Measured on Real Spam: Zero False Positives, 95% Caught. Here Is the Data.

Every anti-spam tool claims it catches "99% of spam." Almost none of them tell you how often they block a real customer by mistake. For a store, that second number is the one that actually costs you money, because a blocked spammer costs nothing and a blocked buyer costs a sale and a refund request. So we built Spam Shield against real spam at real volume, and we measured both numbers. Here is the work.

The dataset was not synthetic

We did not generate fake spam in a lab. We used what live WordPress sites actually receive: the everyday flood of fake orders, bot registrations, contact-form pitches and comment spam that hits WordPress around the clock. The measurement below uses two labeled sets. The main corpus holds 738 submissions taken from one high-traffic store's contact form, each one labeled by hand as spam or genuine. It mixes real spam with real customer messages, real questions about orders, shipping and wholesale pricing, because mixing the two is the entire point. A filter that blocks everything scores a perfect catch rate and is useless. A second cross-site set of 224 comments and product reviews from other WordPress sites checks that the result is not tuned to a single store. Both sets were run through the shipped plugin, version 1.8.7, in July 2026, with the AI tiebreaker and the shared pattern network switched off so the numbers describe the local rules engine alone. The same harness runs before every release, and a drop in precision blocks it.

The numbers

On the 738-item corpus the engine flagged 620 spam submissions correctly, let 88 genuine messages through, missed 30 spam submissions, and blocked zero genuine messages. That is a precision of 100% and a recall of 95.4%. On the 224-item cross-site set: 192 caught, 7 genuine passed, 23 missed, 2 genuine held, for a precision of 99.0% and a recall of 89.3%. A submission counts as caught when the engine blocks it or holds it for review; held messages are never lost and can be released with one click.

And almost zero false positives

The number we are proudest of is the one most tools hide: Spam Shield does not block your real customers. In the main corpus every message it flagged as spam was genuinely spam. A labeled set is finite, and live stores do occasionally surface a genuine message the heuristics misread; that is why anything borderline is held for review rather than deleted, and why a held message can be re-delivered to the right inbox. A blocked spammer costs you nothing. A blocked buyer costs you the sale, the refund email, and the trust. Spam Shield is built so that almost never happens.

The single most telling signal

One pattern stands out across the labeled data. A link inside a contact-form message is one of the strongest single spam signals we measure: people asking about your product almost never paste a URL, and people trying to sell you SEO services almost always do. Spam Shield reads that context the way a human would, which is exactly what a plain keyword blocklist cannot do.

Zero false positives and 95.4% of spam caught on real, labeled data. Real customers left alone. That is the standard we hold the product to, and the data backs it.