Honesty is a feature.
Security products earn trust by being right about what they can’t do, not just what they can. PubSentry is the AI Security Engineer for ad publishers — it blocks invalid traffic before the ad fires, and it tells you exactly where the line of catchable fraud is today, never a fake number to paper over it. Here’s every promise we hold ourselves to.
FPR = 0, by covenant.
Blocking a real reader is the worst thing this product can do — worse than missing a bot. A wrongly-blocked human is a lost reader, a support ticket, and a reason to churn. So a false-positive rate of zero isn’t a target we hope to hit; it’s the constraint everything else is built around.
Conservative by default
When a verdict is uncertain, we serve the ad. The defaults are tuned so the engine would rather miss a borderline bot than risk a single real human — you opt into more aggression deliberately, with eyes open.
A CI calibration gate
Every change to the scoring engine runs against a human test corpus, and the build asserts FPR = 0 on it. If a tweak would block a real reader, it doesn’t ship. The covenant is enforced in CI, not just in copy.
Blast-radius preview
Before any rule or IP block goes live, we replay it against your real recent traffic and show exactly who it would have caught — retroactively. You see the collateral before you commit, not after.
The detection boundary — what we catch, and what we don’t.
Most vendors quote one big recall number and let you assume it covers everything. It never does. Invalid traffic comes in two flavors, and they are not equally catchable. We’ll tell you exactly where the line is.
Obvious bots: ~100% recall, 0 false-positives
Headless browsers, automation tells, datacenter and known-bad ASNs, honeypot trips, impossible behavior — general invalid traffic (GIVT) leaves fingerprints. The deterministic engine catches essentially all of it without ever touching a real reader.
Sophisticated mimicry: ~3% today, by design
A bot on a residential IP with a real browser and human-like timing leaves almost no tell. Catching it on signals alone would mean blocking real people — so we don’t. That ~3% is a deliberate floor, not a bug.
How the gap closes
Sophisticated invalid traffic (SIVT) is closed by scale, not by a riskier gate: the cross-site reputation network (provisional until scale) and the ML moat both sharpen as traffic flows. An entity flagged on any PubSentry site is pre-flagged on yours.
“Awaiting traffic,” not a fake score
Before your site has data, the dashboard says awaiting traffic — not an invented 99.9%. The day-one numbers you see are the day-one numbers we actually have. We’d rather show you a blank than a lie.
We score visitors without hoarding them.
Fraud detection needs signals, not identities. PubSentry is built so the data that could identify a real person is hashed and dropped before it ever lands in storage — the engine keeps the verdict, not the visitor.
Raw IP & User-Agent are dropped
The server enriches datacenter / ASN / geo from the request, then HMAC-SHA256 hashes the raw IP and User-Agent and discards the originals. They are never written to storage in raw form. The verdict survives; the PII doesn’t.
Fingerprints are non-PII
Client-side device fingerprints use a fast fnv1a hash — a non-reversible bucket for spotting the same automation twice, not an identity. The tag only sends client-knowable fields; the server enriches the rest and drops the raw inputs.
Same-origin sessions
Dashboard auth is scrypt-hashed passwords and opaque, same-origin cookie sessions — no third-party tokens, no cross-site cookies. The control room sets no CORS; only your own browser talks to it.
Per-account scoping
Every API call is gated by a session and scoped to the sites your account owns. One account can never read another’s traffic, verdicts, or config — isolation is enforced on every request, not assumed.
The only honest place to stop fraud is before the ad fires.
Once an impression is counted or a click is billed, the damage is already on your account — the fraud happened, and the best a report can do is tell you about it. PubSentry makes the verdict before the ad loads. A blocked impression was never served, a blocked click was never billed, and your AdSense account never sees the spike.
- The verdict lands in under 5 ms, locally, before any ad request goes out
- Suppression across GPT, AdSense, Amazon apstag and Prebid — plus a universal MutationObserver fallback
- The same scoring engine runs in the tag, at the edge, and on the server — a parity test guards it
- It fails open: any error serves the ad. One tag, under 15 KB, ad-network-agnostic, never delays the page
Five promises, in one place.
We won’t block a real human
FPR = 0 is the constraint, enforced by the CI calibration gate and the blast-radius preview. The whole engine bends around it.
We won’t show a fake number
No invented scores, no claim of perfect detection. Before there’s data, the dashboard says “awaiting traffic.” What you see is what we measured.
We won’t hoard your visitors
Raw IP and User-Agent are HMAC-hashed and dropped server-side. Fingerprints are non-PII fnv1a buckets. The verdict is kept; the person isn’t.
The ML never decides on the hot path
Rules decide the verdict; the AI only narrates it. No black box gets to block a reader unsupervised — every block is explainable.
We stop fraud before it’s billed
Block-before-serve, not report-after. A blocked impression was never served and a blocked click was never billed to your account.
We’ll tell you what we can’t catch
~100% on obvious bots, ~3% on sophisticated mimicry today, closing as the network scales. The boundary is in the docs, not hidden.
Protect your account — honestly.
One tag, live verdicts in minutes, and a product that tells you exactly what it can and can’t catch. Free to start — you only see the value once your real traffic is scored.