How it works
A large language model (LLM) writes its answer word by word, picking the most likely continuation at each step. It doesn't check what it writes. When information is missing, it fills the gap with something plausible instead of saying it doesn't know.
The most common causes:
- Missing data: the product page doesn't mention compatibility, so the model assumes it.
- An ambiguous question that the model interprets its own way.
- Outdated knowledge: the model answers from what it learned in training, not from today's catalog.
To limit hallucinations, you ground answers in verified sources (catalog, knowledge base), instruct the model to answer only from those sources, and give it a way out when it can't find the answer: say it doesn't know, or hand over to an agent.
Why it matters in e-commerce
On a retail site, a made-up answer quickly turns into an order. The customer buys based on what the assistant said, receives a product that doesn't match and sends it back. That means a return to process, an unhappy customer, sometimes a dispute, and trust in the assistant that doesn't come back.
A hallucination also commits the brand. A delivery promise, a discount or a feature announced by the assistant reads to the customer as the seller's word.
Further reading: It’s Time to Fire the Chatbot
Real-life example
A shopper asks whether a frying pan works on an induction cooktop. The product page doesn't say. A poorly grounded assistant answers "yes, no problem." The customer orders, finds the pan doesn't heat and returns it. An assistant grounded in the catalog says it can't find that information, suggests two pans whose product pages confirm induction compatibility, or passes the question to an agent.
Common confusion
Not every AI mistake is a hallucination. If the product page lists the wrong weight and the assistant repeats it, it hasn't made anything up: it has faithfully passed on bad data. In that case, fix the source, not the model.