Features
A problem card: the complaint with a number, a weight and a quote
"Worth looking after cleanliness" is technically correct and practically useless. A problem card says how many people it affects, how badly, since when, and what proves it.
What is on a card
A title written in the guests' words, not ours. A category and a confidence level. The number of reviews carrying the signal, together with its share of the whole corpus. Source quotes with a date and a platform. At the end, the estimated effect on the average rating if the problem went away.
Weight computed from data
Card order follows complaint counts, not what sounds most interesting. What hurts the largest number of people sits at the top — including when it is boring.
Confidence, not a claim
The confidence level comes from the number of independent sources and from how unambiguous the statements are. A card built on three reviews from one platform does not pretend to be certain.
Confirmation across platforms
When the same signal recurs on Google Maps and Booking.com, the card gets a confirmation marker. That is the difference between one angry person and a phenomenon.
Step by step
How a single problem card is built
-
1
Collecting the signal
The model assigns categories to every review with text. Statements landing in the same category with a negative sentiment form a candidate card.
-
2
Checking the threshold
A candidate needs at least three independent source reviews. A one-off incident, however emotionally described, stays in the reviews and never reaches the task list.
-
3
Computing weight and confidence
Weight is the number of reviews carrying the signal and its share of the corpus. Confidence comes from the number of independent sources and how unambiguous the statements are — a card resting on three reviews from one platform does not pretend to be certain.
-
4
Selecting quotes
The card takes the statements most representative of the signal, not the harshest. Each is verified against the original and carries a date, a platform and a review number.
In practice
A complaint that looks like an incident month by month
Forty-one complaints about waiting at reception across a year averages three and a half a month — a number any shift manager will call rare. Only the breakdown by day of week showed that twenty-nine of them fell on Fridays. It stopped being a question of service quality and became a question of staffing.
29 of 41
complaints about waiting fell on one day of the week. The problem card shows that distribution, because without it the task would read "improve service".
Why it works this way
Why a problem without a number is not a problem
Without a complaint count there is no ordering, and without ordering a task list is a wish list. A business owner has a finite budget and finite team time — the first thing they need is not "what is wrong" but "what to fix first".
Common questions
How do you know a problem repeats rather than being a one-off?
A card needs at least three independent source reviews. A one-off incident, however emotionally described, does not reach the task list — it stays in the reviews without pretending to be a phenomenon.
Can I change the categories problems are split into?
Yes. The built-in categories are a starting point, and you add up to ten of your own with a one-sentence description for the model. A furniture maker wants to see "installation" and "delivery dates"; a clinic wants "phone bookings".
What does the estimated effect on the average mean?
It approximates how much the average rating would rise if the reviews carrying this complaint did not carry it. It is clearly marked as an estimate and exists to compare cards against each other, not to promise an outcome.
See also
Be first in when we launch
Join the waitlist. We'll write once, on launch day, with a link to create your account.