People usually worry that an AI chatbot will refuse to help. In our experience, the more dangerous failure is the opposite. The chatbot helps too much. It finds something nearby and stretches it into an answer. Every word sounds plausible, and the customer has no way to tell which parts are true.
Most of the work on answer quality over the past months has gone into that failure. Gydr 4.1 is the latest step: when nothing in your knowledge base answers a question, your chatbot now says so plainly, instead of offering an unrelated FAQ or products the visitor never asked for. This piece explains how we think about chatbot hallucinations, and why helping too much is the failure that matters.
Grounded is not the same as right
A common way to make AI assistants trustworthy is to "ground" them: retrieve something from the business's own content and answer from it. Grounding is necessary. It is not sufficient.
Retrieval almost always finds something. Ask a bicycle shop whether it sells electric scooters and a search will happily return bicycle accessories, because they are the closest thing in the catalogue. An assistant that treats "I found something related" as "I found the answer" will recommend a bike light to someone who wanted a scooter, politely and with a link.
So we treat two questions separately. Does the evidence exist? And does it actually answer what was asked? Only a yes to both earns a confident answer. Weak or unrelated evidence creates no obligation to cite it, and it is never presented as though it were the answer.
An empty field is not an invitation
Catalogues have gaps. A product row may have a name and a price and nothing else. To a language model, that blank description looks like a space to fill, and it will fill it fluently.
We have seen how this goes wrong. A product whose description field was empty was described, confidently, with ingredients and a benefit that appeared nowhere in the catalogue. Every sentence named a real product at its real price, so a simple check of "did it mention something we sell?" passed.
Since Gydr 3.3, an item with no description is presented by name and price only. Since 3.6, the chatbot no longer writes its own line under a product whose entry is just a name and price. What the business did not say, the assistant does not say.
Numbers are checked before they are quoted
Prices, phone numbers and stock levels are where a confident mistake costs real money. Since Gydr 3.2.8, your chatbot checks every figure it is about to quote against your content and will not state one it cannot support. Since 3.6, "In stock" or "Sold out" appears only when your catalogue or store says so.
The cards follow the words
When a reply recommends products, the product cards beneath it should show exactly those products: no more, no fewer, no different. It sounds obvious. In practice the reply and the cards were often built separately and drifted apart.
Gydr 3.0 made the cards follow the reply, and Gydr 4.1 tightened this again: each recommended product is described once, and only the products the reply actually names appear with it.
Read the whole source
The opposite mistake matters too. A system that reads only the first part of a long document can tell a customer, with complete confidence, that your business does not carry something that appears two pages later. Since Gydr 3.6, answers drawn from a long document use the whole relevant passage, so the chatbot never denies something your own document says.
What we believe
Saying "I don't have that information" is not a failure of an AI assistant. Inventing something is. A business can recover from a chatbot that occasionally hands a question to a person. It is much harder to recover from one that confidently tells customers things that are not true.
So we hold Gydr to a simple standard. Your knowledge is the authority. The assistant may summarise it, translate it and present it well. It may not add to it.
Frequently asked questions
Why do AI chatbots hallucinate?
Search nearly always finds something related, and a language model fills gaps fluently. Treating a related item as the answer, or a blank field as space to write, produces confident claims your content does not support.
How does Gydr keep its chatbot from making things up?
It answers from your knowledge and says so plainly when nothing answers the question. A product with no description is given by name and price only, figures are checked before they are quoted, and stock status comes only from your data.
Is "I don't have that" a bad answer?
Not when it is true. A customer can recover from a chatbot that admits a gap; it is much harder to recover from one that confidently says something false.
The full list of changes is in the 4.1.0 release notes.