Two sets of rules for one text
A product description in an online shop lives in two places. On the product page, Google Search and the shopper see it. In the product feed — the file a shop sends to Google Merchant Center for free listings and Shopping campaigns — the shopping system sees it.
| Product page in Search | Product feed in Merchant Center | |
|---|---|---|
| AI text allowed | yes | yes |
| Must be labelled | no, only a recommendation to disclose how content was made | yes, in a separate field — structured_description and structured_title |
| Manual fact-check | Google requires it for the text, title, meta description, structured data and image alt text | content rules apply as to any description |
| If you break the rules | risk under the scaled-content policy if many pages are generated without value | product disapproval or limited serving |
So the same text can appear on the site without a label but not in the feed. Most "generate your descriptions with AI" advice never mentions the second column.
The product page in Search
Google's guidance on AI-generated content doesn't ban generated text. It asks for two things.
A manual check before publishing. Google explains that generative models don't retrieve facts but predict a likely sequence of words, so they can be wrong. All AI content should be checked, including title, meta description, structured data and image alt text. For an online shop that means: if the model wrote the description, the page title and the description in the Product markup, all three get checked.
Value for the shopper. Google's spam policies give generating many pages with AI without adding value as an example of scaled content abuse. A thousand descriptions that retell the same six specs in generic words is exactly that. A description with dimensions, material, age range, limitations and the answer to a question shoppers ask in chat is not.
What a model most often breaks in a product description showed up in our own test of three AI assistants: one model called a water-based lacquer "hypoallergenic", although the data said nothing of the kind. For a shop that phrase isn't style; it's a promise a customer can hold you to.
The feed: six ways to fill it and what happens
The Merchant Center page on AI-generated content says all descriptions created with generative AI must be submitted in the structured_description attribute. It has two sub-attributes: digital_source_type set to trained_algorithmic_media, and content with the text itself, up to 5,000 characters. Titles work the same way through structured_title, up to 150 characters.
One detail is easy to miss and sits on the description attribute page: if the feed has both structured_description and description, Google uses only description. Here's how that plays out across six typical variants.
Based on the Merchant Center help pages (description, structured_description and the AI-generated content policy). Results in rows 2 and 5 follow the help page's general rule: not meeting the minimum requirements can lead to disapproval or limited serving.
Row 4 is the sneaky one. The shop generates new descriptions, adds structured_description to the feed by the book, but leaves the old description in place. Merchant Center shows no errors, and shoppers keep seeing the old descriptions. If the point is to show the AI texts, remove description for those products.
Row 6 applies to text feeds (TSV, Google Sheets). There the value is written as trained_algorithmic_media:"description text": the colon separates the source type from the text. If the description has a colon of its own — "Dimensions: 120×45×60 cm" — wrap the whole text in double quotes and double any quotes inside it. Otherwise Google sees extra sub-attributes and disapproves the product. XML feeds don't have this problem, because sub-attributes are separate tags.
One more limit: Schema.org has no structured_description. Product markup on the page still uses the ordinary description, and the AI label lives only in the feed. Before generating, check whether your platform or feed module can export structured_description. If it can't, AI descriptions can't go into the feed — keep the human-written ones there.
What a feed description can't say
The content rules are the same for human and AI descriptions. A model told to "write a selling description" will break them almost every time, because to a model selling copy means price, delivery, the shop's name and capital letters. According to the Merchant Center help page, a description must not include:
- price, discounts, sale dates, shipping terms or the shop's name — they have their own attributes;
- comparisons with other products ("better than X") or descriptions of accessories and similar items;
- links to the shop or other sites;
- capital letters for emphasis, symbols or gimmicks;
- category paths like "Toys > Dolls" — use
google_product_categoryandproduct_type; - anything inconsistent with the product title in the feed or the landing page.
Google recommends putting the most important details in the first 160–500 characters, because shoppers see the rest only after clicking.
Process and prompt
A workflow for a shop that generates descriptions in batches:
- Input data from the product record only. Specs from your database or the manufacturer, not "write a description for a children's slide". The less the model gets from the input, the more it invents.
- Generation with restrictions. The prompt is below.
- A human fact-check. Every property in the description must be in the specs. Anything that isn't ("safe", "hypoallergenic", "strongest") gets cut.
- The site. Description, title, meta description and alt text — after the same check.
- The feed. AI description into
structured_descriptionwithtrained_algorithmic_media,descriptionremoved for that product. AI title the same way intostructured_title. - A spot check in Merchant Center a few days later: product diagnostics and a few listings.
A prompt that follows the Merchant Center requirements (substitute your own specs):
Write a product description for an online shop in Ukrainian, up to 600 characters. Use only the specifications listed below and do not add properties that aren't in the list: safety, hypoallergenic claims, certifications, warranties. Put the most important details in the first two sentences. Don't mention price, discounts, delivery, the shop's name, other products or comparisons with them. No capital letters for emphasis, no emoji, no links. Specifications: …
Restrictions in the prompt reduce the risk but don't remove it: the model predicts words rather than checking them against the specs. That's why step 3 isn't a formality.
The wider context of Google's October changes on AI text is in AI content and Google from October 2026: what actually changed.
Our position
For a shop with hundreds of products, AI descriptions pay off, but only as a draft checked by a person with the specs to hand. The costly mistakes here aren't stylistic: an invented property becomes a promise to the customer, and a badly filled feed quietly switches off the new texts or puts the whole catalogue's serving at risk. If you can't export structured_description, generate descriptions for the site only and keep human-written ones in the feed.
The position doesn't hold if you don't send products to Merchant Center and don't plan to. Then only the Search rules apply: a manual fact-check and value for the shopper, with no separate fields or labels.
Checked on 7 October 2026 against the Google Merchant Center help pages (AI-generated content policy; the description and title attributes and their structured versions) and Google Search Central (guidance on generative AI content, spam policies).