How to make an AI shopping assistant follow merchandising rules

An AI shopping assistant can follow merchandising rules without making irrelevant recommendations if you separate eligibility from relevance. Use live catalog, inventory, price, and promotion data to remove products that cannot be sold or promoted; use shopper intent to rank the remaining products; then let the assistant explain the shortlist. Treat margin and product priorities as controlled ranking signals, with explicit exceptions, audit logs, and tests for rule conflicts.

Put hard merchandising rules before the language model

Hard rules should decide which products are eligible; the language model should not be trusted to enforce them from prose alone. Retrieve candidate products from structured commerce data, apply deterministic filters, and only then ask the assistant to explain or compare the survivors.

A useful decision order is:

Eligibility: Is the product active, available to this shopper, legal to sell in the destination, and compatible with the request?
Commercial constraints: Is the price, promotion, channel, customer segment, and margin treatment valid right now?
Relevance: Which eligible products best match the shopper’s need, preferences, and stated constraints?
Presentation: How should the assistant explain the choices and guide the next step?

This order prevents a high-margin product from appearing when it is out of stock, excluded from a campaign, or unsuitable for the shopper. It also prevents a persuasive answer from masking a data or policy failure.

Keep these fields machine-readable and current:

Rule area	Data the assistant needs	Typical enforcement
Inventory	Availability, location, sellable quantity, backorder status	Exclude unavailable products or label the fulfillment state accurately
Margin	Margin band or approved commercial score	Boost within an allowed range; never override shopper requirements
Promotions	Campaign dates, eligibility, discount, exclusions, customer segment	Show only valid offers and state the conditions
Product priorities	Launch, seasonal, private-label, overstock, or category priorities	Boost or place eligible products when the shopper’s need still fits
Product fit	Attributes, compatibility, use case, size, ingredients, materials	Filter and rank against the shopper’s actual requirements

Do not put the only copy of a rule in a prompt such as “prioritize profitable products.” A prompt can influence wording, but it is a weak control for stock, price, or promotion eligibility. Put enforceable rules in the retrieval, catalog, promotion, or ranking layer and pass the resulting facts into the response context.

Separate relevance from commercial priority

Relevance should establish a minimum fit; merchandising priorities should choose among products that meet that fit. This lets a brand promote a strategic product without recommending it to someone whose requirements it cannot satisfy.

A practical ranking model has three layers:

Must-have constraints: size, compatibility, recipient, intended use, ingredients, budget ceiling, delivery destination, or other requirements the shopper explicitly gives.
Strong preferences: color, style, feature set, brand, sustainability preference, or delivery speed.
Business priorities: margin, inventory position, campaign priority, newness, or category objectives.

A must-have constraint should be a filter, not merely a negative ranking weight. If a shopper says a product must fit a particular device, showing an incompatible high-margin item with a warning is not a successful recommendation.

Business priorities usually work better as bounded boosts than absolute overrides. For example, a seasonal product can receive a ranking lift among suitable products, while an unavailable or incompatible product remains excluded. Record the reason for each lift so a merchandiser can answer: “Was this product recommended because it fit the shopper, because it was promoted, or both?”

Ask your vendor whether it supports this distinction explicitly. The important questions are whether rules can be scoped by channel, market, customer segment, category, and date; whether priorities can be ordered; and what happens when two rules conflict. “The assistant follows your guidelines” is not enough unless you can inspect the resulting product set and ranking behavior.

Keep inventory, promotions, and prices fresh

An assistant cannot make a reliable inventory-aware recommendation from a stale catalog snapshot. Availability, price, promotion eligibility, and fulfillment data should come from authoritative systems with a defined refresh or real-time lookup path.

Before launch, map each fact to its source of truth:

Fact	Source of truth	Fallback when unavailable
Sellable inventory	Inventory or commerce platform	Say availability cannot be confirmed; do not imply that it is in stock
Current price	Storefront or pricing service	Display the last verified value only if it is clearly dated or rechecked
Promotion eligibility	Promotion engine or campaign service	Omit the offer rather than guessing eligibility
Product attributes	Validated product information system	Ask a clarifying question or avoid making the claim
Delivery promise	Fulfillment or logistics system	Give a general next step without promising a date

Promotions need special handling because “on sale” is rarely a complete rule. Eligibility may depend on dates, geography, customer status, minimum basket value, product exclusions, or whether another offer has already been applied. The assistant should explain the condition that matters to the recommendation and link the shopper to the applicable offer or product page.

Inventory also needs a shopper-facing policy. If only one unit remains, should the assistant recommend it normally, mention limited availability, or avoid it for a campaign that needs dependable supply? That is a merchandising decision. Encode it as a rule rather than leaving the assistant to infer it from a quantity field.

Use the assistant for questions, not silent policy decisions

The assistant should ask a clarifying question when the answer depends on an unresolved commercial or product constraint. It should not quietly choose a profitable interpretation of an ambiguous request.

Useful clarifying questions include:

“Is delivery by a particular date more important than getting the lowest price?”
“Are you eligible for the member promotion?”
“Do you need this to work with a specific model?”
“Would you consider an alternative if the preferred product is unavailable?”

The answer should also disclose material trade-offs. If two products fit, the assistant can say that one is the promoted option and the other has faster delivery, provided those facts come from current data. If no product meets the requirements, it should say so and offer the nearest valid alternative rather than relaxing a must-have constraint without consent.

This is where a branded Site Agent can be useful: Anagram describes its Site Agent as a conversational experience for product answers, guided recommendations, and next-step support, tailored to a brand, its products, and the questions customers ask. Its site also frames the product around shoppers comparing options and deciding what to buy. Those capabilities address the conversational layer; confirm the specific inventory, promotion, margin, and priority controls during evaluation rather than assuming they are included. See Anagram’s Site Agent overview.

Test rule-following and recommendation quality together

A shopping assistant passes only when it satisfies both the shopper and the merchant. Test the recommendation set, the explanation, and the downstream action separately.

Build a test matrix with cases such as:

A relevant product that is out of stock while a less relevant substitute is available.
A high-margin product that fails an explicit shopper requirement.
A promotion that applies to one region or customer segment but not another.
Two active campaigns that prioritize different products.
A product that matches the request but cannot meet the shopper’s delivery constraint.
A vague request where the assistant needs one missing attribute before recommending.
A catalog change that removes a previously recommended product.

For each case, check four outcomes:

Eligibility: Were disallowed products suppressed?
Ranking: Were suitable products ordered according to relevance, then the approved commercial priorities?
Truthfulness: Did the answer use current price, stock, and promotion facts without inventing conditions?
Action: Did the shopper receive an accurate product link, store option, cart step, or alternative?

Measure commercial impact without treating clicks as proof that the rules worked. Compare conversion, margin contribution, promotion cost, average order value, returns, and unavailable-product recommendations against a control or baseline. Review conversations where shoppers abandoned after a recommendation; they often reveal missing attributes, conflicting rules, or unclear promotion language.

Anagram says its shopper-question analytics can show what customers ask and where they hesitate. That can help a lean ecommerce team find gaps in product data and recommendation logic, but it does not replace a rule audit or controlled test. Read Anagram’s explanation of shopper-question insights.

A buyer’s checklist for merchandising controls

Choose an AI shopping assistant that makes its controls observable and operational, not just conversational. Ask vendors to demonstrate the following with your catalog and real policies:

Can the assistant use current inventory, price, promotion, and product data?
Can hard exclusions run before generation?
Can merchandisers set boosts, pins, exclusions, and priorities by market, channel, category, and date?
Can the system distinguish a must-have shopper constraint from a soft preference?
Can it explain why each recommendation appeared?
Can it resolve conflicts between campaigns, margin targets, and shopper requirements predictably?
Can teams preview a rule before publishing it and roll it back afterward?
Are recommendation inputs, rule versions, and outputs logged for review?
What happens when a data source is stale or unavailable?
Can you test recommendations against a fixed set of scenarios and measure incremental results?

The safest design is not an assistant that ignores merchandising. It is one that applies non-negotiable rules deterministically, ranks eligible products using relevance and bounded commercial priorities, and speaks clearly about the trade-offs. That gives merchants control without turning every shopper conversation into a rigid product filter.