Blog / AI & search /
A review of AI-generated recommendations after a core update
An AI tool has just given you a list of recommendations after a core update, and the list looks impressively specific. Nice formatting, confident tone, numbered priorities. So who checked it?
Before anyone rewrites the site, ask what evidence the tool actually used. A recommendation can sound plausible while mixing timing, correlation and generic SEO advice. I treat each item as a hypothesis. The update may be useful context, but it doesn’t diagnose a particular page, and it doesn’t prove a suggested change will help.
Check the input before the output
Record what the tool received: date ranges, pages, queries, crawl findings, prompts and any outside material. If it had no site data, it can’t make a site-specific diagnosis. If it summarised a report, compare its claims with the original.
Look for invented causes, missing caveats and recommendations that don’t follow from the evidence. A fluent explanation isn’t verification, especially when a chart has been turned into a confident story about an algorithm.
Validate against the site
For each suggestion, find the affected page group and look at real performance, content, technical state and the user’s task. Separate confirmed issues from possible explanations. A drop in clicks with stable impressions calls for a different question than pages that aren’t being crawled.
Check whether releases, seasonality, inventory or measurement changes overlap the period. Google advises waiting at least a week after a core update finishes before analysing Search Console performance. Use the official timeline and compare like with like.
Choose actions deliberately
Keep recommendations that point to a verifiable problem. Revise the ones that overstate their evidence. Reject the rest. Fix what hurts users even when the update isn’t the cause.
Write down the evidence and expected result before you change anything, then review later without promising recovery.
Keep the work reviewable
Pick one question the current data can answer, rather than listing every concern. A small scope is one a person can actually validate.
State the date, source and conditions for any Search Console observation. Without that, a one-off pattern gets read as a stable one.
Compare one ordinary page with one exception, such as a missing value or an outdated URL. If their behaviour differs, someone should be able to say why.
When two reports disagree, don’t merge them into one answer. Check definitions, timing and scope with the owners, and keep the difference visible until you know where it comes from.
Keep confirmed facts, likely explanations and proposed actions apart in the report. They sit side by side on the page, but they don’t carry the same weight.
Before applying a rule across the whole site, test it on one representative page and check the live result after publication. I’d rather extend slowly than undo quickly.
Know which fact is authoritative and which system merely displays it. A copied number or a generated suggestion should never quietly replace the verified source.
Review the path a visitor or editor actually follows, not just the setting where the team expects the behaviour to start. Another layer, such as a template or a redirect, may change the final result.
Log every change with a short reason and an owner, and keep the earlier state where rollback or comparison matters. The next reviewer then has context without needing a private chat.
A useful handover names the person who can answer a question about the domain, the product or the measurement, instead of pointing at a broad team.
If the evidence is incomplete, say what’s missing and choose a next check that reduces the uncertainty. Don’t fill the gap with a plausible assumption just to make the review look finished.
Revisit the same page after the change and confirm the original problem is actually gone. A successful deployment on the wrong layer fixes nothing.
Keep the maintenance note short but specific: the owner, the source and the event that should trigger another look. A real change, such as a new core update or a template release, is a better prompt than a calendar reminder.
When you share a finding, put its limitation next to the recommendation. A manager reading it should see what is known, what is still uncertain and why the next step is proportionate.
Keep a short list of assumptions you still need to confirm, and say who can settle each one. A convenient working note is not a verified source.
Related reading
- A review process for AI-generated recommendations
- The March 2026 core update: a measured post-rollout review
- How to audit AI-generated metadata across a catalogue
What would you need to see before approving the AI’s first recommendation?