Four thousand products nobody was ever going to write by hand
This is the job generation is genuinely good at. Built from your real attribute data, checked before anything runs, with the categories that actually make money given human attention.
In short: We audit your product data first, then generate descriptions from the attributes it actually contains, then hand the revenue-driving categories to a writer. Nothing gets stated about a product that the data does not support.
This is where generation earns its place
Most large catalogs have the same shape. The top hundred products got proper copy at some point, usually years ago. Everything behind them carries whatever the manufacturer supplied, or one line, or nothing at all. Those pages do not rank, do not answer the question a buyer has at the moment of comparison, and give a shopper no reason to choose the item over the identical-looking one three rows down. Fixing them by hand was never going to happen. Nobody has budgeted four thousand pieces of copywriting, and nobody ever will.
That is a genuine machine job. The task is repetitive, structured, and driven by data you already hold. A model turns a row of attributes into readable prose thousands of times without getting bored or sloppy in the way a person would by hour six. It is not the same as writing. It is closer to typesetting with judgment, and it is worth doing precisely because the alternative is leaving the long tail blank forever.
What goes in
Your attribute feed, category rules, spec sheets, and any tone guidance you already work to. Not a scrape of a competitor and not the model’s general knowledge of the product type.
What comes out
A short lead paragraph, the specifications in a consistent order, the practical fit-and-use notes, and enough variation between near-identical items that the pages are not duplicates.
How it is checked
A sample from every category is read against the source data before the full run. Then a second sample after it. Anything that asserts a fact absent from the feed is a defect, not a nuance.
How it stays current
New SKUs run through the same rules on a schedule, so the catalog does not quietly split into described products and forgotten ones again in eighteen months.
The data audit comes before any writing
Here is the failure everyone should be worried about. Generation from a thin or wrong attribute feed does not produce four thousand vague descriptions. It produces four thousand confidently wrong ones, written in the same assured voice as the correct ones, spread across your whole catalog in an afternoon. A blank field becomes a plausible guess. A millimetre value in a column labelled inches becomes a product forty times the size it is. Nobody notices until a customer receives the wrong thing, and by then it is on every page.
So we start with the feed rather than the copy. We check field coverage category by category, look for units that change meaning halfway down a column, separate genuinely empty values from zeros, and find the attributes that were mapped into the wrong field during some migration years ago. Then we sample real products and compare the data against the actual spec sheet. Usually this turns up a handful of problems worth fixing in the source rather than papering over in the copy, which is a better outcome for you anyway. Sometimes it turns up enough that we recommend cleaning the data before we generate anything. That conversation is uncomfortable and it is the point of doing the audit.
The categories that make the money get a person
Not every product deserves the same treatment, and pretending otherwise is how these projects go flat. Pull your revenue by category and the picture is usually stark: a small set of lines carries most of the margin, and everything else contributes in aggregate. The long tail gets generated copy, done well, from clean data. The lines that carry the business get written by a person who has looked at the product, read the returns notes, and understands why a buyer hesitates between two options.
That human copy is also where the merchandising decisions live. Which objection to answer first. What to say about the thing customers keep getting wrong. When to recommend the cheaper item because it is the right one, which costs a sale and buys a return customer. A model cannot make those calls, because they are commercial judgments rather than language problems. Structure matters here too, particularly for variant pages and category templates, and that is where this connects to content structuring and the wider SEO work.
What we won’t do
Invent a specification, ever
If the feed does not carry a load rating, a material, a compatibility claim or a certification, the description does not mention it. A missing sentence costs you nothing. A confidently wrong one costs you a return, a complaint, and sometimes a regulator.
Publish performance figures we cannot source
No percentages, efficiency gains or lifespan claims unless they come from your testing or a document we can point at. Models will produce a convincing number for any product on request, which is exactly why we do not ask them to.
Write reviews, ratings or testimonials
We do not seed product reviews, write customer questions and answers under invented names, or draft testimonials for anyone. Beyond being illegal in most of the markets you sell into, it poisons the one signal shoppers still trust.
How many of your products have no real description?
Send an export of your catalog and we will tell you what the data can support, what it cannot, and what it would take to fix the gap.
Field coverage by category, units that change meaning partway down a column, empty values that are being read as zeros, attributes mapped into the wrong field during an old migration, and a sample of real products compared against their spec sheets. It usually takes a few days and it decides whether we generate at all.
Will thousands of generated descriptions count as duplicate content?
Not if the source data is rich enough to make each page different. That is another thing the audit tests. Where two variants genuinely differ by one attribute, the right answer is often a single page with options rather than two near-identical pages, and we will say so.
Which products get written by a person?
The ones carrying your margin, plus anything with a safety, compliance or compatibility angle. We pull revenue by category with you and draw the line together. Typically that is a few hundred products written by hand and the rest generated.
What happens when we add new products later?
New SKUs run through the same rules and the same checks on a schedule you set. Without that step, catalogs drift back within a couple of years into described products and forgotten ones, which is the problem you hired us to solve.
