Which data the bot can see
For connected listings, Vidkov uses the title, description, stated price, attributes, photos, and additional seller information. The history of the specific chat is supplied separately. Together, these sources let the system distinguish a new question from a continuation of an earlier one.
The presence of a model in the catalog does not mean the buyer is asking about it. If the buyer names a different item, that name remains a separate subject. General product properties can be verified through public sources, while availability, ordering, and seller promises can come only from seller data.
Why full context matters more than the last sentence
The words “what about this one?” explain nothing without earlier messages. A sentence containing a new product name, by contrast, can change the entire subject. The system can process recent messages from newest to oldest for speed while preserving their real chronological order and the role of every participant.
A previous automated reply is not evidence. If the bot made a mistake, the next turn must return to the buyer’s question and approved sources instead of defending the earlier error.
How to avoid repetition
The buyer can see earlier messages. A new reply should not list the processor, display, battery, condition, and package again when the buyer asked about one different property. A stated fact should return only when the buyer asks again, disputes it, or the new conclusion cannot be explained honestly without it.
What improves accuracy
- One item or one clearly separated offer in each listing.
- A current price and condition with no conflict between title and description.
- Complete photos of labels, included items, and cosmetic condition.
- Additional information containing only facts the seller is ready to confirm.
- A review of the first conversations in confirmation mode.
How to prepare listings for accurate answers
The model sees more than an abstract product category. It receives the seller's available active listings, the current conversation, additional notes, and public sources when appropriate. Answer quality depends on whether these sources agree. A useful product card does not have to be long, but it must contain the facts a buyer cannot safely infer on their own.
Which fields matter most?
Use an exact name, actual price, condition, included items, essential dimensions or compatibility, available delivery methods, and any material feature of this particular unit. General promotional language cannot replace specifics. If one attribute regularly decides a purchase, put it in the description or additional information.
How does multi-message context work?
A buyer may name a model once and then write only “what colour?”, “any others?”, or “that will not work.” Full history links the follow-up to the right subject. The reply does not need to repeat the name every time; it should reintroduce it only when omission would create genuine ambiguity.
How should questions about other products be answered?
When a buyer requests other models, sizes, or variants, the reply should inspect the seller's full active assortment, select every relevant position, and name each one clearly. “There are others” does not help a buyer choose. The current item must not be presented as an alternative when the buyer explicitly asked for a different one.
What do photographs contribute?
Images support visible details such as colour, included accessories, wear, shape, or a printed marking. They cannot prove hidden properties that are not visible. If a photo and text conflict, the seller should correct the card rather than make the buyer investigate the discrepancy during the conversation.
When is web search useful?
Public sources are useful for general model specifications, official compatibility, or an explanation of technology. They do not determine the actual condition of a particular unit, the seller's stock, or willingness to grant a discount. Web research can complement a listing but cannot replace seller-owned information.
How do you keep similar items separate?
Retain distinctions buyers use to choose: generation, size, storage, material, or condition. If several products have the same headline, add a clear distinguishing fact in their description or note. The model can then associate a question with the correct unit rather than borrowing a specification from a neighbouring listing.
What most often damages context
- Several listings with identical names and no usable fact that distinguishes them during an ordinary buyer conversation.
- Different prices or conditions in the title, description, and extra note without explaining which value is current.
- A list of attractive specifications that omits practical answers about condition, included items, dimensions, or delivery.
- Expecting web search to determine a specific item's condition or a seller decision that no public source can know.
- Giving a general brand description instead of a precise list of available alternatives when the buyer requests other options.
Review every card as someone seeing the shop for the first time. The name should identify one position, the opening paragraph should explain its condition and practical value, and later blocks should answer ordinary purchase questions. Do not bury important facts inside a wall of keywords. For variable products, distinguish sizes, colours, or configurations available in this listing from those published separately. When a buyer moves between products, conversation history must support comparison without blending specifications. After a catalogue update, ask brief contextual questions such as “this one?”, “what else?”, and “anything cheaper?” Verify that the reply combines earlier context with the complete assortment. If one failure occurs across several chats, remove the conflicting source data. Clear cards enable shorter answers, stronger reasoning, and natural subject economy because the model does not need to repeat what the buyer already understands.
For a large catalogue, review a weekly sample from different categories. Compare each card with OLX, ask about a distinctive feature, and request alternatives. The answer should name current products without invented stock or redundant category wording. A sold or inactive item must disappear from suggestions, while a new one should appear after synchronization. This short routine checks catalogue freshness, fact scope, and semantic assortment search together.