Particular Audience Particular Audience
Talk to us

The retail operating system

One engine for every retail decision.

Most retailers run four or five separate tools.

Search in one tool, pricing in another, ads bolted on top — nobody owns the order shoppers see, and margin leaks between them. The Unified Decision Engine governs relevance, inventory, margin and sponsored demand in one pass, on every surface, including the AI assistants now doing the shopping.

Start with one surface. Leave the rest of your stack untouched.

Need one thing first? Start with our AI product recommendation engine or AI site search.

Sponsored Products in AI Assistants: Ranking, Labelling, CPM and Measurement

How retailers run sponsored products inside their own AI assistant, ChatGPT app or Claude connector: relevance-first ranking, labelling, CPM and CPC pricing, pacing and holdout measurement.

Sponsored products in agentic and conversational AI

A retailer's sponsored products can appear inside its own AI assistant, ChatGPT app or Claude connector through Adaptive Transformer Search over MCP. Relevance decides whether a sponsored product is eligible; the bid only orders products that are already relevant. Each placement is labelled, priced on CPM, CPC, cost per sale or sponsorship, and measured with holdout tests.

Why the answer is the new results page

A search results page shows dozens of products. An assistant's answer usually shows three to five. With fewer slots, each one carries more of the shopper's attention, and a poor placement costs more trust. That makes relevance the price of entry and makes every slot more valuable to the advertiser.

Retail media in conversation is therefore a ranking problem first and an auction second. The retailer owns both, because the products come from its catalogue through its own search.

How a sponsored product gets into an answer

  • The shopper asks a question in the retailer's onsite assistant, its ChatGPT app or its Claude connector.
  • The assistant calls the ATS MCP server, which sends the request to Adaptive Transformer Search.
  • ATS selects relevant organic products and relevant sponsored products from live campaigns.
  • The retailer's rules decide the mix: how many sponsored slots, which positions and which categories are allowed.
  • Each result reaches the assistant with a sponsored mark, and the assistant states it in the answer.

Relevance decides eligibility. The bid decides the order only among products that are already relevant. The Retail Media AI Architecture whitepaper describes this decisioning layer between the language model and the catalogue, and the ATS whitepaper explains the search ranking behind it.

Labelling and trust

Every sponsored product is labelled in words, for example "Sponsored", next to the product. The assistant must never present a paid placement as its own neutral recommendation, and it should be able to say why a product appeared when asked.

Check each platform's current rules before launch. OpenAI's Apps SDK app submission guidelines and Anthropic's connectors directory FAQ set what a listed app may show. Local advertising disclosure rules still apply. These placements are the retailer's own retail media inside its own app; they are not ads sold by OpenAI or Anthropic.

Formats in conversation

  • Sponsored product in a list: a relevant sponsored item among the products the assistant suggests. Runs on Sponsored Products.
  • Sponsored answer to a search-style request: the shopper's words act as the keyword. Runs on Sponsored Search.
  • Sponsored swap: when an item is out of stock or does not fit, the assistant offers a relevant sponsored alternative, labelled as such.
  • Sponsored bundle item: a sponsored product inside a suggested bundle or basket. Runs with Bundles.
  • Follow-up retargeting: a shopper who searched in conversation sees relevant sponsored products later on site, with consent. Runs on Search Retargeting.
  • Brand guidance: a brand-sponsored explanation or comparison, clearly labelled, linked to that brand's products.

Pricing and auction mechanics

The retailer chooses how to sell conversational placements. Common options:

  • Cost per thousand impressions (CPM): the advertiser pays when the product is shown in an answer. Define an impression as a product shown and labelled in a delivered answer.
  • Cost per click (CPC): the advertiser pays when the shopper opens the product or asks about it.
  • Cost per sale: the advertiser pays when the shopper buys the product within the agreed window.
  • Flat fee for each conversation or sponsorship: fixed placement in a category or topic for a period.

Whatever the model, apply a relevance floor before the bid. Because an answer has few slots, set a cap on sponsored slots per answer and price slot scarcity into rate cards rather than adding more ads. Report conversational inventory separately from site inventory so advertisers can compare like with like.

Pacing and delivery

Conversational inventory depends on how many shoppers ask relevant questions, so delivery can fall short of plan. PA flags a likely shortfall by notifying AdOps and the advertiser. PA does not change bid floors, bids or spending caps; the retailer's AdOps team decides the response. If the retailer chooses to offer make-goods, that is its own ad business decision and cost.

Measurement

Define the events before launch:

  • Impression: a sponsored product shown and labelled in a delivered answer.
  • Click: the shopper opens the product, adds it or asks a follow-up about it.
  • Conversion: a purchase of the product, or the brand's range, within the agreed window.

Return on ad spend alone overstates value. Use holdout tests, in which a random group of conversations shows no sponsored products, to measure the true effect. See how to measure retail media incrementality and closed-loop attribution. Track shopper-side signals too, such as follow-up questions and abandoned conversations, so trust is measured alongside revenue.

What to build

  • The ATS MCP server with the retailer's PA credentials, and live sponsored products campaigns.
  • The retailer's own MCP servers for stock, basket, loyalty and orders. Checkout stays in the retailer's systems.
  • Labelling rules in the assistant's instructions and interface, plus consent rules for any retargeting.
  • Reporting that splits conversational from site inventory, and a holdout group.
  • A listing per platform: a ChatGPT app, a Claude connector or an onsite assistant. See retail MCP product search.

Example conversation (illustrative)

Shopper: "I need a waterproof jacket for hiking, under $150."

The assistant calls ATS search with a price filter. ATS returns four relevant jackets, one of them a sponsored product from a brand campaign. The assistant answers with the four, labels the sponsored one "Sponsored", and adds a link to the retailer's cart for each. The retailer's stock tool confirms which are available for pickup nearby. This is an illustrative flow, not a customer result.

Sources

How Particular Audience does it

Frequently asked questions

Do sponsored products have to be labelled in an AI assistant?

Yes. Label every sponsored product in words next to the product, and follow each platform's app rules and local advertising disclosure law.

Will sponsored products reduce shopper trust?

Irrelevant ones would. Requiring relevance before the bid, capping sponsored slots per answer and labelling clearly keep answers useful. Measure trust signals alongside revenue.

Can we sell conversational placements on CPM?

Yes. CPM, cost per click, cost per sale and flat sponsorships are all options. Define what an impression means in an answer before selling on CPM.

Which brands can buy these placements?

Any brand whose products the retailer sells and allows in retail media, under the retailer's own category and brand rules.

Are these ads sold by ChatGPT or Claude?

No. They are the retailer's own retail media inside its own app or assistant. The platforms' own advertising products are separate.

Does PA run checkout in the assistant?

No. ATS returns products and sponsored products. Basket, payment and orders stay in the retailer's systems.

Related guides

How-to guides · All guides