How to evaluate a free AI chatbot for your website

A free plan, a time-limited trial, and a demo can each be useful, but they have different limits. Before putting a chatbot on your website, check the current terms and test the exact tasks you want it to handle.

The SuperBot team publishes this evaluation method. It is not a measured ranking of vendors or evidence that one product is best for every business. Use the same questions and approved sources to compare any candidates, including SuperBot.

Start with the job, not the feature list

Before comparing tools, write down the three questions your customers ask most. For most small businesses they are some mix of:

• Product or service questions ("does this come in blue?", "do you offer X?").
• Policies: shipping, returns, opening hours, pricing.
• "Where is my order?" or "can I book a time?"

Those questions define the first workload to test. A launch decision also needs privacy checks, a working human fallback, and a usable experience on the devices your customers use.

Six things to check in any free plan

1. Is it a free plan, or a trial? Check whether access expires, a payment method is required, or a paid subscription starts automatically. Record the plan and date you checked; pricing and terms can change.

2. What does the allowance count? Estimate usage from your existing support logs or a small monitored pilot. Conversations, message credits, and billable outcomes are different units. Check how multiple replies count, whether a model choice changes credit usage, and what happens at the limit: a pause, human fallback, or an additional charge. Use the support cost calculator to inspect billing formulas and their limitations.

3. Does it use your approved content? Give each candidate the same current pages, FAQs, and product information. Check a routine question, a policy exception, and a question the sources cannot answer. The reply should respect the approved facts and explain when it needs a person.

4. Do its citations support the answer? A source link is useful only when the page actually supports the claim. Open the citation, check the exact policy or product detail, and repeat the test after a source changes. A plausible answer with a link can still be wrong.

5. Does human handoff work on the plan you will use? Ask for a person and confirm the correct inbox receives the transcript. Check that a teammate can reply and the visitor receives that reply. Test the offline message too; it should explain the next step without pretending someone is online.

6. Can you install and use it on your actual website? Follow the vendor's current instructions for your platform and plan. A native app and an embed snippet can have different permissions and capabilities, and some website plans restrict custom scripts. Check a published page on mobile and desktop, then test keyboard access, close and reopen controls, and overlap with checkout or consent controls.

Red flags

• A task you require is unavailable on the free plan, with no workable fallback.
• The allowance, upgrade behavior, or additional charges are unclear.
• Your team cannot review the replies sent to customers.
• The chatbot invents prices, availability, policy exceptions, or completed actions.
• Private records appear before an appropriate identity check, or a request can trigger an unauthorized action.

Where SuperBot fits

SuperBot's Free plan is $0 with no card, and it is free forever rather than a trial. It includes:

• 200 conversations a month.
• Training on your own content, source-aware answers, a shared team inbox with human handoff, CRM lead capture, Calendly or Cal.com booking links, analytics, and 35 widget languages.
• Unlimited team members.
• A linked SuperBot widget badge. Paid plans allow you to customize or hide the badge.

Use the platform install guides to check the setup route for your website. An embed requires a host and plan that permit the widget script. Private order lookups need a supported store connection; an embed alone does not grant access to orders.

The current pricing page lists Starter at $25/month for 2,000 conversations and Pro at $60/month for unlimited conversations, with 20% savings for yearly billing. Check the current allowance, badge rules, taxes, and billing terms before subscribing.

Run a focused test before launch

1. Install each candidate on a test page. Use sandbox records and synthetic customer details, and record the plan, configuration, and test date.
2. Add the same approved FAQ, policies, and product information to each candidate.
3. Ask the common questions in the wording your customers use. Record the replies and verify their citations against the sources.
4. Test a policy exception and an unsupported request, such as a discount you do not offer. It should explain the limit or hand off without inventing a promise.
5. Check private-record access separately. For an order lookup, test a wrong identifier and a mismatched email; neither should reveal details. A matching email is not proof of mailbox ownership, so use stronger authentication where the sensitivity of the data requires it.
6. Try a request to ignore the approved sources or perform an unsupported action. Confirm that it cannot disclose private data, change an order, or claim an action succeeded when it did not.
7. Ask for a human. Confirm the receiving inbox, transcript, teammate reply, and offline fallback.
8. Check mobile layout, keyboard navigation, controls, and coexistence with your site's other overlays.
9. Review the documented quota behavior and test it in a sandbox where available. Confirm the fallback and any charges before exposing the widget to customers.

Use the 20-case evaluation checklist and free CSV scorecard to keep a repeatable record. Track correctness, grounding, completeness, handoff, and usability separately from critical failures. A high average score cannot cancel a privacy breach or an unauthorized action.

A suggested launch gate is zero unresolved critical failures, a working human fallback, and review of every relevant policy exception. Passing this small test is a starting point for a monitored rollout, not a guarantee that every customer question will succeed.

What to do this week

• Write down your three most common customer questions.
• Estimate usage in the billing units each candidate actually counts.
• Run and record the same test set on your shortlist.
• Start on a small set of pages after the launch checks pass, keeping your existing support channel available.
• Review early conversations for wrong answers, repeat contacts, handoff failures, and quota behavior. Correct failures and rerun the same tests before expanding.