A seller wrote to us last month to say the assistant was broken. I asked what happened. A customer had asked about the return window, the assistant guessed, the guess was wrong, and the seller switched it off that afternoon. One bad answer, one switch.
I hear some version of that story a lot, so we went and looked at the knowledge base behind 11,878 assistants to see what separates the ones sellers keep from the ones they quietly shut down. It is almost never the intelligence. It is what the assistant was handed, or never handed.
What the data says
| Knowledge base trait | Value |
|---|---|
| Median knowledge base size | ~4,500 characters |
| Top decile (the biggest ones) | over 11,000 characters |
| Assistants with explicit Q&A entries | only ~45% |
Those last two rows are the ones worth staring at. More than half of all assistants have no prepared answers at all and lean entirely on a block of background text. Meanwhile a small top slice has grown its knowledge base past the point of being useful.
Two ways to break a perfectly good assistant
1) The knowledge base is too thin. The questions customers actually ask (price, shipping, warranty, how the thing works) are never written down anywhere, so the assistant has to infer them. Sometimes it infers wrong. One wrong answer is usually enough for a seller to decide the whole idea does not work. This is the common failure, because more than half of assistants have no explicit Q&A at all. We found the same gap from the customer's side when we looked at the three questions that go unanswered in DMs.
2) The knowledge base is too crowded. The opposite problem, and a sneakier one. The seller keeps adding and never removes. What builds up is several thousand characters that quietly contradict each other, and the assistant has no way of knowing which version is still true. Contradictions make answers wobble, and nothing kills trust faster than an assistant that says one thing on Monday and something else on Tuesday.
The sweet spot sits in between: a focused knowledge base, roughly the median ~4,500 characters, plus explicit Q&A for the questions you get every single day. An assistant is exactly as good as what you feed it. No better.
What actually works
1) Write down your 10 to 15 most common questions. It is the simplest move on this list and the one that pays the most, and more than half of sellers still have not done it. If you want a starting structure, here is a practical guide to automating your Instagram FAQs.
2) Keep it focused. Replace old information instead of stacking new information on top of it. If your knowledge base has grown large and repetitive, it is probably due for a cleanup rather than another paragraph.
3) Test before you trust. Ask your own assistant the hardest questions your customers ask. If it answers well, hand it the conversation. If it does not, you have just found the exact gap to fill.
4) Fix it instead of turning it off. A wrong answer is a hole in the knowledge base, not a verdict on the assistant. Fill the hole and it gets that question right forever. The speed is worth protecting: our own numbers put an AI reply at nine seconds against roughly an hour for a human, and speed only counts when the answer is correct.
The takeaway
A new assistant is a new hire. Skip the training, or hand it a folder of notes that argue with each other, and it will get things wrong and you will lose patience with it. The fix is not firing the new hire. A focused knowledge base plus a handful of written answers is the entire difference between an assistant that gets switched off and one that sells.
Vardast builds that step into setup. It walks you through a clean, focused knowledge base for your business and helps you add the answers that matter, so it stops guessing from day one. See how Vardast works.
Based on our own analysis of 11,878 assistants: the current version of each knowledge base and its question-and-answer entries. Size is measured in characters of knowledge base text.