Score WordPress Comments for Toxicity, Spam, and Relevance Before They Hit the Queue
WordPress comments still matter on posts that earn them. They also attract spam, drive-by insults, and off-topic walls of text that burn an hour of moderation for ten minutes of useful discussion.
Akismet-style filters catch a lot of spam. They do not score tone. They do not judge whether a comment is actually about the post. Generatify scores comments for toxicity, spam, and relevance before they hit your queue, then approve, hold, or bin based on thresholds you set, on the model you already connected.
This post is about comment moderation only. Not the admin assistant, not helpdesk, not smart compose, not draft posts, not product recommendations. For connecting ChatGPT, Claude, or Gemini to WordPress, see how to integrate ChatGPT, Claude, and Gemini. For why models fail without site context, see why AI models struggle without context.
The problem: the queue is a junk drawer
Default WordPress moderation is binary. Hold for review, or let it through. On a busy post that means you skim a pile of “nice post,” affiliate spam, and one actual question buried in the middle.
Human moderation does not scale when every reply needs a full read. Keyword blocklists miss clever spam and false-positive normal speech. A model that can score toxicity, spam, and relevance gives you a first pass that matches how you already think about comments: is this nasty, is this junk, is this on topic?
The catch is control. You want thresholds you own, not a black-box “AI decided” with no dials.
What Generatify actually ships for comment moderation
Stick to real features. No invented personality profiles or fictional community reputation graphs.
Scores for toxicity, spam, and relevance. Each comment gets scored on those three axes before it sits in your normal queue as an unexamined blob.
Approve, hold, or bin on your thresholds. You set the cutoffs. High spam goes to bin. Borderline toxicity holds for a human. Strong relevance with clean scores can approve. The policy is yours.
Runs on the model you connected. Generatify uses the AI model you already wired up. You are not stuck with a mystery moderation endpoint you cannot swap.
That is the whole surface for this post. Connect a model, set thresholds, let scores drive the first triage.
How to set comment scoring without nuking good discussion
Start stricter on spam, gentler on relevance, and review held items for a week before you widen auto-approve.
Connect a model you already trust. If Generatify is not talking to ChatGPT, Claude, or Gemini yet, fix that first using the integration guide. Moderation cannot score on a model that is not connected.
Open comment moderation settings in Generatify. Confirm toxicity, spam, and relevance scoring is on. Leave auto-bin aggressive only for spam until you have seen a few days of scores.
Set thresholds in writing. Example policy you can adjust: bin above a high spam score, hold above a moderate toxicity score, approve only when relevance is solid and the other scores are clean. Write your numbers down so you can compare next week.
Run a sample of recent comments through the same rules. Past spam, past useful replies, past edge cases. If good comments land in bin, loosen. If garbage still approves, tighten spam first.
Keep a human on held items daily for the first week. Scoring is triage, not a replacement for judgment on borderline tone. Held is the bucket that teaches you whether thresholds match your community.
Revisit after a viral post. Traffic spikes change the junk mix. Thresholds that worked on a quiet blog can be wrong on a post that hits the front page of somewhere.
Do not stack three moderation plugins fighting over the same comment. One scoring layer. One queue. Clear ownership.
What useful scoring looks like in practice
On a technical blog, relevance matters as much as spam. A polite comment that sells a random SaaS tool is still noise. On a news or opinion site, toxicity scoring protects readers without forcing you to read every insult twice. On a product or docs site, spam scoring clears affiliate paste while holding the “this steps fails on PHP 8.3” reports for you.
Approve, hold, and bin mapped to your thresholds mean the queue shrinks to decisions you still need to make. The model you connected means you can change providers later without rebuilding the policy from scratch.
Pitfalls that make AI moderation worse than manual
Auto-binning on relevance alone. Low relevance is often “short thanks” or a partial question. Hold those until you know your audience.
One ultra-strict toxicity cutoff that silences blunt but fair critique. Product feedback is not always soft. Tune with real held samples.
No connected model, then blaming Generatify. Scoring runs on the model you connected. Wire it up first.
Never reviewing held comments. Thresholds drift from reality if nobody looks. Ten minutes a day beats a poisoned comment section.
Expecting scoring to fix registration spam at the form layer. This is comment triage after submit, not a CAPTCHA replacement.
Ignoring context limits. Models do better when they know what the post is about. That is the same lesson as why models struggle without context.
Changing all three thresholds every hour. Change one dial, wait, compare the held pile. Thrashing teaches nothing.
How this fits DaftPlug Pro
Generatify is one of five Pro plugins on a single membership. One licence unlocks Progressify, Lightify, Generatify, Rankify, and Snapshotify. New Pro plugins join existing memberships with no new line item. See pricing.
Comment moderation that scores toxicity, spam, and relevance on your connected model, then approve, hold, or bin on your thresholds, is how Pro sites keep discussion usable without living in the queue all day.
Takeaway
A comment queue without scores is a junk drawer with a notification badge. Generatify scores WordPress comments for toxicity, spam, and relevance before they hit that queue, then approve, hold, or bin on thresholds you set, using the model you already connected. Keep a human on held items until the dials match your community. Then moderation stops being an unread pile. Start in Generatify, connect your model if needed, set conservative spam binning first, and review the held bucket for a week before you widen auto-approve.