Which SEO Changes Should Always Require Human Review
Tier by blast radius times reversibility. Always-human: robots.txt and crawl directives, sitewide canonical or hreflang logic, redirects and URL structure, noindex anywhere, structured-data types with policy risk, and anything touching more than a defined URL-count threshold in one change. Reviewable-in-batch: per-page titles, descriptions, internal links on body content. The line isn't how smart the automation is — it's how expensive the mistake class is and how slowly search engines forgive it.
The question isn't whether AI can write a canonical tag — it obviously can. It's that some mistake classes cost a day and some cost a quarter, and the review budget should be spent where the quarter-sized mistakes live.
The always-human tier, and why each earns it
- robots.txt and crawl directives — one line can deindex a site section, and recovery runs on crawler time, not yours
- Redirects and URL changes — they cascade (old links, sitemaps, canonicals) and layered redirect mistakes compound
- noindex, anywhere — the single highest-consequence tag in the vocabulary; removal is easy, re-inclusion is a petition to a crawler
- Sitewide canonical/hreflang logic — errors here are pattern errors by definition, multiplied by the whole site
- Any change crossing your URL-count threshold — scale converts a small mistake class into a large one; pick a number (50 is common) and enforce it mechanically
Worked example
The title-and-canonical edit again: the title half is batch-reviewable (single page, instantly revertible, bounded downside). The canonical half sits in the always-human tier even though it's one page — because canonicals participate in sitewide logic, and one 'fixed' canonical that's wrong teaches the reviewer nothing about the other 340. Same PR, two tiers, and the process should be able to say so: approve the title in batch, escalate the canonical to a named owner.
A failure worth checking
Tier drift: automation earns trust on titles for three quiet months, and someone proposes 'graduating' redirects to the batch tier. The tiering was never about trust in the tool — it's about the price of the mistake class, which hasn't changed. Revisit thresholds annually against actual incident history; never promote a tier because the automation has been lucky.
Put this to work on your own website.
Fenn finds what your customers ask, drafts the articles and site fixes, and measures what ChatGPT, Claude, Gemini, Perplexity and Grok say about you — with every change waiting for your approval.