Descriptions in your voice
Written against your actual brand guidelines and a few examples you approve, not a generic template applied five hundred times.
AI for Product Catalog Management: Fix the Data, Not the Theme
How-to — — by Mahmoud Zalt
Most store conversion problems are catalog problems. How an AI agent finds the gaps, fills them consistently, and keeps them filled.
Why the catalog is the quiet bottleneck
When sales are soft, most people look at the theme, the ad creative, or the checkout. Those are visible. The catalog is not, and it is quietly upstream of almost everything: on-site search finds nothing because the fields it indexes are empty, filters are useless because half the items lack the attribute, ads underperform because the feed is thin, and marketplace listings get suppressed for missing required data.
None of that shows up as a catalog problem in your analytics. It shows up as a conversion problem, which is why people redesign a page instead of filling in a field. The tell is usually on-site search: if searching your own store for an obvious product returns nothing useful, you have a data problem wearing a design problem's clothes.
The reason it stays broken is not that anyone disagrees. It is that fixing it is four hundred small identical decisions, and nobody wants to spend a Saturday on that. Which makes it an almost perfect candidate for delegation: repetitive, rule-driven, and every single output is something you can eyeball in a second.
Start with an audit, not with writing
The instinct is to start generating descriptions. Do the boring thing first: find out what is actually missing. An agent can walk the whole catalog and produce a gap report far faster than you can scroll it, and the report usually reframes the project.
| What to check | Why it matters |
|---|---|
| Products with no description or under 30 words | Thin pages rank poorly and convert worse |
| Images missing alt text | Accessibility, plus image search and feed quality |
| Missing size, material, colour, or dimensions | Filters silently exclude the item from every refined search |
| Inconsistent naming across similar items | Breaks sorting, grouping, and any automated feed rule |
| Uncategorised or wrongly categorised items | Invisible in navigation, wrong in the sitemap |
| Duplicate or near-duplicate listings | Splits reviews and inventory, confuses customers |
| Missing required marketplace fields | Listings get suppressed without an obvious error |
The audit is also what makes the rest of the work safe. Once you know that 340 items lack a material field and 80 have no description at all, you can hand over a defined batch with a defined success condition instead of asking an agent to vaguely improve your catalog. Vague briefs are the main reason these projects go sideways.
Enrichment, with your rules attached
Filling gaps is where an agent earns its place, and where it can also do damage if you skip the setup. The output has to sound like your store, use your vocabulary, and never invent a fact about a product.
Benefits
Attributes from source data
Pulled from supplier sheets, spec documents, or the existing description, never guessed. What cannot be sourced gets flagged, not filled.
Consistent naming
One convention applied everywhere, so sorting and feed rules stop breaking on the exceptions.
Alt text that describes the image
Written from the actual picture rather than pasted from the product title, which is what most stores do.
Metadata that is not the description
Meta titles and descriptions written for a search result, which is a different job from a product page.
A do-not-touch list
Hero products and anything you have hand-tuned stay excluded until you say otherwise.
The never-invent rule is the important one. A confidently fabricated material or dimension is worse than a blank field, because a blank field is visibly missing while a wrong one gets shipped to a marketplace and comes back as a return. Write it into the tool rules so it binds every run, rather than saying it once in a conversation and hoping.
Attaching those constraints to the employee rather than to a single request is what separates a one-off cleanup from a maintained catalog. That is the whole idea behind hiring one at Sistava: the AI Employee keeps the conventions, the exclusion list, and the memory of what it changed last month, so the next drop of forty products arrives already conforming instead of starting the argument again. Cleanups that are not maintained decay back to where they started within about two quarters.
Keeping it clean, which is the actual win
A one-time cleanup feels great and is worth roughly one quarter. Catalogs decay because new products arrive under time pressure, exceptions get made, and the person who knew the convention leaves. The durable version of this is a standing job rather than a project.
The standing catalog job
- Weekly gap sweep — Run the same audit on anything added or edited since last week, and report what is incomplete rather than fixing it silently.
- Draft the fills for review — Descriptions, alt text, and attributes prepared but not published, so approving a batch takes minutes rather than hours.
- Approve the batch, then it writes — On a connected store the write itself always waits for your yes, so approving a reviewed batch is the publish step rather than a separate risk.
- Flag what it cannot source — Anything that needs a fact only you or a supplier has comes back as a question, never as a guess.
That loop is unglamorous and it is the difference between a catalog that is good this month and one that is good next year. It also makes every downstream thing cheaper: your feeds stop getting rejected, your filters start working, and the next person who searches your store finds the thing they were looking for. Running it on a schedule instead of by memory is one of the things the Sistava feature set exists for, since a recurring job that reports back beats a reminder you keep snoozing.
If your catalog has an unusual shape, made-to-order items, complex variants, technical specifications that matter, you can train a custom AI Employee on exactly those rules rather than accepting a generic enrichment tool's assumptions. That is usually the difference between output you ship and output you rewrite. If you want to judge the writing before you commit to anything, the free tools at Sistava will draft a product description or a batch of alt text for one item, which is enough to tell whether the voice is close or nowhere near.
FAQ
Can AI write product descriptions that do not sound generic?
Yes, but only if you give it your brand voice and a handful of approved examples rather than asking for descriptions in the abstract. Generic output is almost always a briefing failure. Give it the vocabulary you use, the things you never say, and three descriptions you are happy with, and the output converges quickly.
Will it make up product specifications?
It can, which is why the rule against it has to be attached to the tool rather than mentioned once in a conversation. Configure it to source attributes from supplier data, spec sheets, or existing copy, and to flag anything it cannot source instead of filling it. A blank field is visibly missing; a fabricated one ships to a marketplace and comes back as a return.
How long does a catalog cleanup take?
The audit is fast, usually the same day. Enrichment depends on volume and on how much you review, but the realistic shape is a couple of weeks of reviewing batches for a few thousand products. The bigger commitment is the weekly maintenance afterwards, which is what stops it decaying back.
What if I have hand-written descriptions I do not want touched?
Put them on an exclusion list before anything starts. Hero products, anything you have tuned for search, and seasonal copy usually belong there. A good setup treats that list as permanent and only removes items from it when you say so.
Does this work if my products are in a spreadsheet, not a platform?
Yes. Catalog data lives in spreadsheets more often than vendors like to admit, and an agent that can read and write files handles that directly. It can also work through a platform admin screen or a supplier portal that has no API, using browser or desktop control, which is where a lot of the messier catalog work actually lives.
Is this the same as a product information management tool?
Different layer. A PIM is where structured product data lives and is enforced. An agent is what fills it, checks it, and keeps it current. If you have a PIM, this feeds it. If you do not, this is usually a cheaper first step than buying one, because most stores' problem is missing data rather than missing structure.
Catalog work never feels like the priority because nothing about it is visible from the outside. It is also the thing most likely to be quietly costing you sales right now, and unlike a redesign, you can prove the fix worked by searching your own store afterwards and finding the product you were looking for.