AI product description generators: what they get wrong

AI product description generators are good at one thing: turning a few facts into a fluent paragraph in seconds. What they get wrong follows from that. Given thin input, they fill the gaps with details that sound right and may not be. Given the supplier’s text, they produce something close to what every other shop selling the same product has. And they tend to write about how a product makes you feel rather than the size, material and fit shoppers actually need.

None of that means don’t use them. It means using them the way you’d use a junior copywriter: give them the facts, tell them not to add any, and check the work. This sets out the common failures, the rules that apply in the UK and on Google, and a workflow that makes AI descriptions safe at catalogue scale.

What the generators do

There are two kinds: the ones built into your platform, and separate tools and apps. They work the same way underneath.

Built into the platform

Shopify Magic is the obvious example. According to Shopify’s help page, you click the generate icon in the description box, type a prompt with the product’s features, keywords and the tone you want, and it writes the description directly into the field. Shopify suggests tones such as expert, playful, sophisticated, persuasive or supportive, and tells you to read everything it produces closely before saving.

Separate tools and apps

Standalone generators and catalogue apps do the same job in bulk, often for thousands of products at once. The speed is the appeal and the risk: a mistake in the prompt or the input is repeated across the whole catalogue before anyone reads it.

What they’re good at

A consistent structure across a catalogue, a first draft in the right tone, rewriting a supplier’s awkward text into plain English, and turning a list of attributes into readable sentences. All of these start from facts the tool is given.

Only as good as what they’re given

A generator given a product name and a price has nothing to work from but guesswork. A generator given the product’s real attributes, dimensions, materials, care instructions, what’s in the box, can write something accurate. The input is most of the result.

What they get wrong

The failures are predictable, and each one has a cost.

Invented details

Ask for a description of a jacket with no material listed and a generator may call it breathable, waterproof or recycled, because jackets often are. If yours isn’t, that’s false information on a product page, and a return or complaint waiting to happen.

The same words as everyone else

Many shops sell the same products from the same supplier. Feed a generator the supplier’s description and you get a paraphrase of what dozens of other stores already have, which gives neither Google nor a shopper a reason to prefer your page.

Feelings instead of facts

Generators lean towards phrases like elevate your style and perfect for any occasion. Shoppers deciding between two products want the measurements, the fabric and the fit. So do AI shopping assistants: ChatGPT’s shopping research is judged on matching several requirements at once, and it can’t match a requirement the description doesn’t state.

The wrong voice

A generator’s default tone is generic and upbeat. For a brand with a distinctive voice, or a technical product sold to experts, that reads as off, and a catalogue written in it loses whatever made the brand recognisable.

The rules that apply

Three sets of rules matter for a UK store publishing AI-written descriptions.

UK consumer law

The Digital Markets, Competition and Consumers Act 2024 replaced the old unfair trading regulations for practices from 6 April 2025. The government’s summary for businesses explains that giving consumers objectively false information is a misleading action. The CMA’s announcement says it can now fine traders up to 10% of their global turnover, and named objectively false information as a priority in its first year. A generator inventing a feature doesn’t move the responsibility off the store.

Google Search

Google’s guidance on generative AI content allows AI-written content, but warns that using AI to generate many pages without adding value for users may break its spam policy on scaled content abuse. It asks for accuracy, quality and relevance, especially when content is generated automatically.

Google’s product data

The same guidance is specific about ecommerce: AI-generated product data such as titles and descriptions must be specified separately and labelled as AI-generated when you send it to Google, and AI-generated images must carry the right metadata. If your descriptions feed Google Shopping, the feed needs to say which ones were generated.

My view

The rules point the same way as good practice: accurate, specific, and checked. A description that states the real attributes of the product in your own voice passes all three, and it also sells better.

A workflow that works at scale

The safe way to use AI for descriptions is to make the facts the input, keep the model from adding any, and check the output against the facts before anything is published.

Start from structured attributes

Put the facts in fields first: dimensions, materials, weight, compatibility, care, what’s in the box. They usually come from several places, the supplier’s feed, the ERP, your own product team, and merging and checking them is the real work. A PIM is built for exactly this.

Generate with limits

Tell the model to use only the attributes it’s given, in your brand’s voice, at a set length, and to leave out anything it wasn’t told. Give it two or three of your best existing descriptions as examples of the tone.

Check automatically

Compare every number, material and claim in the draft with the attributes it came from. Anything that appears in the text but not in the data gets flagged. This is the step most bulk tools skip, and it catches the invented details before a customer does.

A person reviews what matters

A person reads everything flagged, every high-value or regulated product, and a sample of the rest. The rest publishes with a record of which descriptions were generated, so the labelling Google asks for is easy.

Measure it

Compare returns and conversion on the rewritten products with the ones left alone. If a batch of AI descriptions increases returns, something in the process is adding claims the products don’t live up to.

When to use AI, and when not to

When to get help

If your product data comes from several suppliers and systems, getting it clean enough for AI to write from is a data job before it’s a writing one. AI and automation covers merging and checking product data from several sources before it goes to a PIM or the store, with a person approving what matters. And if the products aren’t being found at all, see why a Shopify website isn’t showing up on Google.