How TrustScope analyzes reviews
Our trust score is based entirely on a venue’s public reviews. On this page, we explain how the score is calculated, what the numbers mean, and what they don’t. Nothing other than the reviews themselves is used to determine the score.
What is collected
For each review we read its star rating, its timestamp, its text, how many photos are attached, how many reviews and photos the author has posted publicly in total, and whether the author uses an uploaded profile picture rather than Google's default avatar.
The sample is the venue's 500 most recent reviews, and a venue is analysed only once it has at least 50. Below that threshold the percentages are too noisy to be worth showing, so no analysis is produced at all.
Review text and author details are deleted as soon as the analysis has been computed. What we keep is Google's own review identifiers plus aggregate numbers, no text, no names, no profile pictures. When you open a metric to see the reviews behind it, they are fetched from Google live rather than from anything we stored.
Pleasant room, excellent espresso. The staff were not especially friendly, which is the only thing keeping this from perfect.
- Google’s own review identifiers
- The aggregate numbers on this page
- The review text
- The author’s name
- The profile picture
Trust Score
The Trust Score asks one question: do these reviews carry the signals of real, engaged reviewers? Each review scores between 10 and 100 points, built from a base of 10 plus four terms:
- Text depth - up to 55 points, on a logarithmic scale that reaches full credit at about 10 words. Most of the credit comes from a review having substantive text at all.
- Reviewer history - up to 15 points, also logarithmic, reaching full credit at around 50 reviews written by that author.
- Attached photo - 10 points if the review includes at least one image.
- Real profile picture - 10 points if the author uses an uploaded photo rather than the default avatar.
The venue's score is the average of its reviews' scores, minus a penalty of up to 10 points for the share of reviews showing an unusual pattern . That penalty is logarithmic too: it reaches its maximum at 10% and is capped there, so a single flagged review out of 500 costs about one point, not ten.
The trust score measures credibility signals, not truthfulness. A low score means the reviews are thin, anonymous or photoless - it is not proof that any individual review is fake.
- Base10 points
- Text depthup to 55
- Reviewer historyup to 15
- Attached photo10 points
- Real profile picture10 points
Unusual-pattern penalty. Per review: 10 – 100. The venue’s score is the average of its reviews, minus up to 10 points for the share showing an unusual pattern - capped at a 10% share.
Unusual patterns
The reasoning: a genuinely busy day produces a mix of ratings, lengths and photos. A wall of identical, empty five-star reviews from new accounts on a single day does not.
of which at least 4 are “Thin 5★” - all four at once:
- A five-star rating
- No more than 2 words of text
- No attached photo
- An author with 20 or fewer reviews
Every thin 5★ review on such a day is counted. The venue's figure is the total across all days, also shown as a share of the sample.
This is a statistical pattern only - not a judgment about individual reviews or their authors. Reviews can match the pattern for entirely innocent reasons, and we make no claim that any particular one is fake or paid for.
(no text)
😊
Sehr lecker
(no text)
Review-quality signals
Four percentages of the sample, shown as bars. They are the raw ingredients behind the trust score, exposed so you can see why it landed where it did. Higher is better for all four.
- Detailed - the review has at least 10 words.
- With photos - at least one image is attached.
- Real profile picture - the author uses an uploaded photo rather than the default avatar.
- Established reviewers - the author has written at least 5 reviews in total.
Aspect Radar
Five topics, chosen by the kind of place it is:
- Restaurants, cafes, bakeries & bars: Food, Service, Atmosphere, Speed, Cleanliness
- Hotels: Rooms, Service, Cleanliness, Location, Value
- Clinics, doctors & dentists: Care quality, Staff, Waiting time, Cleanliness, Communication
- Pharmacies: Advice, Staff, Waiting time, Availability, Prices
- Gyms: Equipment, Staff, Cleanliness, Crowding, Value
- Hairdressers & spas: Skill, Friendliness, Cleanliness, Waiting time, Value
An unrecognised category falls back to a generic set.
A review counts towards a topic if its text mentions one of that topic's keywords. Within the sentences that mention it, positive and negative sentiment words are counted, with a negation in the preceding few words flipping the sense. The axis is the positive share of those mentions, smoothed so that sparse evidence sits near the middle rather than swinging to the extremes. Google's own guided ratings, where reviewers tapped them, are blended in. A topic needs at least 3 contributing reviews to show a value at all.
Each axis is a sentiment ratio, not an average star rating: 50 is balanced, above 50 means more positive than negative mentions. This is deliberate - many popular venues are almost entirely five-star, which would flatten a star-based radar into a useless shape, while sentiment still separates the topics.
- Atmosphere88
- Cleanliness79
- Food84
- Service58
- Speed61
Rating Distribution
The rating distribution counts how many of the collected reviews gave each star rating. It reflects the most recent 500 reviews, so it can differ from the lifetime average Google displays.
The review-volume chart shows reviews per day over the last 30 days, or per month across the collected history. Empty periods are drawn as zero rather than interpolated over. The chart is descriptive only: nothing on it is flagged or coloured as a warning.
Limitations
Stated plainly, because a score whose weaknesses are hidden is worth less than one whose weaknesses are known:
Recency bias - the sample is the newest 500 reviews, not a random sample of all of them.
Sentiment is read in English, using Google's own translation of each review. A review whose wording survives translation is scored as well as a native English one; an idiom may not be. The trust score, the quality signals, the star distribution and the unusual-pattern rule are language-independent.
Mention is not attribution - sentiment is scoped to the sentence mentioning a topic, but within one sentence (“great service, awful food”) both words still count for both topics. It is a fast approximation, not language understanding.
Below 50 reviews no analysis is produced at all.
Signals, not verdicts - everything here describes patterns. It surfaces reviews worth a closer look; it does not determine that any specific review is fake, and says nothing about the person who wrote it.
What this is not
A trust score is information, not a decision. Nothing is granted, refused or decided on the basis of it.
- nothing is granted
- nothing is refused
- no venue is ranked down or removed
- A description of patterns across hundreds of reviews
- A reason to look closer, with the reviews attached to every number
- Reproducible - the same reviews give the same score
- Published - every threshold it uses is on this page
- A verdict on a business
- A claim that any single review is fake or paid for
- A judgment about the person who wrote it
- An automated decision within the meaning of Art. 22 GDPR
If you own a venue and object to its analysis, you can have it removed - see . The legal detail is in .