Prynt Control

The control system the big platforms built in-house, ready for your business

It starts in mirror mode: for 14 days you see what it would have stopped, without touching a real user. Prynt Control applies the methods that Meta, Google, LinkedIn, Uber and others published about fake accounts and fraud, on your traffic, at a price a small team can pay.

$199 a month · 500k identifications · 1-year retention · priority support · everything in Scale

We copy their published methods, not their data or their scale. Prynt never blocks anyone on its own: it returns a decision and you decide what happens.

What Control adds

Nine pieces of a fraud control system

Each one answers a question a small team asks every week: what would we have stopped, what should we ask this person, who is behind these accounts, and how much money is this.

Mirror mode

For the first 14 days Prynt only watches: every request is answered allow. Then the report shows, case by case, what it would have stopped. An admin can stop the mirror or restart it, and new rules can run in evaluate mode before you switch them on.

Use-case decisions

Ready-made decisions for signup, free trial, bonus or promo, login, checkout and extra seat. Each one answers allow, step-up or block, with the reasons.

Friction ladder

Instead of only blocking, the answer says which step to ask for: none, a captcha, verify the email, a card or a passkey, or block. Good users rarely see a step.

Account clusters at signupRolling out

New accounts that share a device, a network and a time window are grouped, so you see "9 accounts from the same place in 20 minutes" as one case, not nine.

A case queue that learnsRolling out

Cases ranked by suspicion, the worst first. A fraud or legit button on each one, and every answer teaches the system about your business.

Account sharing reportRolling out

Which accounts are used by more than one person, and a ready-made "add a user" notice you can show them. An invitation to pay for a seat, not a lockout.

Stripe connectionRolling out

Connect Stripe and your disputes and fraud refunds become fraud labels automatically, so the system learns from the outcomes you already have.

Monthly money statementRolling out

Abused trials and bonuses, shared seats and ad clicks from bots, in money, with conservative estimates and the method shown. Plus prevalence and proactive rate.

Bot share per ad campaignRolling out

The share of visits from automation for each campaign, so you know which ads are paying for bots before you renew them.

Start with zero risk

Fourteen days of watching before anything changes

Control starts in mirror mode on its own. For the first 14 days your users see nothing different and nothing is blocked: every decision is answered allow, and Prynt records the one it would have made on every signup, trial, login and checkout.

  1. 01

    Day 0 — install

    Add the agent and send your account id with each identification. Mirror mode starts by itself on the first request; nothing else changes.

  2. 02

    Days 1–13 — observe

    Every decision is recorded as "would have": would have allowed, would have asked for a step, would have blocked. No user is affected.

  3. 03

    Day 14 — the report

    The mirror report lists what Prynt would have blocked or challenged, with counts, the top reasons and recent examples. Mark the ones it got wrong; they are the first labels it learns from.

  4. 04

    After — decisions apply

    When the 14 days end, decisions apply. An admin can stop the mirror sooner or restart it for 1 to 30 days from the console, and rules you write later can run in evaluate mode first.

Uber documented the same practice for its fraud rules engine: a new rule with the Evaluate status "is being run but won't actually affect the user". Source

Friction ladder

Ask for the smallest step that settles the doubt

Most suspicious sessions are not worth a block. The decision tells you which step fits the risk, and you choose whether to apply it.

  1. 1

    None

    Most of your users. No step, no delay.

  2. 2

    Captcha

    For automation. Cloudflare Turnstile or the one you use.

  3. 3

    Verify email

    A code or link to the address they gave.

  4. 4

    Card or passkey

    Something that costs a fraudster more than a new email.

  5. 5

    Block

    Only with direct evidence, such as a device you labeled fraud. Your code still decides.

No SMS step: phone numbers are cheap to rent and costly for good users, so Control does not use them.

How big platforms do it

Published methods, applied to your business

Large platforms have written down how they control fake accounts and fraud. These are the methods Control applies, with the public source for each one.

Uber

Runs new fraud rules in an Evaluate status first: the rule runs but does not affect the user, so analysts see what it would do before turning it on.

Mirror mode and evaluate mode · Uber Engineering, “Mastermind: Using Uber Engineering to Combat Fraud in Real Time” (2017)

Google

Across more than 350,000 real hijacking attempts, device-based login challenges stopped 100% of automated attempts and over 94% of the attempts rooted in phishing.

Friction ladder · Doerfler et al., “Evaluating Login Challenges as a Defense Against Account Takeover”, WWW 2019

Airbnb

Applies "targeted friction": risky cases get a verification step the real user can pass, instead of an outright block, and the cost to good users is measured.

Friction ladder and use-case decisions · Airbnb Engineering, “Fighting Financial Fraud with Targeted Friction” (2018)

LinkedIn

Groups the accounts registered from the same IP on the same day and scores each cluster as a whole, using only what is known at registration; every account inherits its cluster's score.

Account clusters at signup · Xiao, Freeman & Hwa, “Detecting Clusters of Fake Accounts in Online Social Networks”, AISec 2015

Meta

Classifies accounts by the features of the entities around them (Deep Entity Classification); the paper reports a 27% relative reduction in the estimated share of abusive accounts (from 5.2% to 3.8%) beyond what other methods already caught. Meta also reports prevalence and proactive rate publicly.

Account clusters and the money statement metrics · Xu et al., “Deep Entity Classification: Abusive Account Detection for Online Social Networks”, USENIX Security 2021

Tuenti

Ranked accounts by how likely they were fake (SybilRank): about 90% of the top 200,000 warranted suspension, against about 5% of the accounts inspected from user reports.

Case queue ranked by suspicion · Cao, Sirivianos, Yang & Pregueiro, “Aiding the Detection of Fake Accounts in Large Scale Social Online Services” (SybilRank), NSDI 2012

Netflix

Rolled out paid sharing in 2023, with an "extra member" option for people outside the household; it reported 5.9M, 8.8M and 13.1M paid net additions in the second, third and fourth quarters, crediting paid sharing together with its programming.

Account sharing report and "add a user" notice · Netflix quarterly earnings and shareholder letters, Q2–Q4 2023

Stripe

Trains Radar's fraud models on payment outcomes such as disputes, so the labels come from what actually happened.

Stripe connection · Stripe, “A primer on machine learning for fraud detection”

None of these companies uses, endorses or partners with Prynt. Prynt applies methods they published; it does not have their data or their scale, and it says so.

Results in money Rolling out

A monthly statement, not a dashboard of alerts

Every month Control adds up the money it found, with conservative estimates and the method next to each line, so you can check the arithmetic yourself.

Example SaaS · October Example — illustrative numbers for a fictional business, not customer data
What Prynt foundCountEstimateHow it is counted
Repeat free trials41$1,189Trials on a device that already had a trial in the last 90 days × one month of your cheapest plan ($29)
Seats used by more than one person18$270Accounts active on 2+ devices in different cities at the same time on 5+ days × one seat ($15)
Promo codes claimed by linked accounts23$230Second and later redemptions in the same account cluster × code value ($10)
Ad clicks from automation1,240$521Clicks with a bot verdict × your average cost per click ($0.42)
Total found this month $2,210Only counted where the evidence is direct; nothing is extrapolated
Prevalence2.1%Share of new accounts that look fake or abusive.
Proactive rate87%Share of those Prynt flagged before you noticed them yourself.

Your statement uses your prices and your traffic. Some months it will be smaller than the cost of the plan, and it will say so.

What Control does not do

Plain limits

  • It does not block anyone on its own. Every decision is a recommendation your code applies or ignores.
  • It does not send SMS or ask for phone numbers.
  • It does not promise to remove every fake account. It ranks, groups and explains, so your team decides faster.
  • The accuracy figures we publish come from our lab, not from the field, and they include what Prynt misses.

Accuracy, measured

Price

$199 a month

500k identifications a month, 1-year retention, priority support and everything in Scale, plus the control suite. Mirror mode is on from the first day.

Questions about Control

Is it risky to turn Control on?

No. Control starts in mirror mode for the first 14 days: every decision is answered allow, Prynt records what it would have done, and nothing changes for your users. An admin can stop the mirror or restart it (1 to 30 days) from the console.

Does Prynt block my users?

Never on its own. Each decision is allow, step-up or block, with the reasons and the suggested step. Your code decides what to do with it.

Which verification steps can it ask for?

None, a captcha (for example Cloudflare Turnstile), email verification, a card or a passkey, or block. There is no SMS step.

How is the money in the statement calculated?

Each line shows its method: the count of cases with direct evidence times a price you give us (plan price, seat price, promo value, cost per click). We do not extrapolate, so the estimate is on the low side.

Do I need Stripe?

No. The Stripe connection turns disputes and fraud refunds into labels automatically. Without it, you label cases from the queue or through the API.

Some features say «Rolling out». What does that mean?

Control is being released in stages over the coming weeks. A capability with that badge is not in your account yet; it is turned on for Control accounts as it ships.