Watchtower scores clicks and conversions in milliseconds and stops fraudulent traffic before it ever reaches your advertiser — while it leaves clean users alone.
To the engine, a “click” is an abstract traffic event — an offer, a source, some identifiers, a postback. The same scoring plugs into rewarded offerwalls, affiliate trackers, and the mobile measurement partners app marketers run on.
The original home. Score the click before the advertiser redirect, hold the payout before it clears.
Same shape as an offerwall. Native support, no adapter needed.
Point Offer18 postbacks at Watchtower with their own macros — we translate them.
Maps sub1–sub5, transaction_id and payout onto canonical fields.
The partner-marketing platform (ex-HasOffers). Affiliate S2S — not to be confused with its old attribution arm, now Branch.
More performance trackers out of the box, plus a generic adapter for anything else.
The largest measurement partner. Feed its install/event postbacks to score traffic by media source.
Callback placeholders (gps_adid, idfa, click_id) map straight in.
The rest of the MMP field, so an app marketer can vet partners wherever they measure.
A fully config-driven adapter maps your own tracker’s macros — no code to write.
Running paid campaigns? Grade the partners sending you traffic instead of blocking single clicks.
A clean /v1 REST surface and a lightweight client SDK for everyone else.
No one signal blocks a user. Watchtower gathers evidence across independent families, decays repeats so a farm can’t stack the same trick, and only acts when the breadth of evidence clears a tier. Try it — tick signals and watch the verdict move.
Cheap detectors run before the redirect — IP intelligence, device integrity, velocity, behavioural timing and biometrics, user-agent forensics.
The second and third signal in the same family count for less. Ten flavours of “datacenter IP” are still one idea, not ten reasons.
A real history of clean, non-reversed conversions offsets some evidence — but never launders the farm tells (emulators, rings, impossible travel).
Points become a 0–100 score on a saturating curve, and a block needs at least two independent families — the single-signal false positive can’t happen.
Graph clusters, cross-tenant reputation, attestation and the ML model run async and can hold a payout after the fact.
The same verdicts answer two different questions — depending on whether you run the traffic or buy it.
Refuse the fraudulent click before it costs you an advertiser relationship, and withhold the payout before it clears. Inline and blocking, or async and advisory — you choose what sits on your critical path.
Running campaigns across media partners? Get a rolling quality score per sub-source, worst first, so you cut or renegotiate the bad ones instead of chasing individual clicks.
| Sub-source | Clicks | Quality | Grade |
|---|---|---|---|
| partner_44 | 18,204 | 96 | A |
| partner_07 | 9,881 | 71 | C |
| partner_19 | 6,442 | 38 | F |
Watchtower runs a privacy-preserving reputation network across every tenant. A device farm burned on one publisher is already suspect on yours — before it costs you a cent.
Every layer is an additive second opinion — explainable, measured before it’s trusted, and never a lone blocker.
A logistic model trained on your own labels, emitting a risk band with per-feature contributions. “Network and graph drove this,” with numbers — not a black box.
Pointer timing, tap cadence and pressure. It catches the click with no human interaction, superhuman speed, or metronome-regular rhythm.
Union-find over shared devices and networks surfaces the whole ring — 40 accounts on 2 devices is a farm; the console kills it in one click.
Play Integrity and App Attest verified server-side behind single-use nonces. ships dark
A live feed beside a z-score anomaly monitor that flags a signal spiking above its own baseline — the shape a fresh attack takes first.
A hill-climb over labelled backtests proposes threshold changes optimised for precision. It suggests; a human approves. Never auto-applied.
Every tier runs the full detection engine — you pay for volume, the reputation network, and operational tooling, never for weaker protection. Priced per evaluation (a conversion counts the same as a click); a monthly fee includes a bucket, then a low overage that falls as you grow. Labels are always free.
Recommended list pricing, month-to-month or annual (2 months free). Effective rate runs ~$0.001/eval on Starter down to ~$0.0002/eval at Scale — well under per-lookup fraud APIs, and overage keeps every tier extensible so you stay self-serve to ~30M before Enterprise. The reputation network is the differentiator.
Send a click, read the verdict, send labels back. The system gets sharper the moment your ground truth starts flowing.
Provision a live and a test key. Start every integration in shadow mode — we report, you don’t act yet.
POST canonical fields to /v1/clicks, or point your tracker’s postback at an integration adapter.
Async by webhook, or inline for a blocking answer. Every verdict carries its signals and score.
Advertiser reversals and approvals become ground truth. Labels are free and unmetered — they’re how it learns.
Tell us where your traffic comes from and we’ll set you up with sandbox keys and a quickstart matched to your platform. No card, and shadow mode means zero risk to your funnel.
Prefer to talk first? The assistant in the corner answers integration questions any time.