A behind-the-scenes look at how we keep our review dataset clean — and why most platforms cannot.

The biggest, most persistent critique of online healthcare reviews is the same critique that has haunted every review platform since the invention of TripAdvisor: fakes, planted reviews, and competitor sabotage. When we started building the PublicListen review dataset three years ago, we assumed we would encounter the problem. We did not appreciate the scale.

Approximately 14% of every batch of submitted reviews fails one of our four verification layers. We publicly disclose that number because we believe transparency about our rejection rate is itself the strongest signal of trust. This is a brief behind-the-scenes look at how the pipeline actually works.

Layer 1: Identity verification

Every reviewer must verify a phone number, an email address, and — for reviews above a certain sensitivity threshold — a government-issued ID. The ID is not stored; it is verified against the name provided and then discarded. The purpose is not to know who the reviewer is publicly (all reviews remain pseudonymous) but to confirm that one real person is behind each account. This step alone eliminates roughly 4% of submissions.

Layer 2: Proof of visit

The reviewer must upload some evidence of the underlying visit: a bill, a prescription, an appointment SMS, or a discharge summary. Personal information is redacted before storage. This is our strongest single defence against fake reviews. It also gives us the ability to link a review to a specific date and procedure — which lets us match reviews against outcomes over time. About 6% of submissions fail at this layer.

Layer 3: Pattern detection

Automated systems check writing style consistency, submission timing patterns, IP address clusters, and behavioural signals. Reviews written in bursts, from clustered IP addresses, using similar phrasing, or arriving suspiciously close to a positive counter-review, all get flagged for manual review. About 2% of submissions get flagged and rejected here — but the more important effect is that pattern detection catches the coordinated attacks that would otherwise dominate the dataset.

Layer 4: Editorial review

Anything flagged by the automated systems, and a random sample of everything else, goes through our editorial team. Humans catch what algorithms miss: subtle sarcasm, coded criticism, personal disputes disguised as clinical complaints. About 2% get filtered here.

About 14% of submitted reviews fail one of the four layers. We disclose that number because transparency about rejection rate is itself a trust signal.

What we do not do

Three things we explicitly do not do, and we believe most other platforms should also stop doing:


Why other platforms cannot do this

The verification pipeline is expensive. It slows down the volume of new reviews. It creates friction for the reviewer. Most platforms optimise for volume and speed instead — because more reviews means more traffic means more advertising revenue. We optimise for signal, because we believe signal is what patients actually need.

The final principle

Reviews are useful only if they are true. Truth in reviews requires infrastructure. Infrastructure requires investment. If you see a review platform that is free, frictionless, and infinite in volume — assume that any single review on it is probably fake until proven otherwise. That is the honest starting point.

The most common attack patterns we defend against

Three patterns account for the majority of fake-review attempts we detect:


How our confidence score works

Every published review carries a confidence score, from 1 to 100, derived from the four verification layers. Reviews above 85 are considered fully verified. Reviews between 60 and 85 are shown with a "partially verified" label. Anything below 60 is not published. About 8% of our published reviews carry the "partially verified" label — usually because the proof-of-visit was less definitive.

What patients can do

Two habits that improve the ecosystem: submit reviews for both good and disappointing experiences, and use the "was this review helpful?" signal generously. Both actions strengthen the underlying dataset for the next patient.

The final principle

Reviews are, ultimately, an act of civic trust. Every patient who takes the time to submit an honest, verified account of a visit is investing in the next patient's ability to make a better decision. We do not underestimate what we are asking of reviewers — and we do not underestimate what verified reviews are worth to the readers who rely on them. Trust the pipeline. Trust the verification. And when you have a healthcare experience worth documenting, take four minutes to add it to the collective record.