Shopify AI for product content that waits for human review.

Writing a product description is slow, repetitive work, so it is no surprise that merchants reach for AI to draft them. Shopify built a text generator into its admin for exactly this. The drafting is the easy part. What the public guidance from search engines, shopping feeds, advertising regulators and consumer law has in common is that the words you publish are yours, whoever or whatever typed them. This post collects what those sources say about product content and what a review step should check.

What the built in generator does

Shopify's help centre describes Shopify Magic as a set of tools to help speed up your writing process, and says suggestions are generated from the information you provide. Product descriptions can be generated in the product details section of the admin. The documentation says you can give feedback on suggestions to flag inappropriate or harmful content, and that Shopify Magic tools are free regardless of your subscription plan.

Nothing in that description promises the output is accurate for your product. The tool works from what you give it. If your product data is thin, the draft will fill the gap with plausible language, and a human reviewer has to check that language against the product.

Google: review before you publish

Search is the first place poor copy shows up, so start with what Google says about AI written pages.

Google Search Central's guidance on generative AI content is direct. Generative AI can be useful for researching a topic and for adding structure to original content. But using generative AI tools to generate many pages without adding value may violate Google's spam policy on scaled content abuse. The page says it is critical to manually factcheck and review all AI generated content for accuracy and trustworthiness before publishing. It explains why: generative models predict likely sequences of words and do not retrieve facts, so hallucinations are a real risk.

The advice applies to more than body text. Google says the review covers titles, meta descriptions, structured data and image alt text. It also suggests sharing information about how a piece of content was created, to give readers context, and notes that merchants who use AI for images or product data should label them: e commerce images should carry the IPTC digital source type value TrainedAlgorithmicMedia, and AI generated titles and descriptions in product data should be specified separately and labelled as such.

The spam policies define scaled content abuse as generating many pages primarily to manipulate search rankings and not to help users, however it is produced. The policies also name thin affiliate pages that copy product descriptions from merchants as low value. Its list of examples includes using generative AI tools to generate many pages without adding value for users, and scraping or combining content with automated transformations such as synonymising, where little value is provided to users. It also lists creating many pages whose content makes little sense to a reader but contains search keywords. The consequence it states is that sites violating the policies may rank lower in results or not appear at all, and that Google can apply manual actions that cost significant visibility. On thin affiliate pages, it says a good site adds value through things like additional pricing information, original reviews and rigorous testing, comparisons and ratings. The practical point for a store with a large catalogue: drafting every description in one pass and publishing without reading is exactly the pattern these policies are written around.

Your shopping feed has its own rules

If you send products to Google Merchant Center, its product data specification adds constraints that a generator does not know about. Titles can be up to 150 characters and descriptions up to 5,000. Descriptions should not include promotional text such as free shipping, should not use all capital letters and should avoid gimmicky foreign characters. The data must match what the landing page shows, and wrong data can cause product disapprovals.

The specification is blunt about scope. The description should include only information about the product, and it should not carry links to stores, sales information, competitor details or other products. It asks for formatting such as line breaks and lists where that helps, and for titles on variants to include distinguishing features such as colour or size. These are the kind of mechanical rules that are easy to miss in a generated draft and easy to catch in a review.

A fluent description that promises a feature the product lacks fails twice. It misleads the shopper, and it can put the listing at risk of being disapproved. Both are checkable by a person in seconds when the draft sits next to the product record.

There is a quieter issue for stores with many near identical products. Google's guidance on duplicate URLs says it groups duplicate pages and picks one as the canonical version. It does not frame this as a penalty. It says consolidation combines signals and saves crawl time. Generating many variations that say the same thing in slightly different words gives search engines more pages to group and no more reasons to show any of them.

What shoppers need from a description

Baymard Institute, which runs large scale usability research on ecommerce, found that the top 60 sites mostly get this right: 90 per cent maintain high quality, consistent product descriptions, and 10 per cent do not give enough detail for what users need. Their research identifies three kinds of information shoppers look for: materials and ingredients, dimensions, and compatibility. For beauty products, they report that 50 per cent of shoppers need ingredient information.

The costs of missing details are practical. Users abandon products quickly when the detail they want is absent. Some make wrong assumptions, which leads to returns. A run of poor descriptions can make people leave a site, and some describe sites with thin descriptions as not legit. One shopper comparing chairs says of the one with more information that they are starting to discount the other.

Baymard's recommendations are specific enough to turn into a review checklist: give dimensions with labelled units and alternates in metric and imperial, state compatibility explicitly for products that depend on others, support ambiguous images with text, say which accessories are included and which are optional, and use subheadings to organise detail by feature. An AI draft often reads well while leaving all of these out, or inventing them. One shopper in the research expected the ingredients to be under the description, and did not find them there. A reviewer who knows the product catches that. A reviewer who does not should not approve.

UK advertising and consumer law

In the UK, the Advertising Standards Authority applies the CAP Code, and rule 3.7 requires marketers to hold documentary evidence to prove claims that consumers are likely to regard as objective and that are capable of objective substantiation. The ASA's guidance says the evidence must exist before the claim is published, and that marketers who lack it will have the claim regarded as misleading. It judges how consumers are likely to read the claim, not what the advertiser meant. Examples of objective claims include statements about price, availability and product benefits.

The ASA also draws the line on exaggeration. Obvious puffery and subjective opinion generally does not need evidence, as long as it is not materially misleading. In its example, the line THE ORIGINAL AND BEST SINCE 2004 was treated as subjective opinion and not provable fact. For objective efficacy claims, testimonials are not enough, and for health and beauty products the guidance calls for human trials. A model writing a product description does not know which of its sentences count as claims needing evidence. It will happily write that a moisturiser is clinically proven. A person must decide whether that sentence is puffery or a claim you can back.

The law behind this changed in 2025. The Digital Markets, Competition and Consumers Act 2024, Part 4, section 226 defines a misleading action. It covers false information about a product, and also true information presented in a way likely to deceive the average consumer, even if the information is correct. The Competition and Markets Authority announced on 7 April 2025, the day after the regime began, that it can now decide for itself whether the law has been breached, and can fine up to 10 per cent of global turnover, without going to court. It named objectively false information among the practices in scope.

The same section covers what you leave out. Section 227 treats a practice as unfair when it omits material information that consumers need for an informed decision. The Act's test is the average consumer, described as reasonably well informed, reasonably observant and reasonably circumspect, and the question is whether the practice would lead that person to make a transactional decision they would not otherwise have made. For product copy, the omission matters as much as the claim: a description that reads beautifully and leaves out that the product is not compatible with the model the shopper owns can fall on the wrong side of this test as easily as one that overstates.

The CMA's announcement gives the wider picture of what the regime covers: aggressive sales tactics aimed at vulnerable consumers, hidden fees revealed late, objectively false information, unfair contract terms, fake reviews and drip pricing. Fines can reach 10 per cent of global turnover for breaches of consumer law, 5 per cent for breaching undertakings given to the CMA, and up to 1 per cent for failing to provide information or concealing evidence. Its chief executive, Sarah Cardell, said fair dealing businesses deserve to know their competitors are playing by the same rules. Most online shops will never be a target of that machinery, but the standard it enforces is the one a careful merchant already works to.

• 10%

  maximum fine as a share of global turnover for consumer law breaches

  GOV.UK, CMA, 2025

• 150

  maximum characters in a Google Merchant Center product title

  Google Merchant Center, 2026

• 90%

  of the top 60 ecommerce sites maintain good descriptions

  Baymard Institute, 2021

These sources do not say that AI written copy is unlawful or that any merchant will be fined for a loose description. The regime is new and the CMA's priorities are its own. The point is narrower: the standard of proof falls on the seller, the test is how an ordinary shopper reads the page, and a draft you did not read is a draft you cannot defend.

What a review step should check

Pulling the sources together gives a short list for whoever approves a description. First, check every factual statement against the product record: materials, sizes, compatibility, what is included. Second, find the claims that a regulator would call objective, such as performance, origin, benefit or comparison, and confirm you hold evidence. Third, strip promotional text that feed rules forbid. Fourth, review the title, meta description and alt text as well as the body. Fifth, look at the draft against its neighbours in the catalogue to see whether it adds anything.

Two things are absent from the list. Nothing in it asks the reviewer to judge style, and nothing asks them to guess whether text was written by a person or a model. Google's guidance is about accuracy, trustworthiness and value to the reader, and the ASA and the Act are about what a shopper is likely to understand. All of them care about the page that results. Who typed it matters only because a machine is more likely to produce a confident sentence that nobody checked.

Seeing the draft beside the thing it describes, with a clear choice to approve or decline, is enough. That is also why bulk publishing without review is the risky pattern: the checks above are per product.

Where BYOM fits

In the free BYOM Shopify app, the assistant you already use reads the catalogue live. It proposes the edit on a confirm card. You see what will change. You confirm it, or you leave it. A confirmed change is written to Shopify and recorded.