blog/ai-chatbot-website/).

## Quick answers

### What is the best AI chatbot for a website?

For most people starting out, AI Engine, because it is free, mature and genuinely
well liked: 100,000 installs and a 4.9 rating across 864 reviews, with only 1% of
those at one or two stars. It is a framework rather than a finished product, so you
bring a model key and build the knowledge base. If that sounds like a project you do
not want, that is the reason to look further, not features.

### What is the most installed chat plugin on WordPress?

Chaty, at 300,000 active installs, and it is not AI at all. It is a floating button
that opens WhatsApp or another messenger, so it answers nothing and forwards the
visitor to a person. It is three times bigger than the largest AI option in the set,
which tells you how most site owners are still solving this, and it is a perfectly
reasonable answer for a small shop.

### Is a free AI chatbot good enough?

Often yes, and the honest version of that answer includes what free costs you. You
supply the model key, so the per-token bill is yours, and you finish the product: the
knowledge base, the prompt, the styling and the refusal behaviour. That is a day or
two of work if you enjoy it and an unfinished project if you do not. Free software,
paid attention.

### How do I compare chatbot pricing fairly?

Compare the shape before the number. Per message means your bill follows your traffic
rather than your revenue, so a viral week or a delivery delay costs you money. Per
seat charges for human agents you may not want. A flat licence with your own key
moves the variable cost onto the model provider's public per-token price. The shape
decides your worst month, not the headline figure.

### Do install counts tell you which chatbot is best?

No, and they are the number most easily mistaken for quality. An install count
measures distribution and age, not whether the thing works: the biggest one in this
category answers no questions at all. Review distribution is the more useful signal,
specifically the share of one and two star reviews, because that is where people
describe what broke rather than what was promised.

### Should I switch chatbots if mine is working?

Probably not. Migration costs you the knowledge base, the tuning and whatever your
team has learned about the answers, and the gain is usually a feature you could live
without. The reasons that do justify it are structural: pricing that scales the wrong
way with your traffic, no way to export your own transcripts, or a vendor who cannot
tell you which sub-processors see your data.

## Sources

- [WordPress.org plugin API, the install and rating figures](https://api.wordpress.org/plugins/info/1.2/)
- [AI Engine on the WordPress.org plugin directory](https://wordpress.org/plugins/ai-engine/)
- [Chaty on the WordPress.org plugin directory](https://wordpress.org/plugins/chaty/)
- [Tidio on the WordPress.org plugin directory](https://wordpress.org/plugins/tidio-live-chat/)
- [Tawk.to on the WordPress.org plugin directory](https://wordpress.org/plugins/tawkto-live-chat/)
- [Chatway on the WordPress.org plugin directory](https://wordpress.org/plugins/chatway-live-chat/)

---
*Published by [DFYe](https://dfye.com/). Free to read, index, quote and cite with attribution and a link.*