A chatbot hallucination is an answer that sounds confident but is not supported by your own facts. It might be an invented discount, a return window that does not exist, or a feature you never built. The good news is that most hallucinations have an ordinary, fixable cause, and you can usually find it without a machine learning team.
This playbook walks through how to collect hallucinations, trace each one back to its cause, and fix the right layer instead of guessing.
What a hallucination looks like in practice
Customer: Do you offer a student discount? Chatbot: Yes! Students get 20% off all plans with a valid student email.
If you have never offered a student discount, this answer is a problem. The customer may sign up expecting 20% off, then contact support angry when it does not appear. In the worst case, a regulator or tribunal may treat your bot's promise as your promise. In Moffatt v. Air Canada (2024), a tribunal held the airline responsible for incorrect fare information its website chatbot gave, and rejected the idea that the chatbot was responsible for its own statements.
Hallucinations are rarely random. They tend to cluster around the same topics: pricing, policies, availability, and anything the bot was asked about but never given information on.
Step 1: Collect real examples before changing anything
Do not start by rewriting the system prompt. Start by building a list of wrong answers, each with:
- The exact customer question
- The bot's full answer
- What the correct answer should have been
- Where the correct answer lives (a page URL, a PDF, a policy document), or a note saying it lives nowhere
That last column is the most important. It tells you which kind of fix you need.
Step 2: Sort each hallucination by cause
Almost every hallucination falls into one of four buckets.
- Missing information. The answer is not in anything the bot can see. The model fills the gap with something plausible. Fix: add the information, or tell the bot to say it does not know.
- Retrieval miss. The answer exists in your content, but the bot did not find it when answering. Fix: improve how content is chunked, titled and searched. See when retrieval fails.
- Conflicting sources. Two pages say different things, and the bot picked the wrong one or blended them. Fix: correct or retire the outdated page.
- Prompt pressure. The system prompt pushes the bot to always be helpful, always give an answer, or always sell. Fix: change the instructions.
Write the bucket next to each example. Patterns usually appear quickly. If most of your problems are bucket one, no amount of prompt tweaking will help.
Step 3: Fix the source first, the prompt second
Content fixes are the most durable because they help every future answer, not just the one you tested. For a missing fact, add a short, explicit statement to your knowledge base:
Before: "We have flexible pricing for different customer types."
After: "We do not currently offer student, nonprofit or educational discounts. All customers pay the prices listed on the pricing page."
Notice the "after" version states a negative clearly. Models are much better at repeating a clear statement than at inferring an absence.
Then add a grounding rule to the system prompt, if you do not already have one:
Answer only using the information in the provided context.
If the context does not contain the answer, say you are not sure
and offer to connect the customer with the team. Never invent
prices, discounts, dates, policies or product features.
This kind of instruction does not eliminate hallucinations on its own, but combined with good content it removes the pressure to make something up.
Step 4: Narrow what the bot is allowed to promise
Some topics are high risk: refunds, legal terms, medical or safety questions, and anything involving money. For these, consider a stricter pattern: the bot quotes or links the official policy instead of paraphrasing it, or it hands off to a person. A bot that says "Here is our refund policy page" is less impressive than one that explains it, but far less likely to cost you money.
Step 5: Re-test the same questions, and more
After each fix, ask the exact same questions again, plus a few variations. Customers do not phrase things the way you do. If you fixed "Do you offer a student discount?", also try "Is there a discount for college students?" and "Any cheaper plan for schools?"
Keep these questions as a regression set. Hallucinations have a habit of coming back after a model upgrade, a website redesign or a prompt change.
A checklist you can run today
- Ask your bot ten questions about pricing, refunds and policies, and write down every answer
- Mark each wrong answer with one of the four causes
- Add clear, explicit statements for missing facts, including the things you do not do
- Remove or correct outdated pages that contradict current policy
- Add a grounding rule and an "I don't know" path to the system prompt
- Re-test with the original questions plus rephrased versions
- Save the questions so you can re-run them after every change
When to get help
If your list of wrong answers is long, or you cannot tell which bucket a problem belongs in, it is worth getting a second pair of eyes. ChatLens360 reads your website and documents, asks your chatbot realistic customer questions, and flags every answer that contradicts your own facts, with the source page and a ready-to-paste fix prompt for each issue plus suggestions for what to add to your knowledge base. You can apply those fixes yourself, or talk to us and have the team fix them for you. Either way, start by running a free chatbot audit.