Chatbase alternative: AnswerWeave compared

You already know what Chatbase does, so this page will not explain it to you. It is a good product with real customers, and the reason to look at something else is not that it is bad. It is that Chatbase is no longer built for the buyer this page is written for.

Everything below is checkable. Competitor prices come from each vendor's own pricing page, dated, linked, with the configuration written out in words so you can redo the arithmetic. Claims about AnswerWeave describe mechanisms that are running in the product, not outcomes we would like you to expect.

Who each product is for now

Chatbase's homepage says "Conversational agents for customer experience." The logos under it are large, recognisable enterprise brands. The product runs across chat, email and voice, with agents meant to resolve issues end to end.

That is a deliberate move upmarket, and as strategy it is sound. The budget in this category sits with support organisations, and Chatbase has the brand, the scale proof, genuine multi-channel coverage and a mature platform to go and take it. If you have a CX team, that is a serious piece of software and we are not going to pretend otherwise.

Now read that positioning as someone with one website and no support team. Nothing in it is aimed at you. You would be buying a multi-channel agent platform to run a single widget, and paying for the parts you never open.

AnswerWeave is built for the other buyer. One website, which is where the customers come from. No helpdesk, no seats, no ticket queue, and no resolution rate anyone reports on. A question that goes unanswered there is not an unresolved ticket, it is a lost customer — so the two things that matter are that the answer is grounded in your own content and that the person asking it ends up in your inbox.

Credits versus a flat number

Chatbase bills in credits. There is a reason for that: when one product spans chat, email and voice across several models, a single abstract unit is the tidy way to price it. The tradeoff falls on the buyer. Their plans are denominated in credits rather than messages, and the pricing page does not publish a credit-per-answer conversion — so a traffic estimate does not turn into next month's invoice without help. Extra credits are sold at $40 per 1,000, which is where going over lands.

So here is one configuration, worked out in full rather than summarised: 10,000 answers a month, with the vendor's badge removed from the widget.

Our own figure renders from the live plan catalog rather than being typed into this article, so it cannot go stale the way a pasted number would. The three competitor figures are the opposite case — a dated quotation of somebody else's published page, frozen until someone re-checks it. That asymmetry is deliberate: we can keep ours current, and we cannot keep theirs current, so theirs carry a date and a link and ours does not need one.

Two honest notes on that table. The $459 is not Chatbase's list price; it is what one specific configuration costs, and at lower volume the number is far smaller. We show the 10,000-answer case because that is where a credit model and a flat plan diverge most, not because it is typical of everyone. And if you do not mind SiteGPT's branding on your widget, their $79 Growth plan is cheaper than our Growth plan. We are not going to argue with arithmetic that goes against us.

What happens when it does not know

This is the difference a competitor cannot close by editing a pricing page.

In AnswerWeave, every question is searched twice: a meaning-based vector search that finds the page answering a question worded differently, and a PostgreSQL keyword search that catches the exact product name, policy number or term the first one glides past. The two result sets are fused by reciprocal rank fusion, and the fused result carries a confidence score.

Below that threshold the assistant does not answer worse. It is handed a different system prompt entirely. The gate is deliberately two-sided: it only trips when both the meaning search and the keyword search come back weak, so one strong exact-term match will not push a good answer into the cautious template.

The low-confidence reply is not a hedged version of the confident one; it is written under different instructions. It answers with what the retrieved pages genuinely support, says plainly where they stop rather than reasoning past the gap, and points the visitor at a person — which on a website that sells something is the more valuable outcome anyway.

That trade is not free. A gate tuned to stop the assistant reasoning past your content will also hold back on questions your content covers thinly, and some of those you would rather it had attempted. We think that is the right side to be wrong on when the visitor is a prospective customer rather than a ticket. It is still a trade, and you should test it against your own pages rather than take our word for it.

What Chatbase does internally, we will not tell you, because we have not read their code and guessing would undercut the point of this page. Go and test it yourself: ask their demo something your site does not cover and watch what comes back. What we can tell you is what the category is paid to optimise for.

Citations: what is cited, and what is not attached

Showing sources under an answer is standard now. The straightforward way to build it is to list the top few retrieved chunks. That is easy, and subtly wrong: the closest matches to a question are not the same as the pages the answer was built from, so two unrelated pages ride along and lend the answer credibility it did not earn. We know that failure well — it is what our own code did until we changed it.

AnswerWeave reads the [n] markers the model emitted and resolves those to pages, up to three. Cite one page, get one source. A marker pointing at nothing is dropped rather than rendered as a dead link. When an answer carries no marker at all — smaller models often do not mark their sources — you get the single best-ranked page, not three.

And on a low-confidence reply, the assistant stops attaching a source it was not given. When the answer cites nothing, we do not fall back to the best-ranked page and put a link under it: a sentence saying your content does not cover the question is not made truer by a source beneath it. This sounds like a small thing. It is the proof that the gate is real: a refusal arrives with nothing underneath it. Sources under a good answer are easy to demonstrate, and everyone does. The screenshot worth asking for is the other one: what sits under the answer when the content does not cover the question.

Side by side

When Chatbase is the better choice

A comparison page with no such section is an advertisement. These are real, and if two of them describe you then you will get more out of their trial than ours.

• You have a customer-experience organisation. Seats, queues, escalation paths, and someone whose job is measured on deflection. Chatbase is built around that now. We have deliberately built none of it, and adding it would make us worse at the thing we are good at.
• You need more than the website. Email and voice are real channels in their product. We do one thing — the assistant on your site — and we would rather say so than sell around the gap.
• You want the assistant to take actions, not just answer. Agent workflows that look something up, process a return, or hand off mid-conversation are a different product from a grounded answer with the right page under it.
• You already run a helpdesk stack and the assistant has to live inside it. Fitting into an existing support toolchain is a platform problem, and platforms win platform problems.
• Your buying process needs a large, recognised vendor. An enterprise logo wall does real work for a buyer who has to justify the decision internally. We are early, and our production footprint is small. You should hear that from us rather than find it out later.

Moving over

There is almost nothing to export. The knowledge base rebuilds itself from your own content, so nothing there is trapped. The exception is history: conversation transcripts and any leads captured in your current tool cannot be regenerated from your website, so pull those out before you close the account.

1. Point AnswerWeave at your site — a crawl, a sitemap or a list of URLs. Add PDFs, Word documents, Markdown or plain text for the things that were never published: price lists, policies, the FAQ that lives in somebody's inbox.
2. Wait for the first crawl, then ask it the ten questions your visitors actually ask. Make two of them questions your site does not cover. What it does with those two is the entire evaluation.
3. Paste one script tag before the closing body tag. No plugin, no theme edit, nothing to change in the pages you already have.

<script src="https://answerweave.hazentech.com/widget/widget.js"
  data-bot-id="YOUR_BOT_ID"
  data-api-key="YOUR_PUBLISHABLE_KEY"
  data-api-url="https://answerweave.hazentech.com"
  data-color="#1d4ed8">
</script>

Copy the real one from the bot screen rather than from this page. It comes out with your own bot id and publishable key already filled in, and it always carries the host that is answering today — which is the host that matters, because a wrong snippet fails on your site rather than ours.

Nothing stops you running the two in parallel for a week — one on a section of the site, one on the rest, or simply swap the tag and swap it back. The free plan needs no card, so a real comparison costs you a script tag and an afternoon rather than a procurement cycle.

Questions

Is AnswerWeave a drop-in replacement for Chatbase?

For a website assistant that answers from your own content and captures leads, yes — you re-point it at your site, paste one script tag, and remove theirs. For multi-channel agents across email and voice, or agents that take actions in other systems, no. We do not build those and are not planning to.

Why does this page say $459 when Chatbase advertises $120?

Because $120 is the Standard plan and it covers 4,000 credits. At 10,000 answers a month you buy 6,000 more credits at $40 per 1,000, which is $240. Removing Chatbase branding from the widget is advertised on their page at $1,188 a year, which is a further $99 a month spread across twelve. That is $459 for that specific configuration, counted at one credit per answer — they do not publish a credits-per-answer figure, so that is the assumption most favourable to them, and a higher rate would raise the total rather than lower it. All three figures come from their own pricing page, checked 31 July 2026. At lower volume the total is much lower, and you should check the page yourself before deciding anything.

What does AnswerWeave do when it cannot answer a question?

It says where your content stops. Every answer carries a confidence score based on what the search actually retrieved. When that score is below the threshold and the keyword search has not turned up a strong match either, the assistant is given a different system prompt: it answers with what the retrieved pages genuinely support, says plainly where they stop rather than reasoning past the gap, and offers to pass the question to you. It also stops attaching a best-guess source to that reply. There is nothing to cite, so nothing is cited.

Is there a fee to remove AnswerWeave branding?

No. There is no branding-removal fee on any plan, because the widget does not carry our badge to begin with. It is the kind of charge that lands after you have already committed — Chatbase's own page prices branding removal at $1,188 a year, on top of whatever plan you chose.

Do I need a developer to switch?

Only for the script tag, and only if you cannot edit your site's template yourself. Everything else — adding the website, uploading documents, changing the assistant's instructions — happens in the dashboard.