Blog / AI & search /
Keeping AI-generated recommendations out of client reports until checked
The client report looks sharp. Page three says, confidently, “Fix crawl budget.” Page five says “Rewrite thin content.” The client nods along.
There’s just one problem. Nobody checked whether either issue actually exists on this site.
AI-generated recommendations are handy for exploring possibilities. They shouldn’t become client advice until someone has verified the evidence and the context. Reports shape decisions, budgets and trust. A wrong recommendation sends the client off spending time on work that doesn’t fix their problem. So put a review gate between generated analysis and anything presented as a finding or next step.
Label generated output as a draft
Keep AI-produced notes in a working area until they’re reviewed. Don’t paste them straight into a client deck, ticket or email. Mark the draft as generated or unverified, so a colleague doesn’t mistake fluent prose for an approved conclusion. If the analysis touches client data, only use a tool and workflow approved for that information.
Keep the source material and prompt where it makes sense. The reviewer can then see what the tool was asked and what data it saw. Don’t put private data in a prompt just because it’s easier. Remove identifiers, or use aggregated figures when that’s enough for the task.
Trace every recommendation to evidence
For each recommendation, pin down the observed condition, the source and the affected pages. “Many pages are not indexed” needs a defined set of URLs, a date and a way to reproduce it. “Add schema” needs to say which content is eligible, what structured data is missing or invalid, and whether the information is visible on the page.
Keep observation and interpretation apart. “The crawl found these URLs returning 404” is not the same as “these errors are harming performance”. The second needs context. Are the URLs important? Linked? Recently removed? Supposed to be gone? The reviewer should be able to inspect the original evidence, not just trust a summary.
Ask whether the recommendation fits this particular site. Generic best practice can be irrelevant, or even harmful, on a certain CMS, mid-migration or for a particular business model. Telling a client to create more content is no help if their existing pages already overlap, or they have no capacity to maintain new material.
Check technical details and dates
When a recommendation depends on a feature or setting, check it against current official documentation. Interfaces and policies change. Confirm the exact URL, environment and configuration before advising a client to touch a robots rule, canonical, redirect or security setting. If you can’t confirm how it behaves, say so and ask the right specialist.
Be careful with anything time-sensitive. Search features, Google updates, analytics names and tool capabilities all change. Don’t put a “current” claim in a report based on an old model answer or an unverified timeline. Cite a primary source, or state the date and limits of the observation. And never present correlation as cause.
Rate confidence and consequence
Use a simple category for each item: verified and ready, needs more evidence, specialist review, or do not recommend. Confidence should reflect the quality of the evidence, not how sure the generated wording sounds. A low-risk idea can go in as a hypothesis to test. A change to live URLs or security settings needs a stronger review and a rollback plan.
Think about what happens if you’re wrong. Could the change remove content, expose private information, break a form or change how customers buy? The bigger the impact, the more review and approval it needs. The human reviewer has to have the time, context and authority to reject the output. A rubber stamp isn’t a safeguard.
Write client recommendations clearly
A client-facing recommendation should explain the issue, the evidence, the expected benefit, the trade-offs, the effort and the next action. Skip jargon where plain words will do. Separate what’s known from what should be tested. If a result isn’t guaranteed, say what can be observed after the change and when you’ll review it.
Don’t overstate results or promise a fix will lift rankings. A technical correction might restore intended access or remove a known error, but search performance depends on plenty of other things. Pick a sensible measure for the work, and be honest about what it can’t prove. When I write an SEO audit, every recommendation has to point at something I can show the client on their own site. If I can’t, it doesn’t go in.
Keep an audit trail
Record who reviewed each recommendation, what evidence they checked, what they changed and what the client approved. If a suggestion is rejected, note why. Patterns in rejected output can improve the prompt, or show that automation just doesn’t suit that task.
If the client asks why a recommendation changed between drafts, the audit trail makes the answer easy. You can point to the data, a correction, a policy update or an approval decision. That honesty beats hiding the role of AI, or pretending the first version was a fully verified analysis.
Keep the review checklist short enough to use: source, date, scope, applicability, risk, evidence and owner. Add more only when the work demands it. The aim is to make careful review repeatable, not to create a form everyone fills in without reading.
Related reading
- Writing an SEO audit report people actually read
- How I report SEO results to clients
- AI and SEO: AI-generated images and product accuracy
AI can speed up brainstorming and organising. It can’t own the client relationship, or the consequences of a bad recommendation. I’d rather send a shorter report I can defend line by line than a longer one padded with generated advice. So take one proposed action, trace it from source to final report, and cut anything you can’t support. Which recommendation in your next report needs a human to check the evidence first?