Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

AI in Marketing › AI Applications in Marketing

Generating Product Descriptions from Catalogue Data

How to generate product descriptions with AI from catalogue data: attribute hygiene, templates by category, banned claims, sampling review and refresh.
Guideintermediate11 min readUpdated

Why catalogue copy is a data problem first

A writer drafting one product page can pick up the product, read the manufacturer's sheet, ask the buyer a question and notice that the listed weight looks wrong. A model drafting ten thousand pages can do none of that. It receives a row of attributes and an instruction, and it returns fluent prose whatever the row contains. If the row says "colour: various", the description will choose a colour. If the material field is empty, the description may supply one that suits the category. If two products share a template and one has a feature the other lacks, the model will sometimes give both of them the feature.

That changes where the effort goes. The AI for content guide covers drafting from a brief page by page, where the brief and the supplied sources make the draft checkable. At catalogue scale the brief is the attribute table, and the quality of the descriptions is capped by the quality of that table. A team that spends its time polishing prompts while the catalogue carries inconsistent units, free-text fields and missing values will get polished prose about wrong products.

It also changes what success means. The goal is not a beautiful description of each product but an accurate, specific, readable one for every product, produced at a review cost the team can sustain. The rest of this guide is organised around that goal, and the table summarises where the risks sit.

StageWhat the model doesTypical failureControl
AttributesNothing; this is the inputMissing, free-text or inconsistent fieldsRequired fields per category; controlled values; units in one system
InstructionsFollows a template per categoryOne template for every category; generic benefitsCategory-specific structure, vocabulary and must-mention attributes
GenerationWrites from the rowInvented features, borrowed attributes, embellished claimsUse-only-supplied-attributes rule; banned-claims list
ReviewNothing; people sample the outputReviewers read a handful and approve the batchStructured sample per category and template; fix causes, not instances
PublishingNothing; the platform shows the copyCopy drifts from the catalogue after changesRegenerate on attribute change; version and log every run

Getting attributes ready

Start with an inventory of the fields the catalogue actually holds, category by category, and how complete each one is. Most catalogues have a core set that is reliable because orders depend on it, such as title, price, stock and size, and a long tail that is not: materials, compatibility, care instructions, what is in the box, country of origin. The long tail is where descriptions get their specificity, and it is also where gaps turn into invention.

For each category, decide which attributes a description must use and mark them as required. A description of a rucksack without capacity, weight and material is not worth generating; flag the product for data entry instead. Convert free-text fields into controlled values where possible, so that "waterproof", "water resistant" and "showerproof" mean three different, defined things rather than whatever the supplier wrote. Put every measurement in one unit system with the unit in a separate field, because a model asked to convert units will occasionally convert them wrongly, and a model given mixed units will mix them in the prose.

The catalogue that feeds the descriptions should be the same one that feeds the Product Feed for shopping ads and marketplaces. Two sources of truth guarantee that the page and the feed will disagree eventually, and the category and product pages guide explains why price, availability and identifiers have to match between the page, the markup and the feed. Generated copy that states an attribute the feed contradicts is the same problem in prose.

Finally, give the model the facts the attributes cannot carry, but only as structured fields a person maintains: who the product suits and who it does not, the problem it solves, the one thing buyers most often ask about. These come from the people who know the range, from reviews and from support questions, and they are the difference between a description that restates the specification table and one that helps someone decide.

Templates and instructions by category

A single instruction for the whole catalogue produces the same description shape for a kettle and a pair of running shoes. Write the instruction per category instead, and treat it as a document the category owner signs off. Each one sets the structure (an opening line that says what the product is and who it is for, then the attributes that decide the purchase, then use and care), the length, the vocabulary the category uses and avoids, and the attributes that must appear in every description.

The product page optimisation guide describes product copy in layers: a short description that gives the reason to want the product, a specification table that answers exact questions, and a longer description for the buyer who wants more. Decide which layers the model writes. A common split is to generate the short description and the long description from attributes, render the specification table directly from the catalogue without any model involved, and leave the title to a rule-based template, because titles carry the identifiers and variant names that must be exact.

Match the instruction to the Search Intent of the category's pages. Buyers of commodity items search by specification and want the facts fast; buyers of considered purchases search by problem and want to know whether the product suits them. The instruction for the first kind keeps descriptions short and attribute-led; for the second it asks for the suits-and-does-not-suit line and the use case.

Every instruction carries the same core rule: use only the attributes supplied, and when a fact the template asks for is missing, leave that sentence out rather than filling it. Ask for output in a structured form, the description plus a list of the attributes it used, so that review can compare the two mechanically.

Claims the model must never add

Some sentences are harmless when wrong and some are not. A banned-claims list names the second kind and is written by the people who would have to answer for them, typically the category owners with whoever handles legal and compliance review. It goes into every instruction and into every reviewer's checklist.

Typical entries are claims about health, safety, environmental impact and performance that the business has not substantiated; comparatives and superlatives ("the best", "lasts longer than", "the lightest in its class"); certifications, awards and standards unless they are attributes in the catalogue; compatibility statements beyond the compatibility field; guarantees, warranties and returns terms, which belong to the policy pages; and anything about price, discounts or stock, which change faster than copy does. Categories with regulated products add their own entries.

The list works in two directions. The instruction tells the model not to make these claims; a simple text check then searches every generated description for the banned words and phrases and holds any match for a person. The check will catch things the instruction did not prevent, because models follow negative instructions imperfectly. The accuracy and review workflows guide treats claims of this kind as the highest review tier and explains how claim checks fit into sign-off; this guide applies that tier to a catalogue.

Sampling review at catalogue scale

Nobody reads every description in a large catalogue, and nobody should approve one without reading any. Sampling sits between those, and it works only if the sample is structured. Draw it from every category and every template, weight it towards new templates and recently changed instructions, and include products with the fewest attributes, because they are where invention concentrates.

For each sampled description, the reviewer compares the copy with the attribute row and marks any statement the row does not support, any banned claim the text check missed, any attribute stated wrongly and any sentence so generic it could describe a competitor's product. The structured output from the generation step makes this quick: the reviewer checks the listed attributes against the row and reads the prose for anything extra.

The important part is what happens next. Each error is traced to its cause, and the cause is fixed: a missing attribute is entered, an ambiguous field is split, the category instruction is tightened, a phrase is added to the banned list. Then every description that shares the cause is regenerated, not just the one in the sample. Fixing the sampled description and moving on leaves the same error on every unsampled product with the same cause.

Some categories should not be sampled at all; they should be read in full. Where a wrong statement could harm someone, such as products for children, electrical goods, supplements or anything with an allergy or safety dimension, the review cost of reading every description is the price of generating them.

Illustrative example — not a real result

A first generation run at an outdoor gear retailer

Brindlecove Outdoor, an invented outdoor gear retailer, generates descriptions for 6,400 products across 22 categories. The attribute inventory finds that 31% of products have no material field and that weights are recorded in both grams and ounces. The team makes 5 attributes required per category, converts all weights to grams and holds 1,150 products back for data entry. The first run covers the remaining 5,250. A sample of 260, at least 10 per category, finds 23 descriptions with an unsupported statement; 14 of them come from one cause, the tents template inferring "waterproof" from a field that records only fabric coating. The field is split into coating and rated water resistance, the instruction is tightened and all 410 tent and shelter descriptions are regenerated. The banned-phrase check holds 38 descriptions, mostly for "best" and "lightest". The 180 products for children are read in full. Conversion rate on regenerated product pages is then tracked by category against the previous quarter rather than as one total.

Search guidance and duplicate copy

Google's published guidance on AI-generated content, cited in this page's sources, says its ranking systems aim to reward high-quality content however it is produced, and that using automation, including AI, to generate content with the primary purpose of manipulating search rankings is a violation of its spam policies. Its guidance on helpful, people-first content offers self-assessment questions, and among its warning signs are using extensive automation to produce content on many topics and mainly summarising what others say without adding much value. Neither page describes a penalty for AI use as such.

For a catalogue, the practical reading is about duplication and genericness rather than authorship. The category and product pages guide makes the point that the manufacturer's text appears on every retailer's site, so a page that repeats it gives the engine nothing to prefer, and it describes the usual compromise: unique content assembled from structured attributes, plus written guidance for the products that matter most. Generated descriptions fit that pattern when they draw on attributes and knowledge the retailer actually has. They fail it when every description in a category is the same sentence with the product name swapped, which is generic copy at scale whoever wrote it. Variety in the inputs, not in the wording, is what makes descriptions distinct.

What happens to AI-written pages in search beyond what the guidance states is not documented, and this guide does not speculate. The AI in search results guide covers the search channel itself, including what the spam policies say about content produced at scale.

Keeping descriptions in sync

A description is correct on the day it is generated and starts to drift the day an attribute changes. A supplier switches a material, a product gains a variant, a compatibility list grows, a care instruction is corrected. Hand-written copy has the same problem, but generated copy makes it easier to solve, because the description can be regenerated from the corrected row.

Set the rule that a change to any attribute a description uses triggers regeneration and a check of that description against the banned list. Store every description with the attribute values, the instruction version and the date it was generated from, so that anyone can see why a description says what it says and which descriptions an instruction change will affect. When a category instruction changes, regenerate the category and sample it again before publishing.

Measure the effect by category, not in aggregate. Compare conversion rate (CVR) and returns for regenerated pages against the same category's previous period, using the conversion rate calculator for the arithmetic, and read return reasons and product questions for signs that descriptions are setting wrong expectations. A rise in "not as described" returns in one category is the clearest signal that its attributes or its instruction need attention.

Common mistakes

  • Starting with the prompt instead of the data. Polished instructions cannot rescue missing or inconsistent attributes.
  • One instruction for every category. Every product ends up with the same shape and the same generic benefits.
  • Letting the model fill gaps. A missing attribute should drop a sentence or hold the product, never produce a plausible guess.
  • No banned-claims list, or one only in the prompt. Models follow negative instructions imperfectly; a text check catches what the instruction misses.
  • Reviewing a handful and approving the batch. Sample by category and template, and read sensitive categories in full.
  • Fixing the instance, not the cause. Every product that shares the cause carries the same error.
  • Generating titles and specification tables. Identifiers and exact values belong in rule-based templates rendered straight from the catalogue.
  • Letting copy and catalogue drift apart. Regenerate when attributes change, and keep a record of what each description was built from.

Frequently asked questions

Can AI write product descriptions for a whole catalogue without anyone reading them?

No. A model writing from attributes makes a small share of errors that are hard to predict: a feature carried over from a similar product, a unit converted wrongly, a benefit the product cannot support. Nobody reads every description in a large catalogue, but someone must read a structured sample from every category and template, fix the cause of each error at the source, and read every description in categories where a wrong claim could cause harm.

Will Google penalise product descriptions written with AI?

Google's published guidance says its systems aim to reward helpful, people-first content however it is produced, and that using automation to generate content primarily to manipulate rankings falls under its spam policies. It does not describe a penalty for AI use as such. The practical risk for a catalogue is generic, near-identical copy that adds nothing to the specifications or the manufacturer's text, which is a quality problem whichever way it was written.

Should we replace the manufacturer's description with generated copy?

Not necessarily. The manufacturer's text is often the most accurate source for technical detail, and the problem with it is only that every retailer uses it. A common pattern is generated copy built from the catalogue's own attributes and the retailer's knowledge of who the product suits, with the manufacturer's description kept and clearly labelled lower on the page, and human-written guidance for the products that matter most.

Terms used on this page

Tools to try

Where this page is used

Also related

Sources

  1. Google Search's guidance about AI-generated content — Google Search Central
  2. Creating helpful, reliable, people-first content — Google Search Central

Related