There’s an instinct, when you install a chatbot, to want it to answer everything. A reply feels like success and a decline feels like failure.
It’s the wrong instinct. On your own website, an assistant that answers every question is not more capable than one that sometimes declines. It’s just less honest — and you’ll find out which, eventually, from a customer who acted on something it made up.
In short
A confident wrong answer costs more than an honest gap. IrisAI checks match quality before generating, so the failure mode is “I don’t know” rather than invention.
What a confident wrong answer costs
Language models don’t signal uncertainty the way people do. A guess arrives in the same fluent, well-punctuated prose as a fact. There’s no hedging in the tone to warn the reader.
So when an assistant invents a 30-day return window for a shop that offers 14, the customer doesn’t sense the difference. They plan around it. You find out at the point of complaint, and the conversation starts from a position where your own website told them something untrue.
Compare that with: “I don’t have details on the return window — the contact page will get you a definite answer.” Slightly less impressive. Considerably cheaper.
Why “answer from the content” isn’t enough on its own
The obvious fix is instructing the model to use only your content. It helps, but it doesn’t hold, because of how retrieval behaves at the edges.
Search always returns its best candidates, even when the best candidates are poor. Ask about something your site never covers and you’ll still get passages back — the closest ones available. Hand a model weakly related text and a question, and it will do what it’s built to do: produce a plausible continuation, bridging the gap between what it was given and what it was asked.
The instruction is followed in spirit. The answer is still wrong.
Deciding before generating
The fix is to check match quality before the model is involved at all. IrisAI scores what retrieval found, and if nothing clears the threshold, the assistant is told explicitly that no matching content exists — not to answer from general knowledge, and to point at your contact page instead.
Two details matter. The gate runs before generation, so the model never gets the chance to improvise. And no sources are attached to a decline, because none were used — nobody gets a citation pointing at a page that doesn’t contain the answer.
Tuning the threshold
Set it too low and near-misses get answered — the failure this exists to prevent. Set it too high and the assistant declines things it should have handled, which is frustrating in a different way.
The defaults are calibrated against real sites, and both the similarity thresholds and the decision itself are filterable if your content is unusual. The Search Eval screen shows the actual scores behind each decision, so tuning can be done from evidence rather than guesswork.
Declines are a content report
Here’s the part worth acting on. Every question your assistant declines is a question a real visitor asked that your website doesn’t answer.
That list is unusually valuable. It isn’t keyword research or a guess about intent — it’s people telling you, in their own words, what they wanted and couldn’t find. Review it periodically, write the missing content, and the declines turn into answers.
An assistant that declines gracefully doesn’t just avoid harm. It tells you what to write next.
Every question your assistant declines is a question a real visitor asked that your website doesn’t answer.
Give your visitors an assistant that tells the truth.