How it works
From thousands of reviews to five tasks
The analysis runs in five steps. Each one leaves a trace you can see in the panel — if something goes wrong, you know exactly where.
-
1
Fetch the whole history
Not the latest ten reviews — every one of them, from the first ever posted. Later analyses fetch only what is new, so the second report costs a fraction of the first.
With 4,000 reviews this step takes minutes, not days.
-
2
Normalise the data
Ratings from different platforms are mapped onto one scale, reviews without text are separated from those with text, and duplicates by the same author within one source are merged.
Statistics come from every review; conclusions only from reviews with text.
-
3
Classify the signals
The language model reads a review and marks the fragments that name a specific problem — as character positions, not retyped text. One review can concern several topics at once.
This step uses the cheapest model: it is about volume, not style.
-
4
Aggregate and apply thresholds
Signals group into problem cards. A card is created only once it has at least three independent source reviews and a verified quote. A one-off incident never reaches the list.
Trend is measured on the share of complaints per period, not their count.
-
5
Verify and build the report
A separate layer cuts every quote out of the original review and compares it character by character. A fragment missing from the original drops out, together with the card that rested on it.
Only after this check do the PDF and the share link get created.
Why you can trust it
The model writes sentences, not numbers
A report from a language model is worth exactly as much as your ability to check it. The whole architecture of Mutterbox follows one rule: anything countable is counted by the database, and the model only gets the work it is genuinely good at — naming and organising.
Numbers come from the database
The model never produces a single number that reaches the report. Complaint counts, dates and percentages come from SQL.
Quotes verified literally
A quote is a character range in a specific review, not text from the model. A fragment absent from the original does not pass.
Trend measured as a share
A "rising" problem means a rising share of complaints, not a rising count — otherwise every problem would look rising in a growing venue.
„Felt like a water park from another era, everything old and the outdoor pool was barely heated."
Principles
Four rules that make the report checkable
A report from a language model is worth exactly as much as your ability to verify it. These four rules live in the architecture, not in the terms of service — and they are what separates this from a tool that summarises reviews nicely.
-
Every number comes from the database
The model names and organises; it does not count. "94 complaints" is the result of an SQL query over classified signals — you can reproduce it from the CSV export.
-
Every quote comes from the original
Quotes are cut from the review by character position and compared character by character. A fragment missing from the original drops out along with the card that rested on it.
-
A threshold instead of guessing
A card needs at least three independent reviews. With a thin corpus we say so plainly and show where to find more reviews, instead of forcing out a report.
-
Trend measured as share, not count
In a growing venue the raw complaint count rises for every topic. So the chart shows the share of reviews raising an issue — otherwise growth would look like decline.
Common questions
What people ask before their first report
Where does Mutterbox get reviews from?
From Google Maps, Facebook, TripAdvisor and Booking.com — you point at a profile and we pull the reviews with their history. Collected reviews stay in the database, so a second analysis of the same venue does not pay twice for the same data, and a refresh fetches only what is new.
How long does the first analysis take?
From a few to about fifteen minutes, depending on review volume. Fetching several thousand reviews is the slow step; the analysis and report assembly take seconds. You get a notification when the report is ready — no need to wait at the screen.
How do I know the report is not made up?
Every number comes from a database query and every quote is cut from the original review and compared character by character. You can download a CSV export with source quotes and check any sentence in the report against the original.
I only have 40 reviews. Is it worth it?
Yes, but the report will be marked as indicative. Below twenty reviews with text a single voice weighs too much to call anything recurring — we say so plainly and show you where reviews about your brand live that you are not collecting yet.
Does Mutterbox reply to reviews for me?
No — deliberately. Replying requires access to your business profile, which we do not want to hold, and a reply written by a tool reads exactly like one. The report tells you what to answer and what to fix; you send it.
Can negative reviews be removed?
No, and nobody can do that legally. What can be removed is their cause — that is what the fix list is for, where every task carries the number of complaints it takes away. And do not buy reviews: it shows in the data, because bought reviews cluster in time and never name specifics.
Be first in when we launch
Join the waitlist. We'll write once, on launch day, with a link to create your account.