What customers actually ask a bot (and what you think they ask)
Two findings from production that change how you read your own bot's data: why your list of "top questions" is probably measuring your own interface, and why customers ask for your company number before they ask your price.
Every site owner knows which questions are the common ones. Ask them and they will recite the list. Then you put a bot live, open the log a month later, and find something else entirely.
Two findings from our production conversations change how those logs get read. The first is a trap almost everyone falls into the first time they look at “top questions.” The second changed our own website.
1. Your “top questions” list is probably measuring you, not your customers
The widget offers suggested questions — chips a visitor clicks instead of typing. Standard stuff, everyone has them, they lower the barrier to the first click.
The first time we listed the most common questions on one site, the top was emphatic: “How do I list my company?” and “What does registration cost?” It reads like a clean window into the customer’s head. Decision made about what goes on the homepage.
Then we opened the widget configuration. Those were literally the strings we had put there ourselves, as suggestions.
Nearly half of all conversations open with a click on a suggested question rather than typed text. If you do not separate those clicks out, “top questions” is not data about customers — it is your own interface handed back to you, with the authority of a table behind it.
That is not a flaw in the widget. Suggestions do their job — they lower the barrier to the first click, which is why we keep them. The flaw is in the reading.
Practical rule: when you look at what people ask, look only at what a human typed. Measure the suggested ones separately, and measure them as UI — which chip works, which does not. Those are two different reports answering two different questions. Anyone who mixes them is selling you fiction with percentages on it.
2. Customers ask who you are before they ask your price
Once the suggestions are stripped out and only typed questions remain, the top themes are the expected ones: price, how do I sign up, how long is the trial.
And then, right behind them, a theme nobody planned for: roughly one in six conversations is interrogating whether you are a real operation at all. Who is behind the site. Is there a registered company. How long have you been going. Do you have a tax ID.
Nobody prompted them. There is no chip for it. People type it themselves — and they do it before they ask anything about the product.
The same signal arrived independently from the other direction. In the log of questions the bot could not answer — the most honest report that exists, because nobody writes it, it simply accumulates — one of the most frequent entries was the company’s tax ID and registration details.
The bot was asked. It had no answer. It was recorded.
What we did about it
We dropped the street address from this site’s footer and put the full company name, tax ID and registration number there instead. Not because it looks better, but because the data says people are asking for it and could not find it.
The point is not three numbers in a footer. The point is the order: trust is demanded before the purchase, not after. While you are polishing your product description, the person on the other side is trying to establish whether you exist at all. If the answer is not within reach, they will not ask. They will leave.
If you sell anything online and your company details are not somewhere visible, that is probably the cheapest change you can make this week.
Three smaller things of the same kind
People ask in the evening and at weekends. It is densest between 19:00 and 23:00, and weekends are not quiet. Not because “24/7” is a nice line, but because someone who needs a slot or an answer does it once they are home — not when you are at work. The same system on an internal, employee-facing portal has zero conversations outside business hours. Timing tracks the audience.
Messages are short. The average message is around 40 characters. Nobody writes a paragraph to a bot. They write “how much” and wait. A knowledge base that answers neatly-phrased questions will not meet the way people actually type.
Not one message was in Cyrillic. Not one. One hundred percent Latin script, across every Serbian-language conversation we looked at. If you are configuring a bot for the Serbian market, that is a detail that saves you time.
What to do on Monday
- Open the log and separate typed from suggested. If you cannot separate them, your question statistics are worthless — and that is the first thing to demand from any vendor.
- Read the list of questions the bot could not answer. It is the only report nobody wrote in order to please you.
- Answer the trust question before it is asked — on the site, not in the bot.
- Compare when people write against your opening hours. The gap is your missed revenue, expressed in hours.
None of this came from assuming what customers want. All of it came from the log. Which is the whole point: a bot does not only answer. It is the first honest report on what your customers actually ask — and it usually does not match what you put on your homepage.