Test an AI Support Agent on Your Website (Chatbot Preview)
Paste your URL. A crawler reads up to six pages, a model writes five questions customers ask and answers them from those pages only, or says the answer isn't on the site.
Every vendor's chatbot demo is trained on the vendor's own site, which tells you nothing about yours. The question a support lead actually has is narrower: if an agent only knew what our website says, what would it answer, and what would it get wrong? Sitting through a sales call to find out is a poor use of an afternoon, and asking the team to write test questions first means nobody does it.
This page answers it in about a minute. Paste a URL. WireDesk's crawler reads up to six high-signal pages — home, pricing, FAQ, contact, about — and converts them to Markdown. A model then writes five questions a real customer would ask and answers each one from those pages only, citing the page it used. Where the site doesn't say, the answer is "not on the site". Nothing you paste is stored; the pages are read once and the result is cached briefly so a refresh doesn't crawl you again.
This ran once, on one input. The product runs on your whole site, every conversation.
Show me answers from my whole site — start freeExample: three of five answers for a B2B invoicing platform
Q: Do you charge per user? A: No. Every plan includes unlimited users; plans are priced by the number of invoices sent per month — 50 on Starter, 500 on Growth, unlimited on Scale. Source: /pricing Q: Can I send invoices in more than one currency? A: Yes, on Growth and Scale. The pricing page lists multi-currency invoicing with automatic exchange rates on both; Starter bills in a single currency. Source: /pricing Q: What is your uptime SLA? A: Not on the site. None of the six pages read — home, pricing, FAQ, contact, about, security — mention uptime, an SLA or a status page. An agent would say so and offer to have someone follow up rather than quote a figure. Source: none
What each input means
- 01
Website URL
The homepage is the best starting point: the crawler finds the sitemap from there, then walks same-origin links. A help center on its own subdomain counts as a different site, so run that separately if it's where the answers live.
- 02
What it reads
Up to six pages, highest-signal first — about, pricing, contact, FAQ, hours, locations — then shallow paths. Pages behind a login, JavaScript-only content and PDFs are skipped. Six is enough to judge the agent; the full product reads the whole site.
- 03
How to read the result
Five questions, five answers, each with the page it came from. An answer that says "not on the site" is the tool working, not failing: it means the model refused to guess. Count those — they're your content backlog.
Questions about this tool
Five is enough to see the pattern. The model picks the questions a customer landing on your site would plausibly ask — plan limits, pricing, returns, hours, how to reach a person — rather than trivia the pages happen to contain. Run it twice and you'll get a different five; the answers should stay consistent, because they come from the same pages.
Chat with a website
Ask your own questions of the same six pages, ten messages at a time.
Website FAQ generator
Turn up to five of your pages into a question-and-answer set, with schema.
What happens when an AI phone agent doesn't know the answer?
Why "not on the site" is the behavior to test before trusting any vendor.