On October 1, Google updated its guidance on using AI-generated content. The headline line is this:
“It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing.”
That’s not news to anyone reading this. Of course you should check what the AI wrote.
The news is the next sentence:
“This review also applies to metadata like
Google stretched “content” to cover the parts of the page that almost nobody rereads before publishing. That’s where this gets useful.
What Changed, and What Didn’t
Google’s documentation changelog describes the update as bringing the page in line with the Search Quality Rater Guidelines and with what Google presents at its developer events. So this is guidance catching up to what Google has already been saying out loud, not a new rule appearing out of nowhere.
It’s also not a ranking change. The same page says plainly that rater ratings “don’t directly influence ranking.” Nobody is getting demoted because an alt attribute went unreviewed.
One line on that page is a different kind of thing, though. It still warns that using AI to generate many pages without adding value for users may violate Google’s spam policy on scaled content abuse. That’s a spam policy, and spam policies get enforced. The two ideas sit next to each other on purpose. Unreviewed AI output on a few pages is an accuracy problem. Unreviewed AI output across thousands of pages starts to look like the thing the spam policy describes.
Why Metadata
Google gives its own reason in the same sentence: these elements “can appear in Search results.”
Think about what a searcher actually sees before they decide whether to visit you. The title link. The snippet, which is often your meta description. Any rich result your structured data earned. Your images in image search, described by their alt text. That’s the whole first impression, and every piece of it is metadata.
So the part of your page doing the most work to decide whether anyone shows up is very often the part that got generated in bulk, dropped into a CMS field, and never looked at again. The body copy gets a read-through. The metadata gets a spreadsheet upload.
I covered these elements from the optimization side in the on-page elements note. This is the other side: not whether they’re optimized, but whether they’re true.
Where the Errors Hide
Titles
AI-written titles tend to overpromise. “The Complete Guide.” “Every Option Compared.” “Everything You Need to Know.” The page is a solid 900 words on one piece of the topic, and the title is writing checks it can’t cash.
Or the title drifts from the H1, because they were generated separately and nobody put them side by side.
Google rewrites plenty of titles anyway, but the ones it keeps are your promise to the searcher. Break that promise and you’ve built exactly the quick return to search websites are trying so hard to avoid.
Meta Descriptions
This is where the model fills in specifics that sound right. A starting price. A number of options. “Free shipping.” “Updated for 2026.” None of it was checked against the page, because the model never had to check. It just had to produce a plausible description.
When the snippet says “plans from $19/month” and the page says $29, you’ve lost the visitor’s trust before they’ve read a word. Sometimes before they’ve even clicked.
Structured Data
This is the most dangerous one, because nobody looks at JSON-LD unless something breaks.
The common failures: an aggregateRating with numbers the model made up. A price or availability that doesn’t match the page. A sameAs link pointing to the wrong entity, or to a Wikipedia page or Wikidata item that doesn’t exist at all, because the model produced a URL that looked right.
Google adds an instruction specifically for this one: structured data also has to comply with the general guidelines and the policies for each search feature, and you should validate it. That matters because a mismatch here isn’t only inaccurate. Markup describing things that aren’t on the visible page is a structured data policy problem, which is a different and more expensive kind of wrong.
Alt Text
The model describes what it guessed was in the image. The wrong color. The wrong product model. Two people when there are three. A “smiling team” in a photo of a server rack.
Alt text feeds image search, and it’s what a screen reader reads aloud. A visitor using one gets a confident description of a picture that isn’t there.
If You Sell Online
Ecommerce gets more than advice here. The same page points to Merchant Center’s policies, and they’re specific:
- AI-generated images must carry IPTC metadata marking them as AI-generated (DigitalSourceType set to TrainedAlgorithmicMedia).
- AI-generated product data, like titles and descriptions, must be specified separately and labeled as AI-generated.
If your product feed was written by a model, those are rules, not suggestions. Check that whatever tool generated your feed is actually doing both.
A Review Pass That Doesn’t Become a Project
The core rule: check metadata against the page, not against whether it sounds plausible.
AI output always sounds plausible. That’s the one thing it’s reliably good at. So reading a meta description and thinking “yeah, that sounds fine” catches nothing. You have to check it against something.
Four checks cover most of it:
- Every number or specific claim in the meta description appears on the page. If the snippet mentions a price, a count, or a feature, find it on the page. If you can’t, cut it.
- Every structured data value matches something a visitor can see. Ratings, prices, availability, authors, dates. If it’s in the markup, it should be on the page.
- The title and the H1 make the same promise. Put them next to each other. If one promises more than the other, fix the bigger one.
- Spot-check alt text against the actual images. Open the image, read the alt text, and see if they describe the same thing.
For metadata generated at scale, don’t try to review every page. Review a sample from each template. Errors in bulk-generated metadata repeat by template, so one bad template means hundreds of bad pages. The good news is the same: one fix repairs all of them.
The Tools I Built for This
I’ve built two tools around exactly this review step, because it’s the one that gets skipped. Receipts traces claims and statistics back to their primary sources before they go live. The Schema Builder verifies that every linked entity actually exists before it goes into the markup, which is the sameAs problem above. Neither one replaces reading the page. They just make the review fast enough that it actually happens.
The Point
AI didn’t change what accurate means. It changed how much gets published without anyone reading it.
Google just wrote down the review step that was always part of the job, and pointed it at the first thing searchers see.
