Every store has a report of what sold. Almost none have a report of what someone wanted and could not find. That second list is worth more, and until recently it was genuinely difficult to collect.
A search that returns nothing is a sentence, not a string
Traditional site search logs a query and a result count. If the count is zero you get a row in a table: wide fit hiking boot - 0 results. Useful, but it flattens the interesting part. You cannot tell whether the shopper wanted a boot you do not sell, a width you do not list, or a boot you do sell under a name they did not guess.
A conversational agent captures the sentence around the query. It knows the shopper had already looked at two boots, that they asked about width specifically, and that they left after the answer. That context is what turns a dead end into an instruction.
Roughly one in five zero-result questions is a naming problem, not an inventory problem. Those are free revenue - you already own the stock.
The first week: three passes
Do not try to fix everything. Sort the board by volume and work top-down in three separate passes, because each one has a different owner.
Shoppers say “wide fit”; your variants say “EE”. Map the term and the demand converts immediately. This is a merchandiser's job and takes minutes.
Repeated questions about waterproof ratings mean the spec is missing, not that shoppers are careless. Add it once and the question stops for everyone.
The slowest to act on and the most valuable. Ranked, recurring demand for something absent is the cleanest buying signal you will get.
Measure the close, not the deflection
The temptation is to report how many questions the agent handled. That number always looks good and tells you very little. Tie the conversation to the cart instead: a resolved question that ends in an order is the only evidence that the fix worked.
Attribution also protects you from the opposite error - a high resolution rate on questions that were never going to convert. Volume is not intent.
We build the agent described above. The fastest way to see your own zero-result board is to install it and look at what week one turns up.