Skip to content
All articles

Industry insights

The True Cost of a False Positive (Why FP Matters More Than FN as You Scale)

Once your false negative rate is tolerable, your false positive rate is where the margin actually hides. Here's the CAC math on blocking a legitimate customer, and why the fraud industry's obsession with catching more fraud is the wrong lens past a certain volume.

Every fraud vendor pitch you've sat through has the same opening move. A chart with fraud prevented. A line going up. A dollar figure attached. "We caught $X in fraud losses this quarter." The vendor is telling you about false negatives: fraud that would have happened, now didn't.

What the vendor is not telling you, and what nobody in the fraud industry likes to talk about, is the other axis of the confusion matrix. False positives. Legitimate customers flagged as fraud and either blocked, declined, or given enough friction that they didn't complete the transaction.

At low volume, FP is a rounding error. A store doing $500k a year, with 1,000 orders a quarter, blocks a handful of legitimate customers and the cost is invisible against the fraud it actually prevented. The math recommends aggressive thresholds.

At higher volume, the math inverts. A store doing $30M a year, with 200,000 orders a quarter, blocking 0.5% of legitimate customers as false positives is burning a six-figure annual number in acquired-but-rejected demand, and the fraud it's preventing is often a smaller number than the demand it's destroying.

This post is the argument for why, past a certain scale, false positive rate is the metric you should be optimizing, not false negative rate. And why most fraud tools are pointed the wrong way for a store past the $10-20M revenue mark.


The CAC math, starting from first principles

Start with the cost of a false positive. A legitimate customer tries to buy, gets blocked or friction-gated, and abandons.

What did you spend to acquire that customer? For most Shopify merchants in 2026, blended CAC (the cost of getting a new customer from any channel) is in the $40-$120 range, heavily dependent on vertical. Apparel and accessories sit at $50-$80. Electronics and home goods are $80-$150. Supplements and CPG are often lower, $30-$60, but with faster repeat cycles. Luxury is its own planet, $150-$400+.

A first-order customer blocked is that CAC, spent and then wasted. You paid Facebook, Google, TikTok, or your influencer agency to drive this person to your store. They converted through your funnel. They entered their card. Your fraud tool said no. The money you spent to bring them in is gone and they're not coming back. They went to your competitor, who didn't block them.

That's the floor. $40-$120 per blocked legitimate first-order customer.

Then add the lost lifetime value. A first-order customer who would have completed the transaction has some probability of becoming a repeat customer. LTV minus CAC is the standard measure of whether ecommerce unit economics work. For healthy Shopify stores it's typically 2.5-5x CAC. So a $60-CAC customer blocked isn't just losing you the $60. It's losing you the expected $90-$240 of lifetime contribution margin you would have gotten from that relationship.

Running rough numbers: a false positive on a first-time customer costs $100-$350 in CAC + lost LTV, depending on your unit economics.

Now do the same math for a false negative. A fraudster completes an order worth, say, $120. The cost is the disputed amount plus the processor fee plus the unrecovered merchandise. Call it $200-$250 at typical rates (see our real cost of a chargeback breakdown for the full decomposition).

Those numbers are close. A false positive on a good customer and a false negative on a bad customer cost roughly similar amounts, in the $100-$350 range. Which means the relative importance of FP vs. FN depends on the ratio of good customers to fraudsters in your order flow.

For most legitimate Shopify stores, that ratio is somewhere between 40:1 and 500:1. Which means, per order, an FP is 40-500x more likely than an FN to be the population you're affecting. And the cost structure is similar. So the optimization equation says: for every dollar of improvement you get in fraud caught, you need to make sure you're not spending more than a dollar in legitimate customers blocked.

And this is exactly the tradeoff most fraud tools do not measure or expose.


Why the industry is pointed the wrong way

The fraud industry grew up inside payment processing, where the customer is the bank's problem and the merchant's job is to stop fraud at the door. That's the wrong mindset for an ecommerce brand that just spent $80 acquiring a customer through paid media.

The incentive structure for a fraud tool vendor is also misaligned. "We prevented $50,000 in fraud this quarter" is a good story. "We avoided 200 false positives that would have cost you $40,000 in CAC and lost LTV" is a better story but a harder one, because it requires the vendor to know your CAC and your LTV, which they usually don't.

The result is that fraud tool dashboards are all built around fraud caught, not customers-correctly-cleared. Go look at any fraud tool's dashboard right now. It will show you flagged orders. It will show you rules triggered. It probably won't show you the cost of the orders it declined that would have been legitimate revenue.

That metric is hard to compute. It requires ground truth on declined orders, which you don't have because you declined them. But it can be estimated, and once a merchant sees the estimate the calibration conversation changes.


How to actually estimate your false positive cost

You need three things:

1. A sample of declined orders you can adjudicate. Pull 200-500 orders your fraud tool flagged or declined in the last 90 days. For each, ask the question: was this actually fraud, or a legitimate customer who happened to match some pattern?

You adjudicate this by looking for signals the fraud model didn't see: did the customer reach out to support after being declined? Did they successfully re-order (maybe from a different card, or on a different device, or after clearing cookies)? Did another account with the same email or phone complete a successful order later? Are any of the "matches" that flagged them actually coincidences (same first name, same city, same IP range because of CGNAT)?

After adjudicating 200-500 declines, you'll have a false positive rate for your current fraud tool configuration. It's almost always higher than the vendor's claimed number.

2. Your CAC by channel. Break it out if you can, because the FP cost is different for a $200-CAC Google Shopping customer than a $30-CAC email-list repeat buyer. Most stores can produce a blended number; a subset can break it out by channel.

3. Your LTV estimate. Net contribution margin per customer over 12-24 months. If you don't have this, approximate it as 2.5-4x CAC as a starting point.

Multiply: FP rate × orders per year × (CAC + LTV) = annualized FP cost.

A store with 100,000 orders a year, a 0.4% false positive rate, and a blended CAC+LTV of $200 is losing $80,000 a year to false positives. Add to that the downstream effects (the blocked customer writes a one-star review, the blocked customer tells their friends, the blocked customer's complaint becomes a ticket your support team handles) and the real number is 20-40% higher.

Compare that to your fraud-caught number. At most mid-market stores past $10M, the false positive cost is in the same order of magnitude as the fraud-caught number, and often larger.


When FP cost starts to exceed FN cost

The crossover isn't a universal threshold. It depends on your CAC, your fraud base rate, and your tool's precision. But the mechanics:

Low volume, high fraud rate: FN wins. A small store with 5% fraud rate should be aggressive. Every caught fraud event materially affects the P&L, and the population of legitimate customers being affected is small.

Medium volume, moderate fraud rate: parity. A store at $3-10M revenue with 0.5-1.5% fraud rate sits near the crossover. Small changes in tool precision change which side wins.

Large volume, low fraud rate: FP wins. A store past $10-20M with fraud rate under 0.5% is almost always overspending on false positives. Tightening the precision of the fraud model (accepting a slightly higher FN rate in exchange for a materially lower FP rate) is usually a 3-5x positive-ROI move.

This is the inverse of the conventional fraud wisdom, which says "as your volume grows, the absolute fraud losses grow and you need to be more aggressive." True in absolute terms, false in relative terms. The good-customer population grows faster than the fraud population at most scaling stores, so the economic weight of getting the good customers right grows faster than the weight of catching more fraud.


What to tune

If you've done the math and your FP cost is material, the lever you want to pull is precision: how confident the model is before it declines. In practice that means:

Higher triggered thresholds. If your tool has risk scores, move the "block" threshold up. Catch the unambiguous 99th-percentile cases. Let the 90th-95th percentile cases through to a softer response (manual review, step-up auth, smaller order limit) rather than a hard decline.

Multi-signal gating. A single signal (disposable email, new address, same device as a historical fraud case) should not trigger a block. Require 2-3 signals to co-trigger before a hard decline. Single-signal patterns are where the FP cost lives.

Whitelist repeat customers aggressively. A customer who has bought from you before, paid, and not disputed is a radically lower risk than the population base rate. Most fraud tools don't weight this enough. A repeat customer should need to hit substantially more signals to trigger a block than a first-time customer.

Friction over block, when in doubt. For the middle zone (risky but not obvious), a step-up friction (3DS, additional verification, delayed fulfillment with a hold email) preserves the option to serve the legitimate customer. A block removes it. When you're not sure, add friction instead of cutting off.


The takeaway

The fraud industry measures success by fraud caught. Past a certain volume, that's the wrong metric. What matters is the net economic outcome of the decisions your tool makes (fraud caught minus good customers lost) and at most mid-market and enterprise stores, the lost customers number is competitive with or larger than the fraud-caught number.

The practical test: pull a sample of declines, adjudicate them, estimate your FP rate, multiply by your CAC + LTV. Most stores that run this exercise discover they're spending more on false positives than on false negatives, and the dashboard they've been watching never told them.

Once you see the number, the calibration conversation with your fraud tool changes. Higher thresholds, multi-signal gating, friction-before-block. The fraud-caught number on the dashboard goes down slightly. The customer-retention and conversion-rate numbers go up measurably. And net margin, which is the only thing that actually matters, improves.

Read the full guideShopify fraud protection in 2026What Shopify's fraud analysis, Fraud Control, Flow and Shopify Protect cover, and what comes after the order.
Related use caseGrowing StoresGive support and ops teams a faster way to separate routine returns from the ones that merit manual review during busy periods.

RefundSentry Team

RefundSentry Team is a contributor to the RefundSentry blog, sharing insights on return fraud prevention and e-commerce best practices for Shopify merchants.