Ecommerce AI Assistant

How an ecommerce AI chatbot answers product questions

A product question moving through catalog retrieval to a cited answer
A product question moving through catalog retrieval to a cited answer

An ecommerce AI chatbot should answer from product and policy content selected by the store, not from a general guess about the item.

The workflow has four parts: capture the source, retrieve relevant passages, generate an answer within those passages, and show where the facts came from.

Follow one question through the system

Consider: "Can I use this pump with a 16-bar food-processing line?"

  1. The product page must state pressure and intended use.
  2. The page must be captured and processed into the dataset.
  3. Retrieval must select the matching specification rather than a similar model.
  4. The answer must preserve the 16-bar qualifier and any configuration condition.
  5. The citation must open the product page containing those facts.

If the page lists pressure but not compatibility with the shopper's exact system, the answer should distinguish the published fact from the missing decision.

Product questions need different source types

QuestionBest source
Size, material, or technical limitProduct detail or specification sheet
Product comparisonCurrent pages for each product
Return eligibilityStore policy
Shipping conditionDelivery policy
InstallationCurrent manual or guide
Order statusAccount or order system, not public catalog content

A public website assistant can cover the first five when the sources are available. It should not pretend to know the sixth.

Preserve variants and qualifiers

Test products with sizes, colors, pack quantities, regional offers, sale prices, and "from" pricing. An answer that combines a base product's description with a different variant's price is unsupported even if both facts appear on the page.

Seekdown assistant showing an answer and source cards with sample data
Seekdown assistant showing an answer and source cards with sample data

Open the citation and check that the exact variant or condition appears there.

Define missing-information behavior

When the selected sources do not state a product fact, the chatbot should say so and point to the relevant product page or contact path. A short limitation is more useful than a confident invented specification.

Keep a regression set for the store's most important questions and rerun it after catalog or policy changes.