Two documents from the same company appear to say opposite things, and between them sits every shop owner who has ever asked whether it is safe to let a model write four hundred product descriptions.
One is Google's published guidance, which states plainly that generative AI can be useful and that the problem is volume without value. The other is the instruction given to the people who rate search results, which tells them to award the lowest possible rating to a page whose main content is auto or AI generated. Resolving that apparent contradiction is the whole of this article, and the resolution is more permissive than most sellers assume.
- Google's own documentation does not treat AI involvement as a violation. The named offence is scaled content abuse, defined as generating many pages primarily to manipulate rankings, no matter how they were created.
- The quality rater guidelines do tell raters to apply the Lowest rating to AI generated main content, but the clause is conjunctive. It requires little effort and little originality and little added value together.
- Read side by side, the two documents agree. AI appears in the raters' list as one of several ways to produce effortless content, alongside copying, paraphrasing and reposting.
- For online sellers there is a specific and little known instruction. Google asks that AI generated images carry IPTC metadata and that AI generated product data be specified separately and labelled.
- The practical test is not whether a model wrote it. It is whether the page contains anything that could not have been written about a competitor's identical product.
- Spam actions are described as recoverable once the cause is fixed, which makes this a correctable mistake rather than a permanent one.
Does Google penalise AI written pages?
No. AI generated content is not penalised for its authorship, and the documentation is unusually direct about it. The distinction Google draws is between how content was made and what it is worth, and only the second one is a policy matter.
Google's guidance on AI generated content in Search describes the technology as particularly useful when researching a topic and when adding structure to original content. The warning attached is specific rather than general: using generative AI tools or similar tools to generate many pages without adding value for users may violate the spam policy on scaled content abuse.
Every load bearing word in that sentence is a qualifier. Many pages, not one page. Without adding value, not without a human typing. May violate, tied to a named policy rather than to a vibe. A shop that drafts one product description with a model and edits it has not come near the described behaviour.
The policy itself is worth reading in the original because its final clause is the one that settles the question. The spam policies page defines scaled content abuse as generating many pages for the primary purpose of manipulating search rankings rather than helping users, and describes it as creating large amounts of unoriginal content that provides little to no value to users, no matter how it is created.
No matter how it is created. A human writing four hundred thin descriptions by hand at two minutes each commits the same offence as a model producing them in an afternoon. The policy is indifferent to the authorship and interested only in the output, which is a more coherent position than it is usually given credit for.
Then why do the rater guidelines say the opposite?
They do not, though the sentence is easy to misread and it has been widely misread. The clause is a list of conditions that apply together, not a list of disqualifying properties any one of which is fatal.
The January 2025 revision to the guidelines, reported in detail by Search Engine Land, instructs raters to apply the Lowest rating where all or almost all of the main content on a page is copied, paraphrased, embedded, auto or AI generated, or reposted, and where it demonstrates little to no effort, little to no originality, and little to no added value for visitors.
Read it as two halves joined. The first half lists ways content can arrive without being authored: copying, paraphrasing, embedding, generating, reposting. The second half is the test. A page fails when it arrives by one of those routes and shows no effort and no originality and no added value. Strip out the second half and the sentence would condemn every quotation and every embedded video on the web, which is plainly not what any search engine wants.
The same guidelines, per that reporting, introduced a formal definition of generative AI for the first time and describe it as a helpful tool that can also be misused. That is not the language of a prohibition. It is the language of a factor, which is exactly how it functions in the rating.
So what actually gets a shop into trouble?
Volume without addition, and the shapes it takes in a catalogue are predictable enough to list. The table below sorts the practices a shop actually considers, rather than the abstract cases in the policy.
| The practice | Where it stands | Why |
|---|---|---|
| Four hundred descriptions generated from the manufacturer feed, unedited | Squarely inside scaled content abuse | Many pages, no value added over the source everyone else also has |
| The same drafts, each edited with your measurements, photos and returns notes | Fine | Value added that exists nowhere else, whatever wrote the first pass |
| Auto generated landing pages for every keyword and city combination | Squarely inside scaled content abuse | Pages made for rankings rather than for a reader, the textbook case |
| AI drafted alt text, titles and meta descriptions | Fine | Held to the same standard as hand written ones, no higher |
| Blog posts summarising other blog posts at volume | Inside the policy | Unoriginal content at scale, the repackaging case in the guidelines |
The first two rows are the same tool and the same starting draft, separated only by whether a person added something. That is the entire difference between a catalogue that ranks and one that gets a manual action, and it is a difference of minutes per product rather than of software.
The reassuring part is that this is a recoverable position. A commerce focused explanation of the same updates, Practical Ecommerce's walkthrough of Google's spam updates, notes that traffic declines from spam updates are usually recoverable once the causes are fixed. That is a materially different prospect from a permanent mark, and it means the correct response to a drop is repair rather than migration.
What does Google specifically ask of online sellers?
Two things almost nobody has implemented, both of them labelling requirements, and both of them sitting in the same guidance page as the general advice. This is the most concrete and least known part of the subject.
The first concerns images. Google asks that AI generated images carry IPTC metadata identifying them as such. IPTC is the standard photographic metadata block, the same place a photographer's credit and caption live, and most image generators can write it or it can be added afterwards in any competent image tool.
The second concerns your product data. Google asks that AI generated product data be specified separately and labelled as AI generated, rather than mixed indistinguishably into the fields you supply. In practice that means the attribute you generated is marked as generated, which is a feed decision rather than a copywriting one.
The instinct is to hide the labelling, on the theory that admitting AI involvement invites a penalty. The documentation points the other way. It says sharing how a piece of content was created can give readers more context, and the labelling asks are framed as transparency rather than as confession. Given that the policy does not penalise AI use in the first place, there is nothing being confessed to.
If you are generating product photography rather than prose, the labelling question arrives alongside a set of honesty questions about what a rendered image may claim, which we worked through in our piece on what a generated product image can honestly do.
What about the pages nobody reads?
Category pages, filter combinations and the long tail of variants are where most catalogues quietly breach this policy, and they are the pages an owner is least likely to look at. The product pages usually get some attention. The three hundred generated pages sitting behind them rarely get any.
The mechanism is familiar to anyone who has run a store. A platform offers to create a page for every combination of attribute and category, because more indexed pages sounded like more traffic, and each one is populated by a template with the attribute names substituted in. Blue cotton shirts in medium. Cotton shirts under thirty pounds. Shirts for men in blue. Nothing on those pages was written for a reader, and the policy's phrase for pages generated primarily to manipulate rankings describes them precisely, whether or not a model was involved.
The fix is subtraction rather than writing, which makes it faster than owners expect. Decide which facet pages a customer would plausibly search for, keep those, and stop the rest from being indexed. A handful of combinations genuinely get searched and deserve real copy. The remaining hundreds are navigation, and navigation does not need to be in the index to work.
Variant pages carry the same risk in a slightly different shape. If every colour of the same item has its own page and the text differs by one word, you have near duplicates at scale, and generating those descriptions with a model makes the duplication more uniform rather than less. Either consolidate them onto one page with a colour selector, or give each variant something genuinely its own, which usually means its own photographs and its own note about how that colour behaves.
Location pages are the third case and the one most likely to attract a manual action, because the pattern is so recognisable. One template, a town name substituted, repeated across every place you might deliver to. If you have a real presence in a town, the page can carry real information about it. If you do not, the honest version of that page is a single delivery area page listing where you ship, which is also the version that survives every future update to this policy.
How do you tell if a page has added enough?
Apply one test: could this sentence appear, unchanged, on a competitor's page for their version of the same product? If it could, it has added nothing, and if most of the page could, you have the problem the policy describes.
Run it on a real description and the result is uncomfortable. Phrases about premium materials, careful craftsmanship and everyday versatility survive on any page for any product, which means they are not information. What does not survive the test is specific: the actual measured weight, how it fits somebody who is between sizes, which of the three colours photographs differently from the swatch, why one in twenty comes back.
All of that lives in your business rather than in a model. You have the returns data, the customer questions, the photographs, and the item in your hand. A model has the manufacturer's copy that every other retailer also received, which is precisely why an unedited generation is unoriginal by construction rather than by accident.
This is the same conclusion we reached from the craft side in writing product descriptions with AI without sounding like AI, and from the effort side in how much editing an AI draft needs before it ships. Three different routes to one answer: the draft is free and the specifics are the product.
Should you tell readers a model helped?
On the pages where it would genuinely inform them, yes, and on a product description, almost never. Google's position is permissive rather than prescriptive here, which leaves the decision to you and to what the reader would reasonably want to know.
The guidance says that sharing information about how a piece of content was created can help give readers more context. Note the framing: context, not compliance. There is no required disclosure statement, no prescribed wording, and no ranking consequence attached either way.
The useful way to decide is to ask whether the method changes how the content should be read. On an article that reviews or compares things, it does, because a reader is entitled to know whether a human handled the products. On a technical explainer, it can, because the reader may want to weigh the accuracy differently. On a product description listing your own measurements, it does not, because the specifics came from you and the sentence structure came from a tool, which is no more interesting than which word processor you used.
Where disclosure does become obligatory is separate from search entirely, and it comes from consumer and marketing rules rather than from Google. Anything that reads as a customer's own words, a testimonial or a review, carries requirements no search policy touches. So does a claim about what your AI features do. We covered that second one in what you may claim about AI in your own marketing, and the boundaries there are considerably tighter than anything in this article.
Does any of this change what you should build?
Only in one respect, and it is about how quickly you can edit rather than about what you generate. The gap between the first and second rows of that table is a person adding specifics, and how expensive that is depends entirely on your tooling.
A shop where changing a product description means waiting on a queue will publish the unedited draft, because the alternative is friction at every one of four hundred items. A shop where the copy is directly editable will accumulate the specifics over time, one product at a time, as questions come in and returns get logged. The policy does not care which situation you are in, and the outcome differs completely. Merchants who run a store whose pages they can edit directly tend to end up on the right side of that line for reasons that have nothing to do with intent.
The other thing worth building is the habit of recording what you learn about a product. A note that says three customers asked whether it fits a standard shelf is a sentence no competitor has, and it costs nothing to capture at the moment somebody asks.
What to do on Monday
Sample ten of your product pages and run the competitor test on each, sentence by sentence. Count the sentences that survive. That number, not any detector and not any published rate, tells you where you stand against the policy that actually exists.
If the count is near zero across the sample, you have the scaled content problem regardless of what wrote the pages, and the fix is to add specifics to your best sellers first rather than to rewrite everything. If most sentences survive, the presence of AI in your workflow is not a risk and you can stop worrying about it, which is worth knowing given how much anxiety the subject generates.
Then set the two labels. IPTC metadata on generated images, a separate marked field for generated product data. Both are small pieces of housekeeping, both are explicitly requested, and neither costs you anything in ranking, because the thing everybody fears being punished for was never the offence.