Four ways to fill catalogue descriptions — and none is best in general. What each does well, where its limit sits, and how to choose for your catalogue size.
The question "Textory or ChatGPT?" comes up often, and it is the wrong question. The answer depends not on the tool but on what exactly you are filling: three flagship products or a catalogue of fifteen thousand items, a one-off job or a stream of new cards every week.
There are four ways to fill product descriptions: keep what arrived from the supplier, hire a copywriter, write in an AI chat, or run the catalogue through a feed-to-feed pipeline. We will go through each one — what it does well and where its limit sits. This is not a ranking: each of the four has a situation where it wins.
A catalogue is rarely empty. Most often descriptions already exist — the ones that came in with the supplier's price list. The question of what to fill descriptions with appears once you find out what state they are in.
We measured that. On 8 September 2026 we read 21 public YML feeds from Ukrainian stores — 89,718 products — and counted descriptions that match, word for word, the description of another product in the same feed. There are 19,674 of them, 22% of everything we read, and they appear in 17 feeds out of 21. Another 5,779 products (6%) have no description at all. The method and the extreme cases are in a separate write-up.
That is the state of 21 specific feeds on the date of the measurement, not a picture of the market. But this is exactly the picture the choice of approach starts from.
What it does well. Fills the catalogue in an hour and costs nothing. Technically it is the simplest route: the import module takes the description field from the price list and puts it into the card.
Where the limit is. The same text sits with everyone who works with that supplier. A search engine sees several pages with identical content and shows one of them as the main version — which one is its decision, not yours. There is also the format: a supplier description rarely has headings, a specification list or a meta description.
When it fits. At the start, while the catalogue is still being assembled, and in categories you keep for completeness and do not promote.
What it does well. It brings depth the feed does not have. A person knows what customers ask about this product in chat, why it gets returned, how it differs from the one next to it on the shelf, and how your audience speaks. No automated pass invents any of that — the data is neither in the title nor in the specifications.
Where the limit is. Scale. Time and money grow almost linearly with the number of cards: a thousand products means a thousand texts, and no arrangement changes that. On a catalogue of tens of thousands of items, the manual route hits the calendar before it hits the budget.
When it fits. Flagships, complex niches with a long purchase decision, category pages, ad copy. In other words, wherever a single text pays for itself on its own.
A category in which ChatGPT, Jasper and Copy.ai are examples, not exceptions.
What it does well. Flexibility. You can ask for any tone, rewrite a paragraph ten times, argue with a phrasing, build a structure from scratch. For one page, for a dozen cards, for testing how something sounds a different way, the tool is excellent.
Where the limit is. It is not built around a product feed, and that is by design, not a flaw. The consequences are concrete: you assemble the XML yourself; there are no target name_ua and description_ua fields; keywords and the meta description have to be requested separately and formatted by hand; there is no queue over the whole catalogue with pause and resume; and there is no measurable indicator to compare "before" against "after". On twenty cards none of this is noticeable. On two thousand, every one of those points turns into a manual loop.
When it fits. Individual pages, small batches, a draft structure you then scale with something else.
This is what Textory, our product, does. We covered the mechanics of the rewrite itself separately: how to rewrite an entire catalogue. Here we only cover what makes this route different from the previous three.
What it does well. It closes the loop: it takes XML, YML, CSV, Excel or a feed URL and returns a file ready for import in the platform's format — Prom, Rozetka, generic XML, CSV, JSON. Processing runs through a queue: thousands of cards, real progress, pause and resume. The description, the title, the keywords and the meta description come back in a single call rather than in four passes. The Ukrainian name_ua and description_ua fields are filled along with the rest — in our measurement of 8 September 2026, five feeds out of 21 supplied them, and keywords only one.
The GEO mode stands apart: it restructures the description for citation by generative systems and calculates a GEO score from 0 to 100. The score is deterministic — a heuristic built on six components (a direct answer up front, question-shaped subheadings, fact density, explicitly named entities, a specification list, and length), not a model praising its own text. The same text always produces the same number, so "before" and "after" compare honestly. The details are in the piece on GEO optimisation of descriptions.
You can check the scale without a conversation: textory.com.ua has demo catalogues of up to 11,806 products. Payment is for the actual volume of text, with current terms on the product site. Catalogues are stored encrypted, data is not passed to third parties and does not go into AI training; an NDA mode is available separately.
Where the limit is. The description is built from the product's own specifications. If the feed carries material, size, compatibility and composition, there is something to build from; if the feed is empty, the pipeline has nothing to put into the text, and admitting that at the entrance is more honest than getting invented figures in a description. For flagship products an automated pass is a good draft that deserves editing. And it does not replace assortment, prices and delivery terms: a rewritten description removes a technical loss, it does not create demand.
| Criterion | Textory | Generic AI writer | Manual copywriter | Supplier descriptions as they are |
|---|---|---|---|---|
| Working with a product feed | Takes XML, YML, CSV, Excel or a URL and returns a ready platform feed | Text in a chat, you assemble the feed yourself | None | The feed exists, but the descriptions are not unique |
| Catalogue scale | Thousands of cards in a queue, with pause and resume | One by one or in small batches | Months of work | Instant |
| Ukrainian marketplaces | name_ua and description_ua fields, Prom and Rozetka export presets |
Generic text without the feed's target fields | Depends on the contractor | — |
| SEO structure | Headings, specification lists, keywords and a meta description in one call | Has to be requested separately and formatted by hand | Depends on the contractor | None |
| Readiness for AI search | GEO mode and a deterministic 0–100 score | No measurable indicator | None | None |
| Payment | For the actual volume of text | Subscription or tokens, not tied to the catalogue | For the contractor's time or volume | — |
| Catalogue data | NDA mode, catalogues stored encrypted, not used for AI training | Depends on the service's policy | By agreement | — |
A few rules that cover most situations.
In practice one option alone is rarely the choice. A working arrangement looks like this: the pipeline handles the bulk of the catalogue, a person edits the twenty or thirty cards that matter most, and the AI chat stays for individual pages and phrasings where a dialogue is needed. Importing as it is does not disappear either — it simply stops being the only thing in the catalogue.
The order of work is the same as in any feed job: products with no description at all first, then the largest groups of copies, then a preview on a single product before launching the whole catalogue, then importing one card back into the marketplace — and only after that the full volume.
None of these approaches creates demand. A description affects what the system and the person see on the page, and does not affect assortment, prices, delivery terms, links or competition in the category. We do not quote percentages of sales or ranking growth — nobody who quotes them knows them.
And separately on AI search: nobody, ourselves included, can promise a mention in an AI answer. GEO mode makes the text structurally more suitable for citation, while the decision to cite is made by the model and depends on dozens of factors outside the description.
Choosing between the four approaches is not about picking the best tool, it is about matching volume against depth. A dozen texts are written by hand. Ten thousand go through a pipeline. Everything in between is usually split between the two, and that split is worth planning before you start.
To see what your feed would produce, go to textory.com.ua — the demo catalogues and current payment terms are there. The rest of our products are in the own products section. If the task is wider than the text and the whole path of the feed needs review, the review is free, we reply within two hours, and we work under a contract.
Your opinion helps us create better content
Help others learn about it - share the article on social networks
Thank you for helping us grow
Writes about what he builds himself: online stores on OpenCart, applications on Next.js, integrations and site speed. The articles carry measurements and checks a reader can repeat on their own project, not general advice. Commercial development since 2015.
Answers to common questions on the topic
The difference is not in text quality but in what the tool is built around. A generic AI chat works with text in a dialogue: you assemble the feed yourself, there are no target name_ua and description_ua fields, keywords and the meta description have to be requested separately, and there is no queue across the whole catalogue. Textory works with a product feed: it takes XML, YML, CSV, Excel or a URL, processes thousands of cards through a queue with pause and resume, and returns a file ready for import in Prom, Rozetka, generic XML, CSV or JSON format. Description, title, keywords and meta come back in a single call. On one page the difference is invisible; on two thousand it is the difference between a job and a manual copy loop.
When each individual text pays for itself: flagship products, a complex niche with a long purchase decision, category pages, ad copy. A person knows what the feed does not contain — what customers ask about the product, why it gets returned, how it differs from the one beside it on the shelf. No automated pass has that data. The limit of the manual route is scale: time and money grow almost linearly with the number of cards, and on tens of thousands of items that runs into the calendar.
Yes, and in practice that is what usually happens. The pipeline handles the bulk of the catalogue, a person edits the twenty or thirty cards that matter most, and the AI chat stays for individual pages and phrasings where a dialogue is needed. The order of work is the usual one: products with no description at all first, then the largest groups of identical descriptions, then a preview on a single product before launching the whole catalogue, then importing one card back into the marketplace — and only after that the full volume.
The GEO score in Textory is a deterministic heuristic from 0 to 100, not a model grading its own text. It counts six components with fixed weights: a direct answer up front, question-shaped subheadings, fact density, explicitly named entities, a specification list, and length. Two things follow: the same text always yields the same number, so “before” and “after” compare honestly, and scores for different products are comparable with each other. What it does not show: it does not forecast rankings, it does not calculate the probability of a mention in an AI answer, and it does not verify whether the facts in the text are true.
No, and that applies to all four equally. A description affects what the search engine and the person see on the page; it does not affect assortment, prices, delivery terms, links or competition in the category. A unique description removes a technical loss — the fact that the page adds nothing new to what is already in the index — but it does not create demand. And separately on AI search: nobody, ourselves included, can promise a mention in an AI answer; GEO mode only makes the text structurally more suitable for citation.
Subscribe to our newsletter and receive useful tips, insights and news about web development, marketing and business.
Web developmentOn 21 September 2026, the first day of per-route request tracking, the internal /api/health endpoint received 13,666 requests, 60% of application traffic, and every one of them hit the shared database. Here are the two settings that fixed it, what they cost us, and how to check your own healthcheck.
Web developmentOn 21 September 2026, a type check on our site reported 154 errors, even though the build had been green every time: the check's result was simply thrown away. Inside, we found zeros in link statistics and in the analytics export, sorting that didn't sort, and tests that never ran. Here is how we worked through them, what we changed, and how to check your own project.
Web developmentOn 22 September 2026 we checked the HTTPS certificates on 116 sites we had already measured for market studies, and for 104 of them we compared the state with August. All 116 passed verification, 108 run on 89- and 90-day free certificates, and only one site's certificate has still not renewed, even though with typical settings it should have by now. We walk through the measurements and the limits of the method, and show how to check your own site with a single command.
