How do ads work in LLM shopping assistants? | Particular Audience
In a retailer's LLM shopping assistant, the language model does not choose ads. It calls the retailer's product search, which returns relevant organic and.
How do ads work in LLM shopping assistants?
In a retailer's LLM shopping assistant, the language model does not choose ads. It calls the retailer's product search, which returns relevant organic and sponsored products with sponsored ones marked. The model then writes the answer and labels them.
Why the model should not pick ads
Language models are good at conversation, not at auctions, pacing or reporting. Keep ranking and ad decisions in a decisioning layer the retailer controls.
The flow
Request, product search with meaning scoring, relevance check, auction among relevant products, blended answer, label. See how sponsored products get into an answer.
What the retailer sees
Impressions, clicks and sales for each sponsored product in conversation, reported separately from site inventory.
Run it with Particular Audience
Sponsored Prompts runs on Adaptive Transformer Search and the open-source ATS MCP server. Advertisers pick products or a collection to target, while the retailer's AdOps team sets the budget, rate, dates, eligibility, approval and reporting; eligible items can then appear in chats, prompts and AI searches on the retailer's own assistants where they score highly, and the retailer's interface must show or speak the sponsored label, holding back any result that arrives without it.
What counts as a billable event
A product returned to the assistant is a candidate, not an impression. Bill only on events the retailer's own interface can confirm: a CPM impression is a sponsored product actually displayed (or spoken) and acknowledged by the client, de-duplicated, with invalid traffic removed; viewability and attention are reported separately if agreed. For voice, define started, completed and interrupted or replayed delivery, and agree which one is billed. A CPC click is a tap or click on the sponsored product or a qualifying add-to-cart; a follow-up question is conversation engagement, not a click, and is only billable if named in the agreement. Attributed ROAS on its own does not show incrementality. Impressions in assistants the retailer does not operate are not sold, because the host's delivery event is not available.
Sources
- Adaptive Transformer Search MCP server (GitHub), checked 2026-10-25
How Particular Audience does it
Frequently asked questions
Can the LLM invent a sponsored product?
It should only present products the search returned. Instruct it never to describe an unreturned product as sponsored or available.
Does this need fine-tuning the model?
No. The ads come from search results, not from the model.
Related guides
- Sponsored products in agentic and conversational AI: How retailers run sponsored products inside the AI assistants they operate: relevance-first ranking, labelling, CPM and CPC pricing, pacing and holdout measurement — and why ChatGPT and Claude apps are discovery-only.
- Retail MCP product search with Adaptive Transformer Search: Connect Adaptive Transformer Search to ChatGPT, Claude and onsite assistants through PA's open-source MCP server, with retailer-controlled ranking and checkout.
- What are sponsored prompts?: Sponsored prompts are a retail media format for AI conversations. When a shopper asks a retailer's assistant, voice service or AI search for something,.
- What is agentic retail media?: Agentic retail media is advertising that runs inside AI agents and assistants which act for shoppers: finding, comparing, reordering or swapping products..