Why most website chatbots give wrong answers — and how Kothok avoids it

It’s 11pm. A visitor is on your site, wallet out, and types one last question before they buy. What your chatbot says next either closes the sale or sends them to a competitor. So why do so many chatbots blow that moment — confidently giving an answer that’s just wrong? And what does Kothok do differently? Let’s follow a single real question all the way through.

The question

A shopper on a Dhaka-based store asks, half in Bangla, half in English — the way people actually type:

Why the typical bot gets it wrong

There are two common kinds of website chatbot, and each fails this moment in its own way:

• Scripted botsonly know the exact question-and-answer pairs someone typed in advance. Ask anything slightly off-script — two questions at once, a different phrasing — and they fall back to “Sorry, I didn’t understand that.”
• Generic AI botsare fluent but ungrounded. They’ll happily answer anything, including things about your business they were never told — so they invent plausible-sounding details. Fluent and wrong is worse than “I don’t know.”

How Kothok stays grounded: retrieval before it answers

Kothok uses retrieval-augmented generation (RAG). In plain terms, it doesn’t answer from memory — it looks things up first. When the question arrives, Kothok searches the knowledge base it built from your own content during onboarding (your pages, PDFs, help-centre articles), pulls the passages that actually cover shipping and returns, and only then writes a reply based on that material.

That’s the whole trick behind the good answer above: it quoted your delivery zones and your 7-day window because it read them, not because it guessed. And when your policy changes, you refresh the knowledge base — you never rewrite a script of canned replies.

When it doesn’t know, it says so

Grounding also means Kothok can tell when your content doesn’tcover something — and say “I’m not sure” instead of improvising. For businesses where a wrong answer is expensive — fintech, healthcare, anything regulated — Strict Modetightens this further: answers are limited to verified sources and backed by citations, so the agent only states what your documents actually support. That honesty is the feature. It’s what makes Kothok safe to put in front of customers where accuracy is non-negotiable.

Speaking the language your customers actually use

Notice the shopper switched between Bangla and English mid-sentence. For a lot of South-Asian businesses that’s the norm, not the exception — and a bot that only handles formal English quietly loses those customers. Kothok understands English, Bangla, and Banglish out of the box and replies in kind, so grounded answers actually land instead of bouncing off a language barrier.

And it’s easy to put to the test

The best part: you don’t have to take any of this on faith. The widget is lightweight and installs with a single script tag, and there’s a genuine free tier — so you can point Kothok at your own site, ask it your own trickiest question, and watch how it answers before you ever pick a plan. The same grounded engine adapts across e-commerce, SaaS, and fintech, which is why the approach holds up whatever you sell.

A guess dressed up as an answer costs you the 11pm sale. An agent that reads your content, admits what it doesn’t know, and meets customers in their own language earns the trust that closes it. That gap — between guessing and knowing — is the whole reason Kothok is built the way it is.