Google Says Review AI-Made Metadata Before It Ships. At Store Scale, Who Does That?
Your catalog has four thousand products. An AI tool offers to rewrite every title tag, meta description, image alt text and schema block overnight, and the preview looks good. The question is not whether it can. The question is who checks the output before Google, and your customers, see it.
Google has now said where it stands. On October 1, 2026, its guidance on using generative AI content was updated, and the answer is plain: review it before it ships. For a US store with a real catalog, that sentence hides a hard operational problem.
What Google's guidance actually says
Google's generative AI content guide says AI can help with researching a topic and adding structure to original work. It also warns that using AI to generate many pages without adding value for users may violate the spam policy on scaled content abuse.
On review, the guide is specific. It says to manually fact-check and review AI-generated content before publishing, and it extends that review to the metadata that can appear in Search: title elements, meta descriptions, structured data and image alt text. It adds that structured data should be validated.
For merchants, the guide has two further points. AI-generated images should carry IPTC DigitalSourceType metadata with the value TrainedAlgorithmicMedia, and AI-generated product data, such as titles and descriptions, should be specified separately and labeled as AI-generated in Merchant Center.
The guide also now points to sections 4.6.5 and 4.6.6 of the Search Quality Rater Guidelines, covering scaled content abuse and main content created with little effort, originality or added value. Google is explicit that those guidelines are not a guide to ranking first, and that rater scores do not directly influence ranking. Treat them as a description of what Google considers low quality, not as a lever.
Quick answer: how should a store review AI-generated SEO changes?
Quick answer: Review AI-made titles, descriptions, alt text and structured data before they publish, as Google's October 1, 2026 guidance advises. Check each change against your catalog data, validate the markup, label AI-generated product data in Merchant Center, confirm each page keeps unique value, and keep a diff with a rollback path.
Why the review step breaks at catalog scale
Human review is the default safety net for AI systems. A Berkeley-led study of AI agents in production, Measuring Agents in Production, found that 74% of the deployed agents it analyzed rely on human-in-the-loop evaluation, and that 68% execute at most ten steps before a human intervenes. Its authors name reliability as the primary bottleneck. That study covers agents in general, not SEO agents, but the pattern carries over.
The trouble with a human reviewer is arithmetic. Four thousand products, four fields each, is sixteen thousand changes. Nobody reads sixteen thousand diffs carefully. The review either becomes a bottleneck that stops the work, or it becomes a rubber stamp that approves it. Neither is what Google is asking for.
So the review has to be designed, not assumed. The agent doing the work should not be the only thing checking the work, and the checking should be mechanical wherever it can be.
Five checks a verification step needs
At MnT Future we build agents with verification as a separate role from action, and this is the checklist we think any AI-made SEO change should pass before it publishes.
- Truth against the source. Every claim in a title or description must match the catalog record: material, size, compatibility, price, availability. The catalog is the source of truth, not the model.
- Valid, honest markup. Structured data must validate, and it must describe what is visible on the page.
- Labeling. AI-generated product data and images carry the labels Google asks for in Merchant Center.
- Added value. The change makes the page more useful. A template that stamps the same sentence onto four thousand pages is the pattern scaled content abuse describes.
- A diff and a way back. Every change is logged with before and after, and can be reverted in one step.
None of these needs a person to read prose for each product. Most of the five can be checked by code. The person reviews exceptions and samples, which is a job one person can actually do.
What a measured result looks like
This is the design behind Searchlight, the autonomous SEO and AEO agent MnT Future built. It watches a site, fixes what it finds, and verifies its own work before anything ships. It runs five agent roles, and in its first AI pass it reduced open issues by 70%.
Read that number carefully, because we do. It is a count of open issues on the agent's list. It is not a ranking, a traffic or a revenue claim, and it says nothing by itself about the remaining 30%. A fix counter is useful, but only alongside what stayed open, what was reopened, and what a reviewer rejected.
That is the standard we would hold a vendor to. Before giving any tool write access to your store, ask for four numbers: issues found, issues closed, issues reopened after verification, and changes rejected by a human. A tool that cannot report all four is reporting activity, not quality.
What to do this month
Before the holiday freeze, run three steps:
- Inventory every AI-assisted field on your site: titles, descriptions, alt text, schema, feeds. Note which are generated and which are written.
- Confirm that AI-generated product data is labeled the way Merchant Center expects.
- Put a verification step between any AI tool and your live catalog, with a log and a rollback.
If you want a second pair of eyes, MnT Future offers a free agent-readiness audit that includes how AI-generated product data flows into your feeds and structured data. You can also book a free strategy session to talk through where a verifier belongs in your stack.
Sources: Google Search Central, "Using generative AI content" (updated October 1, 2026) and Search documentation updates; Pan et al., "Measuring Agents in Production" (UC Berkeley, Stanford, IBM Research and others; ICML 2026).