Ad serving within owned chat experiences: A technical overview

Six major retailers have recently launched AI shopping assistants with advertising embedded in the experience. At the same time, more than three-quarters of advertisers report they are likely to invest in agentic commerce experiences including paid AI visibility, sponsored recommendations, feed integrations, and agent-specific measurement. 

Commerce media networks (CMNs) hoping to capture this demand by launching similar programs have two problems to solve:

1. How do I integrate ad serving technology within LLM shopping experiences?
2. How do I commercialize this for my advertisers?

Today, we’ll address the first. This technical guide addresses how to launch in-chat ad serving that prioritizes relevance and low latency while leveraging memory to understand the full consumer journey beyond a single moment in time. Realizing this opportunity requires rethinking how an ad request is created, when an auction should happen, and how sponsored listings fit into the conversation. 

Scoping your approach

This guide is intended to help commerce media networks think through the technical architecture and ad serving logic required to support advertising within conversational experiences. Your team will still need to make several important decisions, especially around:

1. Your fundamental chat experience: What role should the experience play for shoppers, and how should it work?
2. Whether ads should appear at all: Where, when, and under what circumstances does advertising add value to the experience?
3. Your monetization strategy: What types of sponsored experiences and advertiser offerings make sense?
4. How ads inform your learning process: How can advertising generate signals that help you understand shopper behavior, evaluate the experience, and refine your approach?

This guide is a planning framework, not a prescriptive roadmap. It outlines the technical foundation for conversational ad serving while leaving the experience, monetization model, and broader strategy open to iteration.

A practical architecture for conversational ad serving

Traditional sponsored search can operate relatively simply: Query → targeting signals → auction → sponsored result.

But conversational ad serving adds several layers:

Conversation + memory → intent and monetization decision → entity feed + taxonomy → candidate retrieval → relevance and eligibility → auction → sponsored recommendation → next turn updates intent

Each stage introduces new ad serving considerations. The system needs reliable entity data, conversational context, logic to determine when the conversation represents an advertising opportunity, and a retrieval and ranking system that prioritizes relevance.

Consideration #1: Entity feed and infrastructure 

Conversational shopping can often provide rich context for agents, but less explicit guardrails for product selection. Instead of asking for “waterproof hiking boots, size 11,” a shopper may say, “I’m going hiking in Montana next month. What do I need?”

This opens the door to new, relevant product promotions for the consumer. The agent can infer that the shopper may need hiking boots, outerwear, water storage, sun protection, or other gear based on the trip itself. But supporting that kind of request requires an entity feed that the agent can reason against, allowing agents to turn broad shopper goals into relevant product candidates and, ultimately, eligible sponsored opportunities.

Entity feeds therefore need rich, machine-readable attributes including:

• Category and taxonomy
• Brand
• Entity attributes
• Price
• Inventory
• Location availability
• Fulfillment options
• SKU/product ID

TL;DR: entity feeds are the foundational infrastructure for conversational ad serving. If the agent understands the shopper perfectly but the feed lacks attributes, inventory, or usable taxonomy, there is still no reliable way to connect that intent to an eligible product.

There are three primary ways to turn product data into an input for ad serving.

1. Conversation → semantic categories → ad server: The LLM maps the conversation to one or more semantic categories that can be passed to the ad server. This approach is generally easier to constrain and evaluate, aligns closely with existing auction architecture, and requires less complexity from the agent itself. It hinges on a well-structured categorization framework that can account for multiple relevant categories within a conversation.
2. Conversation → products/SKUs → ad server: The agent searches the feed and identifies relevant entities directly. This allows for more precise matching and rich use of conversational context, but requires strong LLM controls, sophisticated evaluation, and clear relevance guardrails.
3. Conversation → categories + SKUs → ad server: The agent identifies both relevant categories and candidate products before initiating the auction. This provides delivery with multiple signals and offers greater resilience when either category classification or SKU matching is imperfect.

Consideration #2: Identify the right ad opportunities within the conversation

True relevancy accounts for both what a user is looking for and their stage in the buying journey. Ad serving technology needs to weigh ad viability against the organic experience based on customer intent. Here’s how that breaks down:

These distinctions matter: 45% expect their heaviest investment in AI agent-mediated environments to focus on lower-funnel conversion and sales. At the same time, there’s room to monetize upper-funnel experiences, though it means thinking creatively about your advertiser offerings.

Caveat: one query doesn’t mean only one advertising opportunity

For example, the query: “Help me make tacos for six people” results in multiple ingredients and, therefore, multiple ad opportunities. Rather than treating the response as a single opportunity, conversational commerce treats each relevant product need as its own commercial decision.

For each ingredient: Identify product need → find organic candidates → identify eligible sponsored products → run auction → rank results

The system needs to identify which parts of a shopper’s task represent independent commercial opportunities and determine whether each warrants an auction.

Consideration #3: Build the full conversation into the ad request

Now more than ever, ad serving must account for consumer history. Take, for example, the chat:

• Message 1: “I need an Italian recipe for dinner tonight. I have pasta, tomatoes, and breadcrumbs: what goes with this, and add it to cart.”
• Message 2: “My kids don’t eat red meat.”
• Message 3: “I’m allergic to fish.”

The system now has an evolving picture of the shopper’s needs, constraints, and likelihood to purchase. The ad request therefore needs to represent that accumulated intent. To do so, ad servers must:

• Use relevant context from across the conversation
• Continuously refine the shopper’s intent as new information appears
• Translate conversational language into structured signals 
• Drop outdated constraints when the user changes direction
• Ask for clarification when the system cannot confidently determine intent
• Determine when and where to surface a product and if multiple ad calls are needed for multiple products

The key change is that the latest chat message is not necessarily the ad request. The accumulated intent state is.

• Turn 1: “I need an Italian recipe for dinner.”
  • Meal: Italian
  • Goal: Dinner planning
• Turn 2: “I have pasta, tomatoes, and breadcrumbs.”
  • Meal: Italian
  • Goal: Dinner planning
  • Existing ingredients: Pasta, tomatoes, breadcrumbs
• Turn 3: “My kids don’t eat red meat, and I’m allergic to fish.”
  • Meal: Italian
  • Goal: Dinner planning
  • Existing ingredients: Pasta, tomatoes, breadcrumbs
  • Exclude: Red meat, fish

That accumulated context now affects both organic recommendations and which sponsored products should be eligible. There are two primary ways to maintain that memory.

1. Rolling conversation context: Pass a summary of relevant previous turns into the ad-serving workflow.
2. Structured intent state: Maintain structured fields that update throughout the session.

A structured approach can make it easier to translate conversational intent into deterministic auction signals because the ad server receives known categories, attributes, exclusions, and constraints rather than an entire transcript. Additionally, memory needs guardrails. Previous conversation should refine targeting only when the system is sufficiently confident that the information remains relevant. If a shopper changes direction, the intent state needs to change with them.

Consideration #4: Keep relevancy central

Once the system understands the shopper and identifies a potential advertising opportunity, it still needs to decide which products should be allowed into the auction. Conversational advertising has to begin with the user’s goal. Ad serving therefore needs two distinct decisions.

1. Determine eligibility: Does this product actually satisfy the shopper’s request? This is based on key search and entity criteria, as well as user-stated constraints. This is where the retrieval approaches established in the product foundation come into play. Semantic matching can help determine whether the product is relevant to the shopper’s actual language, while structured attributes and taxonomy provide hard constraints. Inventory and location can act as further eligibility gates.
2. Rank eligible sponsored products: Once products clear the relevance threshold, commercial signals can determine ranking. Those signals might include:
   • Bid
   • Quality Score
   • Predicted CTR
   • Predicted conversion
   • Advertiser eligibility
   • Expected commercial value

This reinforces that eligibility is based on conversational relevance and commercial ranking, not just the highest advertiser offer. Recommendations inside an agent carry more weight than a traditional placement, especially because, within a chat experience, often only one product will win. A sponsored result that fails to meet the shopper’s stated constraints risks degrading both the recommendation and the broader chat experience.

Consideration #5: Speed is everything

The architecture described so far introduces considerably more decisioning into the response path. All of that needs to happen quickly enough that advertising does not slow the chat experience. Here are paths for ensuring low latency:

1. Parallelize ad retrieval with response generation where possible: The system does not necessarily need to complete each operation sequentially. Depending on the response flow, intent processing, candidate retrieval, and deterministic lookups may run alongside response generation, while retrievals that depend on the response should remain sequential.
2. Maintain intent state: Earlier turns can be summarized or converted into structured fields as they happen. The platform should not need to reprocess the entire conversation every time the shopper sends a new message.
3. Cache common mappings: Frequently used category mappings, product attributes, and candidate sets can be cached where appropriate rather than recreated on every interaction.
4. Use staged retrieval: Rather than running expensive matching against every product, the system can progressively narrow the candidate pool.
5. Avoid unnecessary auction calls: The intent decision from Consideration #2 also becomes a latency tool. If a turn does not meet the threshold for monetization, there is no reason to run a full sponsored-product auction.

Chat may offer some flexibility because retrieval can happen while the LLM is already generating or preparing a response, but that flexibility should be used to support richer relevance rather than accepting slower performance.

Consideration #6: Protect trust 

Consumers are open to monetization inside AI experiences, but that openness has clear boundaries. Koddi research found that 56% of consumers consider it acceptable, or acceptable if clearly labeled, for brands to pay for shortlist inclusion in agentic tools. At the same time, 72% want AI to act as a co-pilot rather than a full autopilot.

The opportunity is to make commercial influence transparent while ensuring the recommendation still contributes to what the shopper is trying to accomplish.

Bringing it together: A conversational ad request in practice

Taken together, conversational ad serving could follow this flow:

1. Establish the entity foundation: Ensure structured data is available to both the agent and delivery system.
2. Understand the conversation: Extract relevant signals from the chat.
3. Maintain the intent state: Combine the latest turn with relevant context from earlier in the conversation and update the structured representation of what the shopper currently wants.
4. Determine whether the moment is monetizable: Assess commercial intent, confidence, suitability for advertising, and whether one or multiple product opportunities exist.
5. Retrieve candidates: Use category search, semantic search, bidder/SKU lookup, or a combination.
6. Apply relevance and eligibility thresholds: Remove products that do not sufficiently satisfy the user’s request, including products that fail inventory, price, location, or attribute constraints.
7. Run the auction: Rank eligible sponsored candidates using bid, Quality Score, predicted performance, and other commercial signals.
8. Integrate the winner into the response: Clearly distinguish sponsored recommendations while maintaining a natural conversational experience.
9. Continue learning from the conversation: Use the shopper’s next interaction to update intent and improve the next recommendation.

The architecture is cyclical: every new turn can change what the shopper wants, whether an advertising opportunity exists, which products are eligible, and which advertisers should compete for that opportunity.

From winning impressions to earning consideration

The ad-serving challenge within chat experiences becomes less about determining what ad belongs and more about understanding what the shopper is trying to accomplish, what the ad server knows about their intent, when to serve a recommendation, and what products satisfy their needs.

For commerce media networks, capturing this opportunity starts with infrastructure: clean product data, persistent conversational context, clear monetization logic, strong relevance thresholds, fast retrieval, and transparent sponsored experiences. The networks that can turn conversational context into relevant, trustworthy auction signals will be positioned to move commerce media from winning impressions to earning consideration inside the decision itself.