All articles Fraud & ATO

How Airbnb Uses Targeted Friction Instead of Blocking — and How to Do It in Your Business

When a booking on Airbnb is paid with a stolen card, the real cardholder eventually disputes the charge. Airbnb’s engineering blog explained in 2018 that the company bears the cost of those chargebacks itself rather than passing them to hosts. So stopping stolen-card bookings matters directly to its own bottom line.

The obvious answer is to block anything that looks risky. The problem is that “looks risky” is a score, not a fact. Some of the bookings above any threshold belong to real guests: a traveler paying from a new country, a first booking made late at night, a card used for the first time. Block them and you lose a customer who did nothing wrong.

Airbnb’s answer was a middle option. In the post “Fighting Financial Fraud with Targeted Friction”, Airbnb trust data scientist David Press described how the company applies extra steps, called frictions, only to the bookings its models flag, and how it decides where to draw the line.

What Airbnb built

The method, as described in the post and in InfoQ’s write-up of it, has four parts.

A model trained on past outcomes. Airbnb scores transactions with machine-learning models trained on past examples of confirmed good and confirmed fraudulent behavior. The model doesn’t decide anything by itself; it produces a score.

Friction instead of a wall. Above a threshold, some transactions are blocked outright, but in most cases the user gets a chance to complete an extra verification. The examples the post gives are all things a real cardholder can do and a person holding a stolen card number usually can’t:

  • Micro-authorization. Two small charges are placed on the card, and the user confirms the amounts, which only the person with access to the card account can see.
  • 3-D Secure. The card issuer authenticates its own cardholder.
  • Billing statement. The user uploads a statement for the card.

An explicit cost for every mistake. The post sets out what each outcome costs. A false positive is a good booking that got friction: some good users won’t complete the step and won’t use Airbnb. A false negative is fraud that scored below the threshold: it costs the full payment plus overhead such as processor fees and the risk of higher card decline rates. A true positive is fraud the friction stopped. The threshold is chosen to minimize the total of those costs, rather than to catch as much fraud as possible.

A measured price for each friction. This is the part most teams skip. To learn how many good users each friction loses, Airbnb ran A/B tests through its experiment framework. It assigned users with low model scores, people very unlikely to be fraudsters, to see the friction at the same stage of the funnel where suspicious users would see it. The drop-off in that group is the cost of the friction to good customers, measured instead of guessed.

The numbers

Airbnb’s post walks through a numerical example that compares blocking with friction. We’d rather send you to the source than paraphrase its figures, so read it in Airbnb’s post. What matters here is the shape of the argument: the comparison is made in money, with the cost to good customers counted on the same scale as the fraud stopped.

The idea doesn’t depend on a number anyway. A friction is worth using when the fraud it stops is worth more than the good customers it loses, and both sides can be measured.

What transfers, and what doesn’t

What doesn’t transfer: Airbnb has years of labeled bookings, a data science team and its own experimentation platform. A business with a few thousand signups a month can’t train a model like that, and its A/B tests will be small. Some of Airbnb’s frictions, such as micro-authorizations, also depend on its payment setup.

What does transfer:

  • A middle answer. Between allow and block there should be a step, and usually more than one.
  • The step should match the risk. A light check for slightly odd traffic, a stronger one where the money is real.
  • Mistakes have prices. Write down what a missed fraud costs you and what a lost good customer costs you. The right threshold comes from those two numbers, not from a feeling.
  • Test the friction on good users first. A small sample of low-risk users tells you how many people a step loses before you show it to everyone you’re unsure about.

How to do it in your business

Prynt Control is built around this pattern. Prynt never blocks on its own: it returns a decision and a suggested step, and your code decides what the user sees.

  1. Start in mirror mode. For the first 14 days Prynt only observes. On day 14 it shows, case by case, what it would have stopped. That is your first look at how many of your flagged users were real customers, before anyone saw a step. Rules can also run in evaluate mode before you switch them on.
  2. Pick the moment with a use-case decision. Signup, free trial, bonus or promo, login, checkout, extra seat: each has a ready-made decision with three outcomes, allow, step-up or block. A risky signup and a risky checkout deserve different answers.
  3. Use the friction ladder for the step-up. Instead of a single yes or no, the answer says which step to ask for: none, then captcha, then verify email, then card or passkey, then block. Like Airbnb’s frictions, each step is easy for a real person and harder for automation or someone running many accounts. Most risky traffic should land on a low step; block should be rare.
  4. Price your mistakes. For each use case, write down what a miss costs (a chargeback, an abused trial, a bonus paid out) and what a lost signup or sale is worth. If a step loses more good customers than the fraud it stops is worth, move it down the ladder: in the console you can cap each use case at a lower step.
  5. Measure each step on good users. Show a step to a small random slice of low-risk users for a week and compare completion with users who saw none. That is Airbnb’s experiment at your scale: a rough number, but a real one.
  6. Send outcomes back. When a case turns out to be fraud or legit, report it with POST /v1/outcomes. Those labels are what let decisions improve over time, the same way Airbnb’s model learns from confirmed cases. Report the result of each step too, with POST /v1/control/steps: a device that passed is not asked again for that use case for 30 days unless something stronger shows up, and one that failed gets one rung more.

Friction only works if the signals behind it rarely flag real people. We publish ours on the accuracy page: in our lab, Prynt flagged 0 of 240 real humans as bots, and detected 89% of stealth automation tools (238 of 268). Lab results are not field results, which is one more reason to start in mirror mode and measure.

The honest part

Targeted friction isn’t free. Every step loses some good users, and Airbnb’s whole point was to measure that loss instead of pretending it was zero. You won’t have Airbnb’s data or its experimentation platform, and your measurements will be noisier. But the habit is the same at any size: ask for a step instead of slamming the door, choose the lightest step that works, and check what it costs the people you want to keep.

Prynt is not affiliated with Airbnb; this post describes Airbnb’s published work.

Try it free

Prynt is device intelligence with a free plan: 5,000 identifications a month, visitor IDs with confidence and bot detection — no credit card. Upgrade to Pro for VPN, proxy, tampering and behavioral signals.

Keep reading