Optimizing product cards for Uzbek e-commerce with AI
The product page is the only page where a purchase decision is made. AI helps get thousands of them into shape, but it does not set the priorities.
Start with attributes, not copy
The typical mistake is ordering description generation immediately while the specification fields stay empty. Yet specifications are what answer most buyer questions and feed the catalogue filters.
Step one is normalizing attributes. AI is genuinely effective here: reconciling supplier chaos where size appears as '42', '42 EU' and 'forty-two' into one field with units.
Step two is filling gaps from existing data: extracting material and colour from the product name or the supplier's description. That is an extraction task where the model is reliable, unlike guessing.
Title and page structure
The page title must contain what people actually type: product type, brand, model, key characteristic. The formula brand-model-type-characteristic covers most commercial queries without keyword stuffing.
Block order on mobile: photos, price and availability, order button, short description, specifications, delivery and payment, reviews. Delivery terms for the visitor's region must be visible without scrolling to the cart.
The model is useful for generating titles and meta descriptions from a template with attribute substitution in two languages. Manual proofreading is only needed for top SKUs that carry the bulk of traffic.
What is specific to this market
Availability and delivery time to a specific region matter more than a detailed description. A buyer in Fergana first wants to know whether the item will arrive and when, and only then reads about the material.
Show payment methods on the page explicitly: Payme, Click, Uzcard, Humo, cash on delivery, instalments. A missing familiar payment method is a common reason for exit at the point where you already paid for the acquisition.
Write prices in UZS as digits with thousand separators, without shorthand like '1.2M'. Shorthand reads poorly on mobile and is parsed worse by search engines and scrapers.
Photos and reviews
Images are the bulk of page weight. Convert to WebP, declare dimensions in markup, and lazy-load everything but the first frame. On 3G that is the difference between viewing the page and leaving.
AI is useful for generating alt text from name and attributes in two languages. It is cheap, speeds up image search indexing and satisfies accessibility requirements.
Reviews supply unique content and defuse objections. The model summarizes them well into a short 'what buyers mention most' block — but it must summarize real reviews; generating reviews is out of the question.
Internal linking and markup
'Similar products' and 'frequently bought with' blocks driven by semantic attribute similarity outperform manual selections and refresh automatically as the catalogue grows.
Product markup with offers, price, priceCurrency UZS, availability and aggregateRating is mandatory. Without it the item never reaches search's product blocks and is parsed worse by AI systems.
Keep markup honest about stock. An item flagged InStock but absent from the warehouse damages both search results and reputation.