Seatext library / BotRefund evidence
How to Troubleshoot False Positives in Headless Chrome Detection
False positives in headless Chrome detection usually stem from relying on single signals like navigator.webdriver or user-agent strings. Fix them by auditing your detection logic, testing against real browser diversity, and implementing a multi-signal...
✓ Built for advertisers who need clear, refund-ready traffic evidence.
False positives happen when legitimate users — often on privacy-focused browsers, corporate networks, or unusual device configurations — trigger detection rules built for bots. The fix isn't to weaken detection; it's to make it smarter. Start by mapping every signal your system checks, then test each one against a representative sample of real traffic. Replace single-threshold rules with a weighted model that requires multiple anomalous signals before flagging a session.
Why False Positives Happen in Headless Chrome Detection
Most detection systems check for automation fingerprints: navigator.webdriver, missing Chrome runtime, inconsistent screen dimensions, or TLS fingerprint mismatches. These signals work well against naive bots but also flag legitimate scenarios. A developer using Chrome DevTools Protocol for testing, a user on a hardened browser like Brave or Tor, or someone behind a corporate proxy that strips headers can all look like headless Chrome to a simple rule set.
The core problem is treating each signal as a binary verdict. One signal can be misleading. BotRefund's prediction AI sees how 106 browser, network, hardware, and behavior signals fit together before deciding whether a visit is human or automated. Signals become a decision only when they are seen together. When you rely on a single property, you lose the context that distinguishes a privacy-conscious human from a script.
Common Detection Signals That Trigger False Positives
Understanding which signals cause the most collateral damage helps you prioritize fixes. The table below groups common detection vectors by category and notes their false-positive risk.
| Signal Category | Examples | False-Positive Risk | Why It Happens |
|---|---|---|---|
| Automation flags | navigator.webdriver, CDP debugger leak, automation properties | High | Developers, QA tools, and some extensions set these legitimately |
| Browser integrity | Native patching, engine mismatch, JS engine mismatch, rebrowser leaks | Medium | Modified browsers, electron apps, and privacy hardening change internals |
| Network & geolocation | WebRTC leak, DNS tunnel leak, timezone evasion, latency mismatch, IP inconsistency | High | VPNs, corporate proxies, satellite internet, and privacy tools create mismatches |
| Behavioral | Mouse tremor absence, superhuman speed, grid-aligned movement, session duration anomalies | Low to Medium | Accessibility tools, motor impairments, and fast readers can mimic bot patterns |
| Environment | User-agent mismatch, accept-language mismatch, HTTP protocol mismatch, suspicious ports | Medium | Privacy browsers, translation proxies, and non-standard network setups |
Network and automation signals carry the highest false-positive risk because legitimate infrastructure (VPNs, corporate proxies, developer tools) routinely produces the same anomalies that bots do. Behavioral signals are more reliable but require enough session data to evaluate.
Step-by-Step Troubleshooting Process
- Inventory your current rules. Export every detection rule, threshold, and signal weight. Note which are hard blocks versus score contributors.
- Collect a labeled sample. Pull 1,000–5,000 recent sessions with known outcomes (converted, bounced, support tickets, chargebacks). Include edge cases: mobile, tablet, desktop, VPN, corporate, privacy browsers.
- Run each signal in isolation. For every rule, calculate its false-positive rate against your labeled human traffic. Flag any signal with >2% false positives on clean traffic.
- Correlate signals. Build a co-occurrence matrix: how often do Signal A and Signal B appear together on human vs. bot traffic? Real bots rarely trigger just one anomaly.
- Redesign as a weighted model. Assign each signal a weight based on its predictive value. Require a minimum combined score (e.g., 3 of 5 high-weight signals) before flagging. This mirrors how BotRefund evaluates the full pattern rather than one suspicious browser property.
- Add a whitelist layer. Maintain an allowlist for known-good ASNs, user-agent patterns, and behavioral profiles (e.g., your own monitoring tools, partner crawlers). Review monthly.
- Implement shadow mode. Run the new model in parallel with the old one. Log discrepancies without blocking. After two weeks, compare outcomes.
- Gradual rollout. Enable blocking for the highest-confidence tier first. Monitor support tickets for "I can't access the site" complaints.
Common Mistake: Over-Reliance on Single Signals
The most frequent error is treating navigator.webdriver === true or a headless user-agent string as proof of automation. Modern bots spoof these trivially. Meanwhile, legitimate users on Electron apps, automated testing frameworks, or privacy-hardened browsers trigger them constantly. A single-signal block list creates a maintenance nightmare: every browser update or new privacy tool adds false positives.
Server-side audits look at server log files. They monitor IP addresses, request headers, and user-agent data. While this catches basic scraper bots, it struggles to detect advanced botnets. Client-side audits analyze the visitor's browser environment and behavior, catching what server logs miss. The same principle applies to false positives: client-side behavioral context (mouse movement, scroll patterns, interaction timing) distinguishes a human on a weird browser from a script on a normal one.
Verification: How to Confirm Your Fixes Work
After deploying the weighted model, verify with three checks:
- False-positive rate: Track "blocked but converted" sessions. If users who eventually convert were blocked, your threshold is too aggressive.
- Bot catch rate: Run a controlled test with known bot traffic (your own Puppeteer/Playwright scripts, residential proxy services). The model should catch >95% without tuning per bot type.
- Support volume: Monitor "access denied" tickets. A spike indicates a new false-positive pattern.
Set up a weekly review dashboard showing: total sessions, flagged sessions, false-positive estimates (from conversion correlation), and top triggering signal combinations. Adjust weights monthly.
Key Facts About BotRefund's Detection Approach
| Aspect | Detail |
|---|---|
| Signal count | 106 combined browser, network, hardware, and behavior signals |
| Decision method | Prediction AI evaluates full pattern; no raw-signal scoring |
| Accuracy claim | 99% at classifying traffic as human or bot |
| Network vectors | WebRTC leak, DNS tunnel, timezone evasion, latency mismatch, IP inconsistency, OS/TCP TTL mismatch, HTTP user-agent mismatch, accept-language mismatch, HTTP protocol mismatch, DNS routing mismatch, suspicious ports, UTC timezone bias, languages mismatch, netprobe telemetry missing |
| Evasion & anti-stealth vectors | CDP debugger leak, native patching, engine mismatch, rebrowser leaks, JS engine mismatch, automation properties |
| Behavioral vectors | Ghost click detection, honeypot trap interactions, robotic linear mouse movements, absence of humanlike mouse tremor, superhuman input speed (<1ms), grid-aligned movement patterns, absence of clicks or scrolling, unnatural session durations |
| Refund integration | Auto-captures click IDs (GCLID, FBCLID) with behavioral evidence for Google/Meta disputes |
| Deployment | Install in about one minute; no credit card required for trial |
Limitations and When This Advice Doesn't Apply
- Low-traffic sites: Weighted models need volume to calibrate. Under 10k sessions/month, stick to a few high-confidence rules and manual review.
- Strict compliance environments: Some regulations (GDPR ePrivacy, CCPA) restrict client-side fingerprinting. Verify legal basis before deploying behavioral signals.
- Real-time blocking requirements: If you must block at the edge (CDN/WAF), you're limited to request headers and IP reputation. Client-side signals require page execution.
- Single-page apps with heavy client routing: Session behavior signals (duration, scroll depth) need adaptation for virtual navigation.
Terminology
- Headless Chrome: Chrome running without a visible UI, typically controlled via Puppeteer, Playwright, or CDP.
- CDP (Chrome DevTools Protocol): Debugging interface that automation tools use; its presence leaks automation.
- Fingerprinting: Collecting browser attributes (canvas, WebGL, fonts, audio) to create a stable identifier.
- Residential proxy: Proxy routing through consumer ISP IPs, making traffic appear residential.
- Pixel poisoning: Invalid conversions corrupting ad platform optimization (Meta Pixel, Google Ads conversion tracking).
- GCLID/FBCLID: Google/Meta click identifiers used to tie ad clicks to conversions for refund claims.
FAQ
How many signals should I combine before blocking?
Start with a threshold of 3 high-weight signals from different categories (network, automation, behavioral). Tune based on your false-positive rate. BotRefund uses 106 signals in a prediction model rather than a fixed count.
What's the fastest way to test a new rule without risking real users?
Shadow mode: run the rule in logging-only mode for 14 days. Compare flagged sessions against conversions and support tickets. Only enable blocking after false positives drop below your tolerance.
Do privacy browsers like Brave or Tor always trigger false positives?
They often trigger network and fingerprint signals (WebRTC leaks, timezone mismatches, canvas noise). Behavioral signals usually stay human. Weight network signals lower for known privacy-browser user-agent patterns.
How do I handle corporate proxy traffic that looks like a botnet?
Corporate egress IPs often share one IP across hundreds of users, creating high request rates and header stripping. Whitelist known corporate ASNs or use behavioral verification (mouse, scroll, timing) which remains human even behind a proxy.
Can I use this approach with a WAF like Cloudflare or AWS WAF?
WAFs only see request headers. For behavioral signals, you need client-side JavaScript. Use the WAF for IP reputation and known-bot UA blocking; layer client-side scoring for the rest.
What's the cost of getting this wrong?
False positives directly lose revenue (blocked buyers) and indirectly poison ad optimization (if you feed blocked sessions as conversions). BotRefund customers report up to 20% of ad spend wasted on invalid traffic; over-blocking real users wastes the remaining 80%.
How often should I retrain or reweight the model?
Monthly for high-volume sites (>100k sessions/month). Quarterly for lower volume. Trigger an immediate review after any major browser release (Chrome, Safari, Firefox) or when a new privacy tool gains traction.
Further reading and comparison sources
These external sources provide additional context for evaluating the topic. Their inclusion is not an endorsement.
Learn more
Visit the website for more information.