Free Trial Abuse in SaaS and AI: Patterns and How to Stop It
Free trial abuse burns AI compute: Stripe blocked 550,000+ high-risk trials at four AI firms in 2 months. Which trials to grant, hold, verify or refuse.

Stripe says its models saw abuse at sign-up rise sharply in the first half of 2026: "During sign-up, AI subscription companies saw a 40% increase in attempted multi-account abuse in a six-month period." The period was January to June 2026, and the most affected companies "saw a 154% increase in that same six-month time period, with some seeing over 600% increases."
The hard part is that the networks that look worst are often real trialers. A corporate security gateway, Apple's relay and a cloud server all look like infrastructure. And serial trials arrive every 7, 14 or 30 days, when the last trial ends, not in an hourly burst.
Below: the six patterns, what six addresses returned from GeoIPHub's API on 20 November 2026 and the trial decision each gets, tested code that counts trials over 30 days, and the friction tiers that cost real users least.
What free trial abuse is, and why AI free credits cost compute
Free trial abuse targets the grant, not the account: whatever you give away before payment, whether days, seats, tokens or API credits. The account is only the wrapper for claiming it again. In seat-based SaaS the cost is seats, storage, support time and trial-to-paid numbers that no longer mean anything.
Stripe describes the mechanics in its September 2026 data post: actors "create many accounts during sign-up to repeatedly claim free tokens, trials, and other new account benefits." Those figures come from Stripe's own models and count attempts; Stripe does not publish a method or a false-positive rate.
For AI products the cost is compute. Stripe's March 2026 first-party fraud post puts it plainly: "When a bad actor spins up five accounts instead of one, they consume five times the compute resources." Its Radar launch post gives a size: "In the first 2 months, across 4 high-growth AI businesses, we blocked more than 550,000 free trials with a high risk of abuse, preventing an estimated $4.4 million in downstream losses from compute costs." Our arithmetic, not Stripe's: that estimate works out to at most about $8 of compute per blocked trial.
Free trial abuse is the trial and free-credit branch of promo abuse, which Stripe files under first-party fraud with account and refund abuse. Coupons and referral bonuses are covered there, and the check before an account exists in our guide to fake signups. This post sits at the trial grant itself, and uses the IP fraud score as one input, not as the policy.
Six free trial abuse patterns and the trace each leaves
The six free trial abuse patterns are serial trials, disposable and aliased email, virtual and prepaid cards, VPN and proxy rotation, cloud automation, and AI credit farming. The email and card patterns leave no trace in the IP address; the other four can.
Serial trials
The same person signs up again with a new email the day the last trial ends. Trace: a repeat IPv4 /24 or IPv6 /64 at trial-length intervals. To go further, cluster the accounts behind them on device and payment signals.
Disposable and aliased email
Throwaway domains, plus-addressing (name+1@) and dot variants on providers that ignore dots let one inbox open many trials. Trace: the email layer, not the IP. The limit is that privacy-minded users alias too, so an alias alone is a weak reason to refuse.
Virtual and prepaid cards
Stripe's free trial abuse docs say: "Common patterns include using prepaid or virtual cards, or repeatedly signing up with the same card across multiple accounts." The catch, from Stripe's first-party fraud post: "Blanket blocking them means rejecting legitimate sign-ups and hurting conversion rates." Trace: the payment layer.
VPN and proxy rotation
VPN and proxy rotation gives each new trial a fresh exit address, which hides the repeat. Trace: is_vpn and is_proxy in the lookup. A fresh residential or mobile exit may carry neither flag, which is the IP layer's main blind spot (last section).
Cloud automation
Scripted signups and API calls run from cheap, disposable servers. Trace: is_hosting set to true, and many trials from one hosting network (ASN) in a month.
AI credit farming and resale
Stripe's September post describes the paid version: "A fraudulent actor may use a stolen card to buy an AI subscription or a block of tokens, then sell that access on secondary markets or in regions where the product isn’t directly available." Free credits get the same treatment, and so do the free API keys themselves. Trace: credits burned to the cap within hours, from hosting networks, and many keys per prefix.
Trial abuse detection at trial start: six lookups and the decision each gets
A GeoIPHub lookup at trial start returns the network type, the detection flags, a fraud score from 0 to 100 and a recommended action. The trial decision then asks what the score doesn't: what does this address get for free? The decisions below are what the basePolicy function in our code returns for each lookup.
Looked up 20 November 2026, 07:58 to 07:59 UTC, keyless. None of these addresses started a trial; each stands in for a kind of network. The decisions follow the API's recommended_action bands; if you use our stricter starting signup thresholds, map 71 and up to refuse the trial. You can check any address's score the same way; the free tier is 1,500 lookups a day with every field and no card, and keyless requests get 30 a minute.
Microsoft's corporate network
{
"ip": "131.107.1.1",
"asn": {
"asn": 3598,
"org": "MICROSOFT-CORP-AS - Microsoft Corporation",
"user_type": "none"
},
"scoring": {
"fraud_score": 0,
"recommended_action": "allow",
"coverage": "unobserved"
}
}
Trial decision: grant. The 0 comes with coverage: "unobserved", which means no data, not a clean bill.
A Zscaler corporate gateway
{
"ip": "165.225.72.1",
"asn": {
"asn": 62044,
"org": "ZSCALER-EMEA - Zscaler Switzerland GmbH",
"user_type": "business"
},
"detection": {
"is_hosting": true,
"is_relay": true,
"relay_provider": "Zscaler",
"relay_type": "enterprise_gateway"
},
"scoring": {
"fraud_score": 24,
"recommended_action": "allow"
}
}
Trial decision: grant, with shared limits (ten times the prefix limit, below). This stands in for a B2B trialer behind a corporate gateway. It is is_hosting: true and user_type: "business" at once, so check is_relay before is_hosting.
iCloud Private Relay
{
"ip": "172.224.226.1",
"asn": {
"asn": 36183,
"user_type": "consumer_privacy_network"
},
"detection": {
"is_hosting": true,
"is_relay": true,
"relay_provider": "iCloud Private Relay",
"relay_type": "consumer_relay"
},
"scoring": {
"fraud_score": 19,
"recommended_action": "allow"
}
}
Trial decision: grant, with shared limits. A consumer relay that is also is_hosting: true. GeoIPHub caps relay scores at 60, so a relay address never reaches block.
A DigitalOcean cloud server
{
"ip": "159.89.100.1",
"asn": {
"asn": 14061,
"org": "DIGITALOCEAN-ASN - DigitalOcean, LLC",
"user_type": "hosting",
"connection_type": "datacenter"
},
"detection": {
"is_hosting": true,
"hosting_type": "cloud"
},
"scoring": {
"fraud_score": 15,
"recommended_action": "allow"
}
}
Trial decision: hold credits. The score says allow; the next section explains why the trial policy is stricter.
A Mullvad VPN server
{
"ip": "193.32.127.71",
"asn": {
"asn": 39351,
"org": "ESAB-AS - 31173 Services AB"
},
"detection": {
"is_vpn": true,
"vpn_provider": "Mullvad",
"vpn_type": "consumer",
"vpn_evidence": "provider_list",
"is_proxy": false,
"is_hosting": true
},
"scoring": {
"fraud_score": 70,
"recommended_action": "step_up"
}
}
Trial decision: verify before credits. GeoIPHub caps a named consumer VPN such as Mullvad at 70 (step_up) unless proxy, Tor, abuse or spam evidence lifts it. The same server's IPv6 address, 2a03:1b20:a:f011:6::f001, returned the same 70 step_up at 07:59 UTC.
An ExpressVPN server also seen relaying
{
"ip": "31.171.152.205",
"asn": {
"asn": 197706,
"org": "Keminet - Keminet SHPK"
},
"detection": {
"is_vpn": true,
"vpn_provider": "ExpressVPN",
"vpn_type": "consumer",
"is_proxy": true,
"proxy_type": "observed_proxy",
"is_hosting": true
},
"scoring": {
"fraud_score": 100,
"recommended_action": "block"
}
}
Trial decision: refuse the trial, offer the paid plan. A VPN alone would stop at 70, but is_proxy: true (a third-party list saw this address relaying) lifts the cap, and it scored 100. Refuse the trial, not the customer: a proxy exit can still buy.
Why a cloud server at 15 "allow" still gets its credits held
A cloud server that scores 15 allow still gets its credits held, because a trial policy asks what an address gets for free, not whether it is risky. A cloud server is low risk for a page view and a poor bet for free GPU time.
A score answers "is this address risky". A trial policy answers "what does this address get for free". 159.89.100.1 came back 15 allow, with is_hosting: true and hosting_type: "cloud": nothing bad is known about it, but whatever signs up from it is software on a rented machine.
So start the trial, release a small allowance, and hold the rest. Real developers test from servers, which is why this is a hold and not a refusal. Release the rest at verification or a usage event, such as the first job that needs more than the allowance, never at signup; a held trial that never verifies spends no GPU time.
We expected a named VPN to be the hardest case. It was the plain cloud server, which the score lets through. The next DigitalOcean trials inside 30 days are what the ledger below counts.
One trial per person: count trials per /24, IPv6 /64 and hosting network over 30 days
To prevent multiple free trials from one person, count trial grants per network block over the trial's lifetime, because a serial trialer returns when the last trial ends. GeoIPHub's API scores one address and counts nothing; each request takes one address, and the ledger is your code.
The hourly counter for account creation is in our guide to checking the IP before the account exists. This ledger counts trials, over 30 days:
import { isIPv4, isIPv6 } from "node:net";
const DAY = 864e5;
const WINDOW = 30 * DAY;
const LIMITS = { v4_24: 3, v6_64: 2, asn: 5 };
const SHARED_FACTOR = 10;
const RANK = ["grant", "hold_credits", "verify", "refuse_trial"];
function expandIPv6(addr) {
const [head, tail = ""] = addr.toLowerCase().split("::");
const h = head ? head.split(":") : [];
const t = tail ? tail.split(":") : [];
const fill = addr.includes("::") ? 8 - h.length - t.length : 0;
return [...h, ...Array(fill).fill("0"), ...t].map((g) => g.padStart(4, "0"));
}
export function trialKeys(ip, lookup) {
const mapped = ip.match(/^::ffff:(\d+\.\d+\.\d+\.\d+)$/i);
const v4 = mapped && isIPv4(mapped[1]) ? mapped[1] : ip;
const keys = [];
if (isIPv4(v4)) keys.push(["v4_24", v4.split(".").slice(0, 3).join(".") + ".0/24"]);
else if (isIPv6(ip)) keys.push(["v6_64", expandIPv6(ip).slice(0, 4).join(":") + "::/64"]);
const d = lookup?.detection;
if (d?.is_hosting && !d?.is_relay && lookup?.asn?.asn) keys.push(["asn", `AS${lookup.asn.asn}`]);
return keys;
}
export function basePolicy(lookup) {
if (!lookup) return "grant";
const action = lookup.scoring?.recommended_action;
if (lookup.detection?.is_tor || action === "block") return "refuse_trial";
if (action === "step_up") return "verify";
if (action === "review") return "hold_credits";
if (lookup.detection?.is_hosting && !lookup.detection?.is_relay) return "hold_credits";
return "grant";
}
export class TrialLedger {
constructor() {
this.grants = new Map();
}
recent(key, now) {
const kept = (this.grants.get(key) ?? []).filter((t) => now - t < WINDOW);
this.grants.set(key, kept);
return kept.length;
}
decide(ip, lookup, now = Date.now()) {
const d = lookup?.detection;
const shared = Boolean(d?.is_relay || d?.is_mobile || lookup?.threat?.is_cgnat);
let decision = basePolicy(lookup);
const over = [];
for (const [level, value] of trialKeys(ip, lookup)) {
const key = `${level}:${value}`;
const seen = this.recent(key, now);
const limit = shared ? LIMITS[level] * SHARED_FACTOR : LIMITS[level];
if (seen >= limit) {
over.push(`${key} has ${seen} trials in 30 days (limit ${limit})`);
const floor = level === "asn" ? "verify" : "hold_credits";
if (RANK.indexOf(decision) < RANK.indexOf(floor)) decision = floor;
}
}
return { decision, over };
}
record(ip, lookup, now = Date.now()) {
for (const [level, value] of trialKeys(ip, lookup)) {
const key = `${level}:${value}`;
this.grants.set(key, [...(this.grants.get(key) ?? []), now]);
}
}
}
Call ledger.decide(ip, lookup) at trial start, act on decision, then ledger.record(ip, lookup) once the trial is created. What it does:
- Keys. Every trial counts against its IPv4 /24 (an IPv4-mapped IPv6 address is folded back to IPv4) or its IPv6 /64. It also counts against
AS<asn>, but only whenis_hostingis true andis_relayis false, so Apple's and Zscaler's networks never pool into one hosting bucket. - Limits. Ours, not benchmarks: 3 per /24, 2 per /64 and 5 per hosting ASN in a rolling 30 days. Relays, mobile carriers and CGNAT get ten times the prefix limit. Never refuse a relay or gateway on volume.
- Escalation. Over a prefix limit, the trial is at least held; over the hosting-ASN limit, it must verify. Velocity never lowers a stronger decision, and a held trial still counts.
- No lookup. If the lookup fails (
null), the base decision is grant and the prefix keys are still counted. That is our choice; a team that prefers to fail closed can return"hold_credits"there.
We ran it under Node 24.20.0 on 20 November 2026 against lookups saved at 07:58 to 08:00 UTC; 8 tests pass. The timeline is synthetic: 198.51.100.0/24 and 2001:db8::/32 are documentation addresses with no lookup, and the rest are today's saved lookups replayed on invented days.
day 0 198.51.100.10 no lookup grant
day 0 159.89.100.1 15 allow hold_credits
day 3 165.225.72.1 24 allow grant
day 4 172.224.226.1 19 allow grant
day 5 131.107.1.1 0 allow grant
day 6 2a03:1b20:a:f011:6::f001 70 step_up verify
day 7 31.171.152.205 100 block refuse_trial
day 9 198.51.100.23 no lookup grant
day 10 159.89.101.1 15 allow hold_credits
day 12 2001:db8:4a:7::1 no lookup grant
day 14 104.248.50.21 15 allow hold_credits
day 16 2001:db8:4a:7:9c3e::2 no lookup grant
day 18 198.51.100.47 no lookup grant
day 19 185.220.101.47 100 block refuse_trial
day 21 2001:db8:4a:7:51::3 no lookup hold_credits v6_64:2001:0db8:004a:0007::/64 has 2 trials in 30 days (limit 2)
day 22 159.89.100.1 15 allow hold_credits
day 24 159.89.101.1 15 allow hold_credits
day 26 104.248.50.21 15 allow verify asn:AS14061 has 5 trials in 30 days (limit 5)
day 27 198.51.100.88 no lookup hold_credits v4_24:198.51.100.0/24 has 3 trials in 30 days (limit 3)
day 40 198.51.100.91 no lookup grant
Four lines carry the lesson. On day 21 the third trial in one /64 is held, though each address is new. On day 26 the sixth DigitalOcean trial in 30 days steps up to verify. On day 27 the fourth trial in the /24 is held. On day 40 the same /24 is granted again, because day 0 and day 9 have left the window.
Store two more counts per key: trials that hit the credit cap and trials that converted. A /24 or hosting ASN with many trials, many caps and no conversions is a farm; the IP check alone cannot tell you that.
The ledger lives in memory and grows without bound. In production, store timestamps per key in your database. For wider IPv6 rotation, see escalating to /56 and /48.
Countermeasures by friction tier, and what each costs a real user
To prevent free trial abuse at the lowest cost to real users, use five tiers of friction, from lightest to heaviest: grant full credits, hold credits, verify, keep the costly perk for paid plans, and refuse the trial. Use the cheapest tier that covers the risk.

- Grant full credits. The default for relays, corporate gateways and no-data addresses. Costs a real user: nothing.
- Hold credits. Create the account now, release a small allowance, and release the rest on verification or a first real action (the first value event). Look the IP up again at that moment, and store the score, the action and the lookup time. Costs a real user: a short wait before the full trial.
- Verify. Email confirmation, a phone code, a card on file, or a CAPTCHA for scripted starts. Costs a real user: a step, and some drop off at every step.
- Restrict the costly perk to paid plans. Keep the most expensive model or feature out of the free tier. Costs a real user: they evaluate a smaller product.
- Refuse the trial. Tor exits and anything scored block, such as a proxy exit, with the paid plan offered instead. Tor is the one flag the code refuses on by itself, whatever the score (the demo's day 19 exit scored 100). Costs a real user: the trial, though not the purchase.
Hold credits instead of refusing: a hold can be lifted the moment a person proves real.
Card or no-card trials: what a card buys you, and what it invites
Requiring a credit card for a free trial filters casual repeat trials and lets a payment network link accounts by card. It also costs signups, and it opens your trial form to card testing.
Stripe's docs describe the control: "Radar blocks subscriptions from customers who are unlikely to pay at the end of the free trial." It evaluates trial starts that collect a payment method. A card-less free tier or a credit grant at signup sits outside that control. Stripe's March post announces a separate Radar feature for account registration and login, in early access; without something like it, the IP check is what you have at the grant.
Adding a card has a known side effect. Stripe's card testing docs say: "Card testers use both card setup and payments to determine whether the stolen or enumerated card information they have is valid or not." Card setup "is a method preferred by fraudulent actors, because card validation and authorizations during card setup don't typically show up on cardholder statements."
Stripe is also honest about our layer: "simple firewall rules or filters based on a single heuristic such as IP addresses are usually not sufficient to prevent card testing on their own." That holds for trials too: the IP decides the friction, and the card checks run beside it. More on how card testing hits checkout.
Where the IP layer goes blind on free trial abuse
GeoIPHub can't vouch for a residential or mobile exit it hasn't observed; that exit passes as a home line. One person on a clean home line passes every IP check, however many trials they open. GeoIPHub flags proxies we have observed on residential and mobile lines, nothing more.
Shared addresses cut the other way. CGNAT, carriers and relays put many real users on one address, which is why the ledger gives them ten times the limit and never refuses on velocity. A lookup also describes the address now, not next week, so look it up again when credits are released.
The email, phone, card and device are your other layers. See device signals beside the IP for how they combine.
Paste a trial-start address from your own logs into the IP fraud score check and see which tier it lands in.
Frequently Asked Questions
What is free trial abuse?
One person or operator claiming a free trial, free tier or free credits repeatedly, usually through new accounts, to use a product without paying. In AI products each abused trial costs compute.
How do you detect free trial abuse?
Check the IP at trial start for Tor, proxy, VPN and hosting networks, count trials per IPv4 /24, IPv6 /64 and hosting network over 30 days, and link accounts on email, card and device signals. No single signal is enough.
Is free trial fraud the same as a consumer subscription scam?
No. Free trial fraud here means customers abusing a business's trial. A consumer subscription scam is a business tricking consumers into hidden recurring charges.
Should I require a credit card for a free trial?
A card filters casual repeat trials and lets payment tools link accounts by card, but it costs signups and invites card testing, where fraudsters validate stolen cards at card setup. Holding credits until verification is a cheaper middle step for card-less trials.
Does blocking VPNs stop free trial abuse?
Not on its own, and it turns away privacy users. GeoIPHub scores a named consumer VPN at most 70, a step-up, unless proxy or abuse evidence lifts it. Verify VPN trialers instead of refusing them.
How many free trials per IP address is too many?
There is no standard. Count per IPv4 /24, IPv6 /64 and hosting network over the trial length, start low (we use 3, 2 and 5 in 30 days) and give relays, mobile carriers and CGNAT ten times the limit.
How do AI companies stop free credit farming?
Hold most credits until verification or a first real action, look the IP up again at that moment with a service such as GeoIPHub, and count trials per network block. Restrict the most expensive features to paid plans.
The team behind the GeoIPHub API
Written by the GeoIPHub team, the people who build the GeoIPHub IP intelligence API and its scanner. Drafts are prepared with AI assistance and edited by the team; API examples are real lookups against the live API unless labelled illustrative.


