benchmarks · methodology

Content spam detection benchmark notes.

Updated May 12, 2026

Reported accuracy

99.4%

Internal benchmark, English-heavy content spam set.

Primary output

0-1

A calibrated spam probability, not a hard label.

Recommended shadow run

7 days

Validate thresholds against your own traffic before enforcing.

Common review band

0.60-0.90

Queue borderline submissions instead of auto-dropping them.

What the benchmark measures

The benchmark is built around content spam: comments, contact-form messages, short signup notes, generic promotional replies, link spam, and AI-assisted submissions that read fluently but add little value. It is not an email-spam benchmark, malware benchmark, payment-fraud benchmark, or account-takeover benchmark.

Accuracy is computed against a labeled set of ham and spam examples. Because most real systems care more about false positives than raw accuracy, production teams should inspect precision and recall at the thresholds they plan to enforce.

How to validate it on your traffic

Run Siftfy in shadow for at least one normal traffic week.
Log the returned probability, your current moderation outcome, and a human review label where possible.
Pick two thresholds: one for automatic drop and one for human review.
Manually inspect false positives before blocking anything automatically.
Recheck thresholds after a spam wave, launch, or major content-type change.

Recommended threshold bands

Probability	Suggested action	Why
0.90-1.00	Drop or quarantine	Usually obvious commercial, bot, or link spam.
0.60-0.89	Review queue	Borderline text where false positives are more likely.
0.00-0.59	Allow, with normal abuse monitoring	Most legitimate submissions should live here.

Known limits

English content is the strongest path today; non-English content can degrade.
Very short messages have less signal and may need stricter product-side rules.
The API scores content, not IP reputation, disposable email status, device fingerprint, or payment risk.
Spam tactics change, so thresholds should be reviewed regularly instead of frozen forever.

To test the API, read the predict endpoint docs, try the spam probability tester, or compare providers in the spam detection API guide.