Seatext library / BotRefund evidence
Detection Rate vs Accuracy in Bot Detection: What Each Metric Tells You
Detection rate measures the percentage of actual bots that a system catches, while accuracy measures the overall percentage of correct decisions across both bots and humans. Because bot traffic is often a small slice...
✓ Built for advertisers who need clear, refund-ready traffic evidence.
Quick verdict
Detection rate (also called recall or true positive rate) answers: "Of all the bots that visited, how many did we flag?" Accuracy answers: "Of all visits — bots and humans — how many did we classify correctly?" When bots are rare, accuracy stays high even if the system lets most bots through. For ad-fraud refunds you need a high detection rate backed by evidence that satisfies Google and Meta.
| Criterion | Detection Rate (Recall) | Accuracy |
|---|---|---|
| What it measures | Share of real bots caught | Share of all visits classified correctly |
| Formula | True Positives / (True Positives + False Negatives) | (True Positives + True Negatives) / Total |
| Why it matters for ad fraud | Directly shows how much bot click spend you can prove | Can look impressive while missing most bots |
| Risk if used alone | May come with many false positives (blocking humans) | Hides poor bot catch-rate when bots are rare |
| BotRefund approach | 106 independent signals fed to AI to maximize catch-rate | Reported 99% accuracy from corroborated evidence |
| Practical takeaway | Ask for detection rate on your traffic mix | Treat as a secondary sanity check |
Why the distinction changes your refund outcome
Google and Meta refunds require proof that specific clicks came from bots. A system with 99% accuracy but 40% detection rate leaves 60% of bot clicks unproven — money you cannot recover. BotRefund's documentation emphasizes that each of its 106 checks (such as Empty Font Canvas, Suspicious Ports, Monitor Sync Anomaly) adds independent evidence, and the AI weighs the complete pattern instead of trusting a single rule [S1][S3][S6]. This design aims to push detection rate higher without inflating false positives.
The three-step process works like this: first, each signal adds one objective fact about the visit (independent evidence). Second, BotRefund tests whether other signals support the same story (cross-checked context). Third, the prediction AI weighs the complete pattern instead of trusting a raw rule [S1][S3][S6]. This corroboration is why BotRefund reports 99% accuracy while still targeting a high detection rate.
How the metrics behave when bot share is low
Imagine 1,000 visits with 50 bots (5%). A detector that flags 10 bots and 5 humans has: detection rate 20% (10/50), accuracy 98.5% (985/1000). The accuracy number looks great; the detection rate reveals the real problem. This is why the MIT Sloan study cited in search results warns that "bot detection models may return a high rate of accuracy, but that's due to a critical limitation in the data used to train them."
When bot traffic drops to 1%, a detector that labels everyone human achieves 99% accuracy and 0% detection rate. Always ask for detection rate on your traffic composition. The lower the bot share, the wider the gap between accuracy and detection rate.
How BotRefund's 106 signals feed detection rate and accuracy
BotRefund groups its 106 independent checks into categories: hardware & GPU fingerprinting, network/VPN/geolocation evasion, biometric & behavioral interactions, and console/debug evaluation [S1][S3][S6][S7]. Examples include:
- Empty Font Canvas — detects mismatch between claimed device and graphics/font rendering [S1].
- Suspicious Ports — flags proxy rotation or location masking that makes network facts disagree [S3].
- Monitor Sync Anomaly — spots timing and movement patterns that scripts struggle to reproduce [S6].
- Ghost click detection — catches clicks without the natural sequence of human intent [S2][S4][S5][S7][S8].
- Honeypot trap interactions — watches for bots responding to hidden page elements [S2][S4][S5][S7][S8].
- Robotic linear mouse movements — flags unnaturally straight pointer paths [S2][S4][S5][S7][S8].
- Absence of humanlike mouse tremor — looks for missing micro-jitter [S2][S4][S5][S7][S8].
- Superhuman input speed (<1ms) — identifies interactions faster than a person can perform [S2][S4][S5][S7][S8].
- Grid-aligned movement patterns — detects movement snapping to precise lines [S2][S4][S5][S7][S8].
- Absence of clicks or scrolling — highlights sessions too static to be real [S2][S4][S5][S7][S8].
- Unnatural session durations — catches visits too short, too long, or too uniform [S2][S4][S5][S7][S8].
Each signal is independent evidence. The AI prediction step combines them, so a single anomaly does not trigger a verdict. This reduces false positives while keeping detection rate high.
Key facts from BotRefund's detection architecture
| Fact | Detail | Source |
|---|---|---|
| Independent checks per visit | 106 signals across browser, network, device, behavior | S1, S3, S6 |
| Reported overall accuracy | 99% from AI weighing complete pattern | S1, S3, S6 |
| Customer refund success rate | 83% of customers get a refund | S2 |
| Average ad spend recovered | From Google and Meta billing disputes back to 2017 | S2 |
| Refund approval rate | Approved rate across client claims submitted to ad platforms | S2 |
| Setup time | About one minute to add to website | S2 |
| Bot click budget impact | Up to 20% of Google and Meta ad budget | S2 |
Limitations of each metric
- Detection rate alone ignores false positives — blocking real users hurts revenue and trust.
- Accuracy alone masks poor bot catch-rate when bots are a small fraction of traffic.
- Precision (of flagged visits, how many are truly bots) matters for refund evidence quality; BotRefund's cross-checked context step aims to keep precision high.
- No single metric captures the full picture; ask for detection rate, precision, and false-positive rate on traffic similar to yours.
Terminology cheat sheet
- True Positive — Bot correctly flagged as bot.
- False Negative — Bot missed (classified human).
- False Positive — Human wrongly flagged as bot.
- True Negative — Human correctly passed.
- Recall = Detection Rate = TP / (TP + FN).
- Precision = TP / (TP + FP).
- Accuracy = (TP + TN) / Total.
- F1 Score — Harmonic mean of precision and recall; useful single number when you need balance.
Decision framework for choosing a bot detector
- Define your goal: refund recovery, analytics purity, or both.
- Request detection rate, precision, and false-positive rate on a sample of your traffic.
- Verify evidence format: video proof, signal logs, and API exports that Google/Meta accept.
- Check integration time and ongoing maintenance (BotRefund cites ~1 minute setup) [S2].
- Run a free audit first; BotRefund offers a live bot audit on a demo call [S2].
Practical scenarios
- High-value PPC campaigns — Prioritize detection rate and evidence quality; accept slightly more false positives if they are reviewable.
- E-commerce checkout — Prioritize precision and low false positives; blocking a real buyer costs more than a few bot clicks.
- Lead-gen forms — Balance both; use honeypot traps and behavioral signals (ghost clicks, linear mouse movements) that BotRefund lists as independent checks [S2][S4][S5][S7][S8].
Frequently asked questions
Can a 99% accuracy claim be misleading?
Yes. If bots are 1% of traffic, a detector that labels everyone human achieves 99% accuracy and 0% detection rate. Always ask for detection rate on your traffic composition.
What detection rate should I expect for sophisticated bots?
Public benchmarks vary widely. BotRefund's 106-signal approach targets advanced bots that spoof fingerprints, rotate proxies, and mimic human timing. Ask vendors for results against headless browsers, residential proxy networks, and CAPTCHA-solving services.
How does false-positive rate affect ad refunds?
High false positives weaken your evidence pack. Google and Meta reviewers look for clean separation. BotRefund's cross-checked context step (independent evidence → cross-check → AI prediction) is designed to keep false positives low while maintaining detection rate [S1][S3][S6].
What evidence do Google and Meta require?
Timestamped click data, IP and fingerprint logs, behavioral video replay, and a clear narrative linking each signal to bot behavior. BotRefund's platform exports this package for dispute submission [S2].
How often should I re-audit detection performance?
Quarterly, or after major traffic source changes. Bot operators adapt; a detector that caught 90% last quarter may drop to 60% without model updates.
Does BotRefund guarantee a refund?
No vendor can guarantee platform approval. BotRefund reports an 83% customer refund success rate and a published refund approval rate across submitted claims [S2]. The free audit lets you assess evidence quality before committing.
Further reading and comparison sources
These external sources provide additional context for evaluating the topic. Their inclusion is not an endorsement.
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.