Feature deep-dive
The reflection critic
Last updated
Most AI writers ship one-shot. Mentionwell doesn't. Every draft hits a reflection critic — a separate model from the writer — that grades the article on four axes and triggers an auto-rewrite when the score drops. The critic is the editorial layer that makes the difference between 'AI slop' and shippable content.
What the critic grades
- Accuracy. Are the claims grounded in the research sources? Are the statistics defensible?
- AEO readiness. Does the article have a TL;DR, question-led headings, FAQ, citation list? Will an answer engine lift the lead paragraph cleanly?
- Factual grounding. Are numbers, names, and dates from the citations rather than the model’s training data?
- Brand fit. Does the tone match the site’s brand profile? Are competitor blocklist entries respected?
The 0.8 threshold
The critic returns a confidence score 0.0–1.0 across all four axes. Below 0.8, the article enters the rewrite stage with the critic’s comments as input. Above 0.8, it proceeds to metadata + FAQ + images. We picked 0.8 after measuring against editorial accept/reject decisions on ~5,000 articles internally.
Why “writer” and “critic” are different models
A model grading its own output is biased toward accepting its own choices. Mentionwell routes the writer and the critic to different frontier providers — one drafts, another grades. This is the same pattern Anthropic documented in their constitutional-AI work and that we’ve seen empirically: cross-model critique catches mistakes the original model missed.