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.

Retail MCP Product Search: ATS MCP Server and AI Shopping Apps

Connect Adaptive Transformer Search to ChatGPT, Claude and onsite assistants through PA's open-source MCP server, with retailer-controlled ranking and checkout.

Retail MCP product search with Adaptive Transformer Search

Particular Audience's open-source ATS MCP server lets an assistant search a retailer's catalogue through Adaptive Transformer Search, with relevant sponsored products returned and labelled in the same call. The retailer runs it with its own credentials, keeps ranking and retail media placement in ATS, and combines it with its own stock, basket and order servers to build shopping apps for ChatGPT, Claude or an onsite assistant.

The ATS MCP server

Particular Audience publishes an open-source Adaptive Transformer Search MCP server. A retailer runs it on its own infrastructure, with Python or Docker, and authenticates it with the client ID, shortcode and secret that PA issues for that retailer's Search API. It reads only that retailer's catalogue. It is separate from the public information server described on the connection page and from the DiscoveryOS connector, which is still planned.

The server exposes three read-only tools:

  • Search: natural-language product search with optional filters and pagination.
  • Filtered search: search with filters that must match, such as category, brand or size.
  • Sorted search: search with a chosen sort order, such as price or popularity.

Results come from Adaptive Transformer Search, so the retailer's boosts, merchandising rules and sponsored products carry into the assistant's answers in the same way they appear on site. The Adaptive Transformer Search whitepaper explains the vector, vertical-tuning and reinforcement-learning approach behind the ranking.

Retail media inside assistant answers

Retail media is a first-class part of the ATS MCP server, not an add-on. Sponsored products come back through the same ATS search call as organic results, so an assistant does not need a separate ad integration.

  • Relevance first: a sponsored product is eligible only when it matches the shopper's request. A bid cannot place an irrelevant product in an answer.
  • Marked in the data: each sponsored result arrives marked as sponsored, so the assistant can label it in words.
  • Retailer rules: the retailer sets how many sponsored products an answer may contain and where they rank against organic results.
  • One campaign, every surface: the same Sponsored Products and Sponsored Search campaigns run on site, in the onsite assistant and in the retailer's ChatGPT or Claude app.

The guide to sponsored products in AI assistants covers ranking, labelling, pricing, pacing and measurement in depth.

Build a shopping app for ChatGPT, Claude and other assistants

A rich shopping app usually combines several MCP servers behind one assistant:

  • Product discovery: the ATS MCP server returns relevant products, including sponsored ones that match the shopper's intent.
  • Retailer systems: the retailer's own servers handle stock, store pickup, loyalty, basket and orders.
  • Planning and information: PA's public MCP server answers questions about PA products, case studies and guides. It does not read retailer data.

The same set of servers can power an onsite chat assistant and an app listed in an assistant's directory. OpenAI's Apps SDK builds ChatGPT apps on MCP, and apps are submitted for review under its submission guidelines. Anthropic lists remote MCP servers in its connectors directory, and the directory FAQ explains its review. Other assistants that accept remote MCP connectors follow a similar pattern. Each listing belongs to the retailer and must pass that platform's review. PA works with the retailer on the integration; a listing is not an off-the-shelf PA product.

The retailer keeps control. Ranking, boosts and sponsored placement stay in ATS. Shopper sign-in, payment and checkout stay in the retailer's own systems. PA does not complete purchases on a shopper's behalf.

An illustrative flow: a shopper asks for "comfortable walking shoes under $100". The assistant calls ATS search with a price filter, then calls the retailer's stock tool for the nearest store, and answers with products that link to the retailer's cart.

Define the smallest useful read contract

Specify the allowed catalogue, store or market scope and the permitted query fields. Identify which constraints are hard requirements and which are ranking preferences. State how the response represents an unknown attribute, an unavailable item and an empty result.

For each returned offer, require a stable identity, the relevant variant, merchant destination, currency, price context and availability timestamp when those fields are supported. Separate a product match from a current offer. A product can be relevant while an offer is not eligible for the shopper’s region or requested fulfilment option.

Establish authority and isolation

Document the caller, resource audience and permitted scope. The MCP security guidance requires careful token handling; do not pass a token intended for one resource through to another as a shortcut.

Test that an identity allowed to query one retailer cannot read another retailer’s catalogue or operational data. Keep product content as data, even when a description contains instructions. Define logging that supports diagnosis without retaining unnecessary customer information or credentials.

Test useful failure behaviour

A synthetic request asks for a dishwasher-safe mug in a particular market. A candidate with an unknown dishwasher attribute must remain unknown; it must not satisfy a strict requirement merely because its name looks suitable. A stale offer should trigger the documented refresh or unavailable path rather than a fabricated current price.

Add fixtures for unsupported filters, empty results, expired credentials, denied scope, timeouts and malformed upstream responses. Document retry rules and distinguish a safe repeatable read from any operation that could change a basket or order. Do not add purchasing actions to a read-only design by implication.

Agree evidence before launch

Ask the implementation owner for the actual contract, deployment status, supported version and test results. Validate ranking quality separately from tool connectivity, and check latency against the agreed requirement rather than an invented public SLA. Only describe the integration as live after the intended environment and permissions have been demonstrated.

Sources

How Particular Audience does it

Frequently asked questions

Is the ATS MCP server live?

The server is published as open source and connects to the PA Search API with credentials issued to each retailer. Whether it is deployed for your catalogue depends on your PA agreement and setup; confirm the environment before describing it to shoppers as live.

Can the public PA MCP connector search our live retailer catalogue?

No. The public server only answers questions about PA. Catalogue search uses the ATS MCP server with your own credentials.

Does a product-search tool also support checkout?

No. The ATS tools are read-only. Basket, payment and orders stay in your own systems, each with its own authorisation and confirmation.

Can we list a shopping app in ChatGPT or Claude?

Yes, as your own app or connector, submitted under each platform's review rules. ATS can provide the product discovery inside it alongside your own MCP servers.

Related guides

  • Sponsored products in agentic and conversational AI: 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.
  • What is agentic commerce?: Understand agentic commerce across discovery, comparison and authorised action. Map product data, permissions, checkout boundaries and useful retail measures.
  • ACP, UCP and MCP: what retailers should compare: Compare ACP, UCP and MCP by the job each performs. Separate commerce capabilities, tool connections, platform access and retailer implementation requirements.

How-to guides · All guides