Chatbase on the free plan: we fed it our own site and asked it twenty questions we already knew the answers to

How we tested

A support bot has one job, and it is easy to fake in a demo. So we did not ask Chatbase to be impressive. We asked it to be right about things we could check.

Before the account existed we froze a fixture: one agent trained on okaneland.com by pointing the crawler at the bare domain, then twenty factual questions whose answers live on specific pages of this site, plus three trap questions about things the site does not cover. Each answer was graded exact, partial, refused or invented against the key. The gates were fixed at the same time: a live agent at $0 with no card, how much of the site the free tier would hold, at least fifteen of twenty exact, whether answers name their source, all traps refused, credits consumed per message against the meter, latency, the free-tier friction, the widget on a page we control, and what the terms say about your data.

Everything ran on the Free plan through the dashboard, on the default Auto model, with the instructions Chatbase wrote for us during onboarding left exactly as written. One disclosure: the site it was trained on is ours, which is the point. We know what every page says, so we can grade to the sentence.

The door

Signing in takes a Google account. Then six screens of survey, where you heard about Chatbase, company size, your website, what the agent will do, your tech stack, where you will deploy, and a pricing page you leave through a small “Continue for free” link at the top right. The crawler found 107 links on okaneland.com from the domain alone and listed them by section for review. Seven minutes after the dashboard first loaded, the agent was live.

It arrived with a banner: “Limit reached! You’re using 1 MB of 1 MB in your plan.” The free tier holds one megabyte of training content. The crawler had trained on 80 of the 107 links, failed on five non-page assets, and dropped 22 for size.

Which 22 is the part that matters. The cap does not rank pages. It keeps whatever the crawler reached first and drops the rest, and the rest included the homepage, the subscribe page, the start-here page, four studies and the Open Design review. Two of our twenty questions had their source page in that pile. We graded them anyway, because a reader on this plan would meet the same hole.

The model picker lists 36 models. Three are selectable at $0: Auto, GPT-5.6 Luna and GPT-4o Mini. Every Claude, every Gemini, and the rest of the OpenAI line are greyed with a trial button beside them. Auto costs one credit per message.

The twenty questions

We asked them one at a time in the playground, timed each reply from send to the moment it stopped changing, and read the usage page after.

Eighteen exact. It knew Codex on Free passed twelve of twelve, picked GPT-5.6 Terra and spent 37 percent of its meter. It knew Cursor Pro and Claude Code tied at eleven and which gate they both failed, and described the double booking correctly. It had the Fable price, the two hardware keys, the 2,122 and the 1,024 posts, the three cuts in eight months, the Clay and Creatify free tiers to the credit, and that Akira Sumi is a pen name. Asked what The Proof is, it said hands-on reviews that show what works, what it costs and where it breaks, which is close enough to the masthead that we would have signed it.

The partial was our fault as much as its own. Asked how many servers the MCP tax study measured, it said eleven. The article says ten in prose and lists eleven rows in its table. The agent counted the table.

The refusal was the plan’s fault. Asked what our two-run Open Design session cost, it said the excerpts did not contain that figure, then offered a real but unrelated number from a Studio interview. The page with the answer was one of the 22 the cap dropped. It did not invent the figure, which is the behaviour you want from a bot that cannot see the page.

The traps came out two for three, and the third was ours. It refused to produce a phone number and volunteered the public contact email instead. It refused to name a crypto exchange and explained that Coinbase appears on the site only as a cited source. Then it answered the Windsurf question in detail, and it was right to: the field guide covers Windsurf’s fate and Devin Desktop’s prices, and we had forgotten. The fixture is published with that error left in.

The meter

Measure	Result
Messages used	25 of 50
Credits per message	exactly 1
Median time to a finished answer	4.8 seconds
Slowest answer	16.1 seconds
Dashboard to live agent	7 minutes
Dashboard to graded results	26 minutes

The credit meter is the plainest thing about the plan. Twenty-three playground questions, one from the embedded widget, one more for a screenshot: twenty-five messages, twenty-five credits, and the usage page agreed to the unit. Fifty a month is under two a day, which is a number for testing rather than for a site with visitors.

Where it breaks

One megabyte is smaller than a small site. Okane Land is around a hundred pages of text. The cap took 80 of them and left the front door outside. Anything you sell will be documented on more pages than that, and the crawler will not tell you which ones it dropped until you open the fetched-links panel and click the “22 overlimit” tab.

Visitors do not get the sources. In the dashboard every answer has a “Show sources” control that opens the page URL and the passage it used, which is how we checked our grades. Switch the channel on and embed the widget, and that control is gone. The visitor gets the answer, a thumbs up, a thumbs down and “Powered by Chatbase.”

The channel is off until you switch it on. The embed code is withheld behind a “Channel disabled” notice until you enable the chat bubble, which is sensible and undocumented. Once on, the bubble loaded on a local page and answered correctly. We switched it off again afterwards.

The instructions are written for you, and not obeyed. Onboarding generated 1,400 characters of instructions from the site, including “Never use em dashes in your replies.” The twentieth answer used one. The same instructions tell the agent to refuse anything outside the business, which is why the traps worked; edit them before you trust them.

Fourteen idle days and the agent is gone. The pricing page says free agents are deleted after 14 days of inactivity. A test bot you build on a Sunday and forget is not there the Sunday after next.

Real cost

Free is $0 and did everything above. Hobby is $40 a month for 700 credits, 40 MB of sources, the advanced models, integrations and attachments. Standard is $150 and adds the API. Chatbase’s terms call fees non-refundable except after an uncured breach, and its privacy policy says your data is not used to train AI models, which is the sentence to look for before you upload a knowledge base to anyone.

The verdict

Situational, and the situation is clear. As a way to find out whether a bot trained on your own writing will answer correctly and shut up when it should, the free plan is the best $0 we have spent on a support tool: eighteen of twenty to the number, both real traps refused, nothing made up, five seconds a reply. As a thing to put in front of customers, the free plan is a demo. One megabyte will not hold your docs, fifty messages will not hold a week, and the visitor never sees where an answer came from.

Two things to do before you pay. Open the fetched-links panel and read the overlimit tab, because that list is what your bot does not know. And rewrite the instructions the onboarding wrote for you, because the agent is following them where it matters and ignoring them where it does not.