Chatbot Not Answering Correctly? A Diagnostic Checklist
Wrong answers, missing answers and confident nonsense have different causes and different fixes. Work through them in this order rather than changing settings at random.
The short answer
- Identify the failure type first. Saying nothing, saying the wrong thing, and answering a different question are three separate problems with three separate fixes.
- The most common cause of "I don't have that information" is not the AI but the content — pages that indexed as empty because they render in JavaScript or are scanned images.
- Confidently wrong answers usually mean outdated content is still indexed and still being retrieved, not that the model invented something.
- Fix the source content and re-index before touching model settings. Changing the model rarely fixes a retrieval problem.
Start by naming the failure
Support teams describe every one of these as the chatbot not working, but the fixes have nothing in common. Before changing anything, reproduce the problem and put it in one of four buckets.
- It refuses — the bot says it does not have that information, for a question your site clearly answers.
- It is wrong — the bot answers confidently with something incorrect or out of date.
- It answers a different question — the response is fluent, on-topic, and not what was asked.
- It is vague — the answer is technically accurate but so hedged it helps nobody.
When it refuses to answer
This is the most common complaint and, encouragingly, usually the easiest to fix. In the large majority of cases the content was never indexed, even though the dashboard reported a successful sync. A crawler that fetches a page containing no readable text records a success and stores nothing.
- 1Check whether the page indexed with actual content, not merely that the crawl succeeded. A page count is not a word count.
- 2View the raw page source rather than the rendered page. If your text is not in the HTML, the content is injected by JavaScript and the crawler never saw it.
- 3Check any PDFs for a text layer by trying to select text in them. If you cannot, it is a scan and needs OCR.
- 4Check whether the answer only exists inside an image, which is common for pricing tables and comparison charts.
- 5Search your knowledge base for the exact words a customer would use. If your page says fulfilment and they say delivery, retrieval may genuinely miss.
- 6If the page cannot be crawled reliably, paste the text in directly as a knowledge source instead.
Pasting text directly bypasses the crawler entirely and indexes the same way, which makes it the dependable fallback for JavaScript-heavy sites and scanned documents.
Read the knowledge base guideWhen it answers confidently and wrongly
The instinct is to blame hallucination. In a properly grounded system that is rarely what happened. Far more often the bot retrieved a real passage from your own content and answered from it faithfully — and that passage was out of date.
Old pricing pages, superseded policies, last year's campaign landing page and an archived blog post that mentioned a discontinued feature are all still, from the retrieval system's point of view, perfectly good sources. It has no way to know which of two contradictory passages you would prefer.
The fix is unglamorous. Find the contradicting source and remove it, or correct it. Then re-index and retest the exact question. Deleting a page from your website does not remove it from the index until you re-crawl.
When it answers a different question
This is a retrieval precision problem. The question and the retrieved passage were similar enough numerically to be considered a match, but not similar enough in meaning to be useful. It happens most often when one topic is spread across many pages that all mention it in passing, so no single passage is clearly the best.
It also happens when a passage is too long. If a section covers shipping, returns and warranty in one block, it will be retrieved for all three and answer none of them cleanly.
- Consolidate the authoritative answer onto one page and remove the passing mentions from the crawl.
- Split long mixed sections into shorter ones, each under a heading that names its single topic.
- Add the customer's phrasing to the page explicitly, including the terms your industry uses casually.
- Where two products have similar names, state the distinction in plain words on both pages rather than assuming context makes it obvious.
When answers are vague or over-hedged
Usually this reflects the source rather than the bot. Marketing copy is written to be broadly appealing and specifically non-committal, and a bot trained on it answers the same way. If your shipping page says delivery is fast and reliable rather than three to five working days within the UK, the bot has nothing concrete to offer.
Go through the questions your team gets most and check that each has a specific, numeric, unambiguous answer somewhere in the indexed content. This is often the single highest-return hour of work in the whole process.
Fix in this order
Working through the cheap, high-yield causes first prevents the common mistake of changing model settings to compensate for a content problem, which hides the symptom without fixing anything.
- 1Confirm the content is genuinely indexed, with text, not merely crawled.
- 2Remove or correct any outdated source that contradicts the right answer, then re-index.
- 3Rewrite the relevant sections to be self-contained, specific and headed by their topic.
- 4Add the customer's own vocabulary to the page.
- 5Only then adjust confidence thresholds or escalation rules, which change when the bot answers rather than what it knows.
- 6Review unanswered questions weekly and treat that list as a content backlog, because that is what it is.
What good looks like
A healthy deployment does not answer everything. It answers the routine reliably, declines cleanly when it should, and hands over with enough context that the person picking up the conversation does not start from nothing.
If your bot never says it does not know, that is not a sign of quality. It is a sign that it will eventually invent something at the worst possible moment.
Frequently asked questions
Why does my chatbot say it doesn't have information that is on my website?
The page most likely indexed as empty. The three usual causes are content rendered by JavaScript and therefore absent from the raw HTML, PDFs that are scans with no text layer, and information that exists only inside images. A successful crawl reports the page was fetched, not that any text was stored — check the indexed word count, not the page count.
Why does my chatbot give outdated answers?
Because the outdated page is still in the index. Retrieval cannot tell which of two contradictory passages you consider current. Remove or correct the old source — an archived campaign page, a superseded pricing page, an old blog post — and re-index. Deleting a page from your site does not remove it from the index until you re-crawl.
Is a wrong chatbot answer always hallucination?
Rarely, in a properly grounded system. Far more often the bot retrieved a genuine passage from your own content and answered it faithfully, and that passage was wrong or out of date. Check what your content actually says before assuming the model invented anything.
Should I change the AI model to improve answers?
Almost never as a first step. Most quality problems are retrieval problems — the right passage was not found, or a wrong one was. A more capable model cannot answer from a passage that was never retrieved. Fix the content and re-index first; model changes rarely fix content gaps.
How often should I review chatbot answer quality?
Weekly at first, then monthly once it stabilises. The single most useful artefact is the list of questions the bot could not answer, which is an honest description of the gap between what customers want to know and what your website says.
