What invalid traffic is
Part of the traffic you pay for never becomes a customer — bots, click farms, visits with no intent behind them. AdPix separates that traffic out, and it is deliberately built to miss some fraud rather than accuse one of your real customers.
What the term means#
You pay ad networks to send people. Some of what arrives is not a person, and some of it is a person who was never going to become a customer. The industry calls both invalid traffic.
The working definition is narrow: a visit that cannot be taken as evidence that a real human was interested in your business. It says nothing about intent — and that is exactly where the term is most often misread.
AdPix does not install a new tag to find this traffic. The data already being collected — dwell time, scroll depth, clicks, conversions, device fingerprint, source and channel, network and country — is read a second time, with fraud in mind. The real advantage is that AdPix sees the same IPs and fingerprints across many independent businesses, so it can recognise a pattern none of them could see alone.
Two layers: GIVT and SIVT#
The industry standard splits invalid traffic in two, and AdPix implements both layers.
| GIVT — general invalid traffic | SIVT — sophisticated invalid traffic | |
|---|---|---|
| What it is | traffic a simple list identifies | traffic only its behaviour gives away |
| Examples | search crawlers, uptime bots, requests from a data center, proxies, Tor | a click farm staffed by real people, incentivised users, an ad click that leaves after two seconds |
| How it is detected | list matching — known data-center, proxy and Tor address ranges | several signals that have to fire together |
| Certainty | high and cheap | needs corroboration |
In AdPix, GIVT is a simple flag set on each session as it is collected: did any event arrive from a known data-center, proxy or Tor address range. Sources carrying a lot of it are then held out of the "normal" the behavioural detectors are trained on, for a technical reason — if machine traffic stays inside normal, the model learns that machine behaviour is normal and then nothing looks anomalous. Obvious bots, the ones that announce themselves in the User-Agent, are quarantined even earlier, at collection time, and never enter your reports at all.
SIVT is the hard target and the reason the module exists. Here the other side may be a real human, in a real browser, on a real home connection. No list catches that. What gives it away is the pattern: hundreds of sessions with almost identical dwell times, one device fingerprint seen across several unrelated businesses, dozens of visits from one address block inside a single hour, none of which ever scrolled.
Invalid is not the same as malicious#
This is the most important sentence on the page. The invalid label is not a moral judgement, it is an accounting one: this visit should not count as genuine interest, and if you paid for it you probably should not have.
Things that are entirely invalid with nobody acting in bad faith:
- your own uptime monitor opening the home page every minute;
- your engineering team's QA bot;
- your ad agency verifying banner placements from one shared IP;
- a price-comparison crawler reading a product page;
- and a person who clicked the ad, realised it was not what they wanted, and left.
And things that genuinely are hostile: a click farm burning a competitor's budget, a network using residential proxies to pass bots off as users, traffic manufactured by a dishonest publisher to earn commission.
AdPix counts both in the same column, because intent cannot be recovered from data. What can be recovered is that your money went to something that does not become a customer. If the first kind is deliberate and known on your property, the answer is a referral exclusion or a false-positive report — not a debate about the definition.
AdPix deliberately favours precision over recall#
Read this before the first number, because everything the product does follows from this one decision.
Fraud detection has two error modes. One is missing fraud. The other is declaring your own real customer fraudulent. The second is the expensive one: tell someone once that a channel which genuinely drives sales is fake, watch them switch it off, and the product has both cost them money and lost their trust permanently. AdPix takes that position, and the whole scoring path leans towards precision.
In practice that means:
- An entity that converts is never auto-flagged. Conversion evidence is not only
purchase: the property's registered key events, session revenue and the conversion rate rolled up to the source itself all count. The bar is 2%. - The absence of conversion is never evidence of fraud. That is the deliberate mirror of the rule above: a zero conversion rate means bad targeting or an awareness campaign, and on its own it never causes a flag.
- Volume floors apply. Below three sessions nothing on any scoring path goes beyond review. A referrer with one visit always sits in the tail of the distribution, and "confirmed fraud" on a single visit is not defensible to a customer.
- A shared IP is not fraud. Corporate networks, mobile-carrier CGNAT and public Wi-Fi all share addresses. Fingerprint sharing is weighted above IP sharing, and IP sharing alone never flags anything.
- A VPN is not fraud. It is a weak contributing signal, never a verdict.
- Traffic inside your own market is never "sanctioned origin". AdPix infers each property's home market from its own data — the country supplying at least half the sessions — and excludes it from that overlay. For a business in Iran selling to Iranians, Iranian traffic is simply normal traffic.
Four verdicts, and what the invalid rate actually counts#
AdPix never returns a binary fraud/clean label. Every traffic slice gets one of four verdicts:
| Verdict | What it means |
|---|---|
| Clean | no evidence worth acting on |
| Review | something is there, but not enough to state automatically — a human decides |
| Potential | credible evidence, short of certainty |
| Confirmed | high model agreement plus structural corroboration |
And one detail that decides how you read every number on the dashboard: the invalid rate counts confirmed and potential only. Traffic pending human review is shown separately and is never counted as invalid. The rate is also session-weighted, not source-weighted — ten small flagged sources out of a hundred is not "10% invalid traffic".
What this product is not#
- Not a real-time filter. Scoring is batch and runs roughly daily. Detection latency of up to about 24 hours is accepted, because the job is investigation and ad-network accountability, not edge blocking.
- Not a guarantee. It reduces and exposes invalid traffic; it does not claim perfect detection. The other side evolves too.
- Not your decision. What it produces is evidence: what, how much, on which signal, and how much of your budget. Blocking a network or disputing a charge is still your call.
Where to go next#
If your property's plan includes fraud protection, start with the overview and severity tiers to learn what each number on the page says and what each tier asks you to do. To understand why one specific slice was flagged, read why this was flagged. If you think a verdict is wrong, sensitivity and false positives is the right path.
Frequently asked questions#
Does invalid mean someone is attacking me?
Usually not. Invalid means the visit cannot be counted as a real, interested person — an uptime monitor, a crawler, an agency checking ad placements from a shared IP, or someone who left after two seconds. Only a subset of it is hostile.
Does AdPix block invalid traffic?
Not in real time. Scoring is a batch job that runs roughly once a day, so the product is built for investigation and for holding ad networks accountable, not for blocking at the edge. You can export what it finds as an exclusion list for Google Ads or Meta.
Why is a source with zero conversions marked clean?
Because the absence of conversion is never evidence of fraud. Bad targeting, brand-awareness campaigns and plain browsing all convert near zero, and no model can tell a bot from an uninterested human on that basis. Conversion is used only in the positive direction: a source that converts is treated as real.
So does AdPix catch everything?
No, and it does not claim to. This is an arms race, and every threshold that keeps false positives low also lets some real fraud through. AdPix chooses that side of the trade deliberately.
Thanks — your feedback helps us improve the docs.