Why rules are no longer enough
Classic rules (“>5 clicks from an IP”, “geo not in whitelist”) only catch primitive fraud. Modern bots rotate residential IP and mimic scroll and time-on-site. ML looks for patterns, which a person would not write in if-else: correlations of device fingerprint, referrer, hour, and navigation speed.
Three pillars of ML antifraud
- Device / browser fingerprint. Not just User-Agent — canvas, WebGL, timezone drift, language mismatches. One “person” with a dozen fingerprints is a farm signal.
- Behavioral biometrics. Mouse trajectory, scroll rhythm, pauses between actions. Synthetic traffic is often “too perfect”.
- Anomaly detection. Unsupervised models catch spikes without labeled fraud — useful for zero-day schemes.
AI Selena in ClikBy: what we claim honestly
AI Selena — an ensemble of models for assessing click and visit quality on smart links: conversion score, statuses confirmed / bot_server / bot_filter. The system analyzes technical and behavioral signals on redirect and on the site after the click.
Мы не position Selena as a replacement for bank-grade transaction ML, do not guarantee “100% blocking of all bots,” and do not issue certificates the product does not have. The value is — seeing junk traffic in the marketing funnel more clearly and not feeding it into retargeting.
“ML in antifraud is an arms race. Our job is not to promise magic, but to give marketers a measurable click-quality layer before they burn through a quarterly budget.”
How not to overdo ML at rollout
- Start in observation mode: watch the bot share for 2–4 weeks without hard blocks.
- Check scoring against business metrics (lead quality, sales), not only against “% bot”.
- Document false positives — in fintech/e-com they cost more than a missed bot.
More on the terms: antifraud glossary.
Supervised vs unsupervised in practice
Supervised models need labeled bot/confirmed data — they are accurate where the pattern has already been seen. Unsupervised catch new clusters but produce more false positives. Selena combines the approaches in an ensemble.
Drift and retraining
Bots adapt to filters within weeks. ClikBy updates scoring on the platform; clients should revisit campaign thresholds and Ads exclusion lists every quarter.
