◆Invalid-traffic detection for performance advertising
Real is not the same as worth paying for.
Somewhere in last month’s invoice are clicks you would not defend. Nobody on either side of that invoice can point at one and say which.
Metaphasic reads every click and returns one status. Valid, or invalid with the reason named and the evidence attached. It weighs device, network, behaviour and reputation against one another rather than one signal at a time, which is what it takes against traffic that is now generated by machine.
Then it learns to read intent. A belief, 1 to 10, that a real click meant to act. Not a promise of a sale, and it sharpens the more of your traffic it sees.
Fraud detection tells you what to block. Metaphasic tells you what has intent.
Fraud is deterministic. Intent is not.
A click is fraud or it is not. There is no degree of fraud, and there is no amount of it Metaphasic calls acceptable.
Every other tool in this category sells you a slider anyway. It collapses the two, returns one probability, and hands you the threshold — the hardest part of its job, priced as your liability. Set it at 70 and you have agreed to a level of fraud. Set it at 90 and you have agreed to a different one. Nobody should have to pick the cutoff.
So we don’t. The status is decided: valid or invalid, with the reason named, and we are accountable for it. Intent is a belief, 1 to 10, and we say that out loud rather than dressing it as a finding.
You find out which one you bought on the call where a partner asks why you rejected their traffic. Seventy-three out of a hundred is not an answer to that question. Invalid — DataCenter, and here is the block it came from is an answer.
Every click comes back decided.
One status per click: valid, or invalid. Invalid names one reason — the highest that fired, out of eighteen in four families — and carries the evidence for every reason that fired, not only the one we named. Valid carries a degree of intent, 1 to 10, once you turn that on. There is nothing to tune, because we never hand you a number. When an evaluation can’t complete we return Unscored and say so. We never default a click into a status it didn’t earn.
The eighteen names are below, defined in a sentence each. Read them once and you will recognise every answer that can come back.
{
"status": "valid",
"intent": 8
}{
"status": "invalid",
"reason": "DataCenter",
"evidence": [
{ "category": "DataCenter", "evidence": "IP block registered to a cloud hosting network" },
{ "category": "Proxy", "evidence": "exit node matches a known proxy pool" }
]
}intent returns only with Detection + Intent, and only on a valid status — our belief, never a promise of a sale.A real person, and nothing contradicted it. This is the click you agreed to pay for. With Detection + Intent on, it carries a degree of intent, 1 to 10.
Something contradicted it. One reason is named — the one that outranks the others — and everything else that fired comes back alongside it, with its evidence.
All of them, defined. No invalid click comes back on your traffic without one of these eighteen names on it, and each one arrives with the evidence behind it.
A server presenting itself as a browser — there is no device behind the request at all.
The request comes from datacenter infrastructure, not a residential or mobile network.
A script is driving the request, with no human input anywhere in it.
A known crawler or automated agent, declared or detected.
The declared user agent doesn’t match how the device actually behaves.
The user agent string doesn’t parse to a real browser or device.
The client has been altered to hide or misstate what it is.
The app or SDK making the call isn’t the one it claims to be.
The device’s own signals contradict each other.
The mobile app sending the request fails validation against its own listing.
The request came from a domain not authorized to send it.
A device configuration known to be compromised or open to abuse.
Routed through a proxy or VPN that hides where it actually came from.
The origin has been deliberately concealed, so there is nothing behind it we can attest to.
This user has been flagged before, and that history stands against this click.
This source has a settled history that counts against it, independent of this click.
Volume or pattern from this source exceeds what real use produces.
This click has already been counted once.
One tag. One call. A verdict you can point to.
Drop in one tag.
One script on the page. No SDK to maintain, no separate stack to run.
Call the API.
One request per click. The status comes back with its reason and the evidence behind it, in the same response.
It’s on the record before you need it.
The status, the reason and every piece of evidence are written to that click’s record when it happens — not reconstructed later, because later there is nothing left to reconstruct from.
Six months from now, a clawback is a document you export. Not an argument you try to win from memory.
Records stay exportable for twelve months. Built for anyone who pays for clicks — brands, ecommerce sites, advertisers, affiliate programs, and the networks that supply them — and evaluated at the click, the traffic you actually pay for.
Detection is one tag and one call. There is nothing else to build for it.
A verdict only settles it if the other side trusts the device.
Most of what gets called fraud in a partner conversation is one side’s tool against the other side’s word. Two vendors, two numbers, one call, and somebody ends it eating the difference.
If the tool you run has a slider, your advertisers are being offered a degree of fraud. Ask them how they feel about that.
So a brand runs it on the clicks it buys, and the networks and publishers that sent them run it on their own traffic — on the same terms and the same price. Same two statuses. Same eighteen reasons. Same evidence — returned to whoever ran the check, before a buyer sees it rather than after.
Which is why the reasons are published and the price is published. A device only settles an argument if both sides can read it, and neither side got a different answer.
A fixed rule set is a target that sits still.
Fraud generation is automated now — scripted, distributed, cheap to iterate. Every rule that can be written down can be tested against until it passes, and the testing costs the other side almost nothing.
Metaphasic weighs device, network, behaviour and reputation against each other, and keeps learning as your traffic changes under it. A signal that stops meaning what it meant last month stops carrying weight this month. Nobody has to ship you a new rule, and you don’t have to be the one who notices.
The reading we won’t give you on day one.
Turned on, every valid click carries an intent score from 1 to 10: our belief that a real click meant to act. Never a promise of a sale.
It isn’t a tier you buy on day one. It becomes available in your account once the model has learned enough of your traffic, and then you switch it on yourself — at the published price, like everything else here.
We could hand you a number on the first day. It would be a guess dressed as a reading, and the entire argument of this page is that we don’t hand you those. It arrives calibrated on your traffic, or it doesn’t arrive.
Priced in public.
Two layers, because they do different jobs. Detection tells you what a click is. Intent tells you how strongly we believe a click meant to act. Both are priced per thousand evaluations, published, and applied to everyone at the same lines — at the first band, reading a million clicks is $150. Everyone else in this category makes you book a call to learn that number, and quotes the company next to you a different one.
What the click is. From the first call.
Every click comes back valid or invalid — invalid with the reason named and the evidence behind it, written to the record as it happens.
- Valid or invalid. One reason. Evidence attached.
- Evidence built for the clawback conversation.
- One tag, one call.
Plus how much intent it carries.
Everything in Detection, plus an intent score from 1 to 10 on every valid click — our belief about intent, not a promise of a sale. Not something you buy on day one: it becomes available in your account once we’ve learned enough of your traffic, and you turn it on yourself.
- Intent score, 1–10, on every valid click.
- A belief about intent, never a promise of a sale.
- Unlocks as we learn your traffic. You switch it on.
Shaped to your stack.
Custom deployments, volume commitments, and integration work for platforms that need something built to fit rather than dropped in.
- Custom deployment and SLAs.
- Volume commitments.
- Integration engineering.
All three go to the same place: an email, and a person replies with access. You start at the price above, which is the same price as the company next to you.
* Detection + Intent calibrates to your own traffic. It becomes available in your account once we’ve learned enough of it, and you enable it yourself. Detection runs from the first call.
Volume
The price steps down as your month-to-date volume crosses each line. Applied automatically, to everyone, at the same lines.
| Evaluations per month | Detection | Detection + Intent* |
|---|---|---|
| Up to 10M | $0.15 CPM | $0.20 CPM |
| 10M to 100M | $0.12 CPM | $0.16 CPM |
| 100M to 1B | $0.09 CPM | $0.12 CPM |
| Over 1B | Talk to us | Talk to us |
Up to $5,000 per purchase, raised on request. You set the auto top-up threshold and level when you buy.
Know what pays.
Run your own traffic through it and read the verdict yourself. You already know which of your sources you would defend and which you would quietly drop. That makes you the only qualified judge of whether we told them apart.
The status, the reason and the evidence are written to a click’s record at the moment it happens. Nothing writes them afterwards. Whatever you buy between now and the day the tag goes on is unrecorded, and there is nothing to export from a month nobody was reading.
Tell us what you want to run. A person replies — not a sequence — and starting is one tag and one call.