One market
Reviews come from Tucson, AZ only. Patterns may not transfer to other markets.
Responsible AI
Signals for people, not verdicts.
HomeLens AI surfaces patterns in customer language so managers know where to look. Every alert is a starting point for human judgment.
Language guardrails
How we describe outputs.
We say
- Review-based risk signal
- Needs management attention
- Operational concern and complaint signal
- Relative concern tier within this corpus
- Prompt for human investigation
We never claim
- Bankruptcy or financial-failure prediction
- Legal liability
- Verified misconduct or fraud detection
- Verified safety violation
- A verdict on any individual business
01 — Known limitations
What the data can and cannot tell you.
Historical data
Data runs from Nov 2006 to Jan 2022 and may not reflect how providers operate today.
Self-selection bias
Yelp reviewers are not a random sample of customers. Very satisfied and very unhappy customers are over-represented.
Star ratings as labels
Models learn from star ratings, an imperfect proxy for operational risk. Three-star reviews were excluded from training.
Unverified allegations
A review is one customer’s account. Detected safety or conduct language is a claim to investigate, not a confirmed event.
Relative tiers
Provider tiers rank providers against each other within this corpus. They are not absolute probabilities of any outcome.
Human oversight
Every serious signal goes through a person.
Signal raisedModel score exceeds the operating threshold
Analyst reviewRead the full review and its context
Provider dialogueHear the business’s side before acting
Proportionate actionCoaching, process fix, or no action
Monitor and retrainTrack drift and refresh the model
Privacy
Data privacy and anonymization.
Providers are masked
Businesses appear only as stable codes such as Provider_0235. Names and Yelp business IDs are not shipped with the app.
Text is masked before modeling
URLs, emails, phone numbers, dollar amounts and explicit star phrases are replaced with tokens before any model sees the text.
Volume threshold
A provider needs at least 20 reviews before it is monitored, which avoids noisy, unfair conclusions from a handful of reviews.
No raw data deployed
The website ships aggregates and a small model only. The raw Yelp dataset is never published.
Pasted text is not stored
Review text you type into the analyzer is processed for the request and not saved. The site sets no cookies and loads no analytics or ad trackers.
Dataset terms
The Yelp Open Dataset is used for academic purposes under its terms and is not redistributed.
Questions about how a signal is made?
Read the methodology or inspect the analysis endpoint.