How to Generate Product Descriptions with AI

Generate product descriptions with AI while sticking to catalog facts

Generating product descriptions with AI means rewriting the title, short description and long description from data the store already has: measurements, materials, SKU, GTIN, brand and availability. The model organizes and formulates the information; it must not add certifications, origins or performance claims that are not shown on the product page. Start with the fields in the store system, lock what is factual, and leave only the wording to the model. This keeps the product page consistent with the catalog, feed and search assistants.

A Google Shopping crawler and an assistant reading a catalog do not “understand” the product: they read fields. In the feed, they look for id, title, description, link, image_link, price, availability, brand and gtin, or identifier_exists when the manufacturer’s GTIN is unavailable. What is usually missing is not style: it is the measurement in millimeters, the material in the title, the brand matching the packaging, or a description that does not contradict the attributes. Generic copy fills the screen while leaving blank the fields channels use to filter and compare products.

Why manually written descriptions are no longer enough for traditional search and assistants

A manually written description may work for the website, but channels and assistants need the same facts repeated in separate fields without contradictions.

Anyone updating three hundred product pages manually tends to copy the same paragraph from one item to another, leave the short description empty and put measurements only in the image. In WooCommerce, the path is Products, then the product, then Description and Short description; attributes are under Products › Attributes. If the volume appears in a “Capacity” attribute but is missing from the description, the feed may still export weak text. Merchant Center › Products › Feeds reads description as a product field, not as an image caption.

Assistants that cite a catalog prefer verifiable statements: material, size, brand and condition. Storefront language such as “superior quality” or “great gift idea” does not help someone answer, “Is it glazed ceramic and dishwasher-safe?” Generating product descriptions with AI is useful when the model reorganizes facts already present, not when it replaces the catalog with brochure-style language.

For people with little time for daily routines

Someone opening the store system between an order and serving a customer does not have a copywriting session ahead of them. They need a workflow that confirms what is true, measurement, material and GTIN, launches the rewrite and reviews the text, rather than spending an evening rewriting every product page. The operational goal is to fill empty fields and align the title and description, not to produce a novel for every item.

The risk of generic AI-generated descriptions: specifications that do not exist

The risk is not flat writing; it is a sentence that introduces a certification, origin or performance claim absent from the merchant’s materials.

A model trained on web content fills gaps. If the title says “300 ml ceramic mug” and the web contains mugs described as “food-certified” or “handmade in Umbria,” the text may attach those words to the wrong product page. In the feed, that sentence becomes the description. Merchant Center does not display an “invented origin” warning: it displays a product with a claim the store cannot support if the packaging or a customer comparison challenges it.

A description that adds a certification missing from the product page is not richer copy: it is false data that the feed will carry to Google Shopping and assistants.

The same applies to numbers. “It weighs 180 g” is a fact if it appears in WooCommerce, the price-list PDF or the packaging photo. If the model infers it from a similar product, the weight enters the description and may conflict with shipping_weight. Cosmetics have an additional limitation: therapeutic or clinical effects cannot be attributed when the brand has not documented them. “Cleanses” and “moisturizes” do not become “treats” or “prevents” simply because the model has seen those words on other product pages.

An artisan selling ceramics with around forty products

With a small catalog, every mistake is visible. Forty glazes, two clay bodies and three capacities: if the rewrite confuses stoneware with earthenware, or adds “dishwasher-safe” to a piece the workshop has not tested, the product page is inaccurate. In this situation, the model can only rephrase the glaze, color, capacity in milliliters and use stated by the workshop. Everything else stays out, even if it sounds good.

What must remain with the merchant and what the model can do

The merchant keeps the measurable facts; the model handles the sentence, the order of information and the language of the product page.

Before launching any rewrite, it is useful to define a closed list. This is not a creative brief: it is the boundary of what is true.

1. Lock identity: SKU, feed ID, brand exactly as shown on the packaging, and GTIN only if you read it from the label, PDF or CSV column.
2. Lock physical details: measurements, weight, material, color, size, package contents and condition, new, refurbished or used.
3. Lock constraints: what is not declared, including origins, certifications and additional performance claims.
4. Ask the model for a title that repeats the brand, product type and a useful measurement, and a description using only points 1-3, without empty slogans.
5. Review line by line: every number and material must already appear in an attribute, the original title or a source document.

The GTIN should not be “completed” by searching online. Read it from the store’s materials, packaging photos, price list or CSV, and it must pass the GS1 check digit. If it is missing, the feed can use identifier_exists with the correct value under the channel’s rules, rather than a code found in another listing.

What the model can do without changing the facts

It can shorten an all-caps title, remove messy HTML, place the measurement next to the product name, write a description that answers typical questions, what material it is made from, how large it is and what it is intended for, and align the website language with the feed fields. It cannot decide that a fabric is “organic” when the attribute says only “cotton.” It cannot invent the brand story. It cannot copy reviews or an aggregate rating unless they reflect real reviews the store has collected.

How a rewrite with fact verification works

A rewrite with fact verification extracts claims from the proposed text, compares them with the product page and a brand knowledge base, and discards anything that does not match.

The claim verifier treats brands, certifications and brand history as sensitive data: if they are not on the product page or in public sources already associated with that brand, they are not included. A consistency gate compares numbers. If the web reports 28 cm while the merchant’s title says 30 cm, 30 cm wins: the store’s data takes precedence over the web. If the text fails, the process is retried using the error feedback; in a tool built this way, the rewrite credit is not charged when the text is rejected.

In practice, this is how you work on a WooCommerce product page.

1. Open the product and check the attributes under Products › Attributes: visible name, value and whether the attribute is used for variations.
2. Copy only what you see into a “facts” block: “glazed ceramic,” “300 ml,” “blue,” brand and SKU.
3. Generate the text while explicitly asking the model not to add origins, standards or performance claims.
4. Compare every sentence with the facts block. An extra word that introduces a new fact is an error, not an improvement.
5. Align the short description, the first two useful sentences, with the long description, then export the feed fields: title, description, brand, gtin, price and availability.

The consistency gate: why merchant data takes precedence over the web

The consistency gate is the rule that a measurement, material or brand found online never overrides what the store has declared.

Without this rule, the product page becomes a collage. The title says “100% wool,” a third-party ecommerce site sells a similar item described as “wool-acrylic,” and the model combines the two into “wool blend.” The feed sends a description and title that conflict. Google Shopping does not “correct” the item: it displays what it receives. An assistant summarizing the product may cite the blend, while the customer receives 100% wool. The damage affects returns and trust, not just the wording.

If a measurement found online contradicts the merchant’s title, the merchant’s data wins and the external number does not enter the product page.

The same logic applies when filling empty fields. Specifications, materials, dimensions, weight, a verified brand and a certain Google category can be completed with facts found online and a visible source; if a fact contradicts the title, it is not written. Completing a field is not the same as inventing. Again, the GTIN is not searched for online: it is either in the store’s materials or the absence is declared.

For anyone syncing the same catalog to Amazon, eBay and Etsy

Anyone publishing the same product on several storefronts has a copying problem, not an inspiration problem. If each channel receives a different description, customer questions multiply and returns may quote sentences that do not appear on the website. The useful sequence is one master product page in WooCommerce, one rewrite based on the facts, and then a copy in each marketplace’s format. Synchronization by SKU prevents duplicates: the same code should not generate two products just because the title was reformulated.

Before-and-after example using a real catalog product page

A useful rewritten product page repeats the same numbers and materials already present in the title, attributes or price list, and removes statements that cannot be verified.

Before, source data visible in the store system: SKU TAZ-300-BLU, Capacity attribute 300 ml, Material attribute glazed ceramic, color blue, workshop brand, no GTIN in the column, price EUR 18.00 and availability in stock:

Title: BLUE CERAMIC MUG 300ML SUPER DEAL. Description: Beautiful premium-quality mug, great gift idea, fast shipping across Italy. Food-certified and handmade in Umbria. Dishwasher-safe at 75 degrees.

After, staying with the merchant’s facts, no certification, origin or dishwasher cycle declared on the product page:

Title: Blue glazed ceramic mug, 300 ml. Description: Glazed ceramic mug, blue, with a capacity of 300 ml. Brand as shown on the product page. New product, in stock. For everyday use with hot and cold drinks according to the workshop’s stated information; no certification, geographic origin or dishwasher resistance is declared. Internal identifier SKU TAZ-300-BLU; manufacturer GTIN not present in the catalog.

In the feed, the aligned fields become: id (SKU), title (new title), description (new text), link (product page URL), image_link, price (18.00 EUR), availability (in stock), brand and identifier_exists indicating that the GTIN is absent. Include aggregateRating only if it reflects real reviews the store has collected.

From description to channel readiness for Google Shopping, ChatGPT and Meta

A correct description is a feed field; channel readiness is the combination of title, description, image, price, availability and identifiers without contradictions.

An ACP feed is a product file that some assistants use to read catalog, price and availability, with the same identifying fields as a shopping feed: id, title, description, link, image_link, price, availability, brand, gtin or identifier_exists. It is not a blog post: it is a table. If description is empty or contradicts title, the assistant has little precise information to cite.

For Google Shopping, the operational path is Merchant Center › Products › Feeds, checking rejections involving the description field and identifiers. For Meta, image quality and completeness already matter. For photos, it is worth keeping an eye on Google's image-quality guidance, which may include minimum side dimensions: an excellent description with an unreadable thumbnail is still an incomplete product page. Anyone starting from a price list can turn its rows into product pages; the transition from PDF price list to store-ready product pages is designed to avoid rewriting by hand what the document already contains.

Common mistakes when generating product descriptions with AI

The most frequent mistakes are copy that invents facts, empty feed fields and GTINs copied from other product pages. Each has a recognizable message and a specific correction.

Error 1. Merchant Center message: “Missing value: description.” This happens when WooCommerce has only a title or text inside images. Resolve it by completing Description, and, if used, Short description, with sentences repeating the material and measurement, then realigning the feed under Merchant Center › Products › Feeds.

Error 2. Message: “Invalid GTIN” or failed check digit. This happens when a code seen online is pasted in or zeros are added at random. Resolve it by reading the GTIN only from the packaging, PDF or CSV and checking the GS1 digit; if the manufacturer does not provide one, declare its absence instead of inventing it.

Error 3. Store-side message: HTML description displaying visible tags on the page (<p>, duplicated <br> tags or styles pasted from a document). This happens when content is pasted from a generator or spreadsheet. Resolve it by cleaning the HTML and keeping simple paragraphs; the feed needs text, not a layout page.

Error 4. Silent conflict, without a channel warning: title “30 cm” and description “28 cm.” This happens when the model combines web sources. Resolve it by enforcing the consistency gate and checking the numbers against Products › Attributes.

Error 5. Image rejected or at risk of rejection because one side is too small for Google's image-quality guidance. This happens with price-list thumbnails. Replace the file rather than lengthening the description to “compensate.”

• Do not ask the model to “make it more sellable” without a locked list of facts.
• Do not reuse the same description for variations that differ by size or color: the title must distinguish them, and grouping should use item_group_id only when the variations are confirmed.
• Add reviews or ratings only if they reflect real reviews the store has collected.
• Do not translate into another language while leaving measurements in a different unit from the product page.
• Do not copy therapeutic claims onto cosmetics or supplements.

How to start with a WooCommerce catalog scan

Start by reading the public catalog in read-only mode, product by product, without changing anything until a field is confirmed.

An anonymous scan of up to one hundred products, without an account, shows two separate scores, traditional search and assistant readiness, along with field relevance and completeness. It is not a promise of visibility: it is a list of gaps, such as an empty description, a missing or invalid GTIN, a shared image or an all-caps title. From there, decide what to rewrite and what only needs cleaning. Mechanical corrections, Title Case, HTML, new/refurbished/used condition, a no-GTIN declaration and a proposed Google category, remain subject to explicit confirmation and can be undone for thirty days.

For an overview of readiness for assistants, the AEO checker for WooCommerce products reads the public catalog and returns scores for each product page. Anyone who needs to expose a catalog to ChatGPT can find the expected fields on the page about the WooCommerce product feed for ChatGPT. In both cases, description is only one field among others: without id, price and availability, it cannot go very far.

How Katapic rewrites product pages without replacing the seller’s judgment

Katapic rewrites WooCommerce product pages with claim verification, lets merchant data take precedence and charges one product-page credit only when the text passes the checks.

The engine does not invent product specifications. It includes a claim verifier for brands, certifications and brand history, a brand knowledge base containing public facts and sources, and the consistency gate described above. A second attempt uses feedback; if the text does not pass, the credit is not charged. A rewritten product page costs 1 credit, and an image costs 1 credit. There is no subscription: credits do not expire, and there is a one-time free trial with no card required.

The initial scan is anonymous and read-only. Synchronization with WooCommerce is bidirectional by SKU and does not create duplicates, the official plugin is on WordPress.org. Variations can be grouped with item_group_id after confirmation, a Google taxonomy category can be suggested for confirmation, and the Shopping feed can be exported as a CSV containing only ready product pages. Bulk actions show the credit cost before confirmation. Completing empty fields uses sourced facts; the second credit for that action is charged only if at least one field is written. Details about the method and its limits are available on Katapic: a catalog tool, not an editor signing on your behalf.

Frequently asked questions

How can you create AI product descriptions without inventing specifications?

First lock the facts in the store system, measurements, materials, brand, SKU and GTIN if present, and ask the model only to rephrase them. Delete every number or certification that does not appear in an attribute, original title, PDF or package. A claim verifier and consistency gate are designed for this: store data takes precedence over the web.

Can AI write WooCommerce product pages on its own?

It can suggest a title and description in seconds, but WooCommerce remains the source of truth. Attributes, price, availability and images cannot be guessed. The correct workflow is a read-only scan, confirmation of the facts, rewriting and then synchronization by SKU. Without confirmation, silent automatic generation is a quick way to publish errors in the storefront and feed.

How much does it cost to rewrite product descriptions without a subscription?

It depends on the tool. On Katapic, a rewritten product page costs 1 credit and credits do not expire; there is no subscription or renewal. If fact verification rejects the text, the credit for that rewrite is not charged and the process is retried using the feedback. An image, if needed, costs a separate credit. You choose the volume rather than paying a monthly installment.

What is the difference between a website description and one for search assistants?

On a website, you can use a workshop or boutique tone. For assistants and feeds, you need sentences a system can cite without misrepresenting the product: material, size, brand, condition and availability. Slogans and urgency such as “super deal” do not help when someone has to compare two products.

Where can you check whether a WooCommerce catalog is ready for assistants?

Start with a product-by-product reading of the public catalog, with separate scores for traditional search and assistant readiness, plus missing fields. Katapic’s WooCommerce AEO checker performs this reading without an account, for up to one hundred products. Then correct descriptions, identifiers and images before exporting the feed to Shopping or assistants.

Can you start from a PDF instead of product pages already in the store?

Yes, if the price list already contains a name, price and at least one measurement or material. Extract the rows and embedded images, match them to products and generate the description from the facts found, not from an empty template. If present, the GTIN must be in the PDF or packaging photo. A PDF without numbers produces only prose; it is not a product page.

What should you do if Merchant Center reports a missing description after rewriting?

Open the product in WooCommerce, check that Description is not empty and that the plugin or feed maps that field to description. Regenerate the file, upload it through Merchant Center › Products › Feeds and wait for processing. If the text appears on the page but not in the file, the problem is the mapping, not the AI. If the text exists but contradicts the title, correct the product page before resubmitting the feed.

Read more

• WooCommerce AI product descriptions