An Approval Checklist for AI-Generated SEO Changes
Approve an AI-generated SEO change only when five things are true: you can read the exact diff (not a description of it), the blast radius is enumerated (which URLs, which tags, how many), the claim behind the change cites its evidence, a tested rollback exists, and the change class is one you've pre-approved for automation at all. Anything failing one check gets a human edit or a rejection — never a shrug-merge.
AI-generated SEO changes fail differently from human ones: not one bad edit, but one bad pattern applied confidently across 400 pages. The approval step is where pattern errors die cheap. This checklist is what 'approval' concretely means, beyond scrolling and clicking merge.
The five checks
- Diff, not description — you are reading the literal before/after, no summarization layer in between
- Blast radius enumerated — the exact URL list or count, per change type; 'various pages' is an automatic stop
- Evidence attached — the observation that motivated the change (a query, a duplicate-title report, a crawl finding), linked
- Rollback tested — not 'git revert exists' but a named, rehearsed path back, including cache and sitemap consequences
- Change class pre-approved — the type of edit is on your allowed-for-automation list in the first place
Worked example: a title and canonical edit
Proposal: retitle /pricing-guide from 'Pricing' to a descriptive 58-character title, and point its canonical to itself instead of the homepage (a real class of misconfiguration). The checklist run: diff shows the two literal tag changes — pass. Blast radius: one URL, two tags — pass. Evidence: duplicate-title report and the canonical pointing at '/' — attached, pass. Rollback: previous values stored, revert rehearsed in staging, cache purge steps written — pass. Change class: metadata edits are on the pre-approved list — pass. Approve, merge, monitor.
The same proposal with 'canonical fixes across 340 blog posts' fails check two until the 340 are listed, and probably fails check five — bulk canonical rewrites belong in the always-human tier.
A failure worth checking
Approval fatigue is the real adversary: fifty small diffs a week and the reviewer starts pattern-matching titles instead of reading tags. Two defenses: batch low-risk classes into one weekly review with a summary diff, and keep high-risk classes rare enough that their review stays genuinely attentive. An approval process everyone skims is a merge button with paperwork.
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.