Stories / Case Study

Contextual signals, bought by an agent

How a buyer's AI agent describes a campaign in plain language and activates a contextual segment on DSPs/SSPs.

FM Florian Mühlhans · November 16, 2026
Back to Stories

This post covers the main content of the presentation we gave at the AAO DACH Chapter Monthly Meeting on 12 November 2026. Thanks again to David Porzelt and Philipp Wolde for giving us the opportunity to talk about our agent.

What a signals agent is

In the Ad Context Protocol (AdCP), a signal is a targeting segment: something a buyer attaches to a campaign to decide where the ad runs. A signals agent is a data provider that is used by another AI agent, not through a website or a sales contact.

Three parties are involved. The buyer agent plans a campaign and describes what it should reach. The signals agent turns that description into a segment. The third party is the place where the media is bought: a platform, meaning a DSP or SSP, or the seller’s own agent.

Our signals are contextual. A segment is a list of pages whose content fits the topic. It contains no user profiles. That means no cookies are needed and no user tracking is required.

What works today

The whole flow is four tool calls. We ran it against production with one example, a campaign for people planning a hiking holiday in the Alps.

  1. Describe: The buyer agent sends a brief in plain language: "Find signals for outdoor enthusiasts planning a hiking holiday in the Alps". About ten seconds later it has three drafts: a narrow one, a normal one and a broad one, each with a name, a description and the platforms it can be activated on. Nothing is created at this point.
  2. Refine: The first drafts were close, but the campaign was about the Dolomites and about summer. So the buyer agent sends the drafts back with one sentence: "Narrow it to hut-to-hut trekking in the Dolomites and leave out skiing and winter sports". It gets three new drafts built from the first ones. The narrow one is now called "Dolomites Hut-to-Hut Trekking Planning", and its description states that skiing and winter sports content is excluded. This step can be repeated as often as needed.
  3. Activate: The buyer agent picks one draft and one platform. Only now do we create the segment in the True Relevance platform and register it on the platform the buyer chose. In our example that was Index Exchange, and the answer came back after six seconds, together with the segment ID the buyer targets in their deal. The same call works for the other platforms we support, PubMatic and The Trade Desk among them.

Where the segment lands depends on the buyer's login. A buyer with their own seat on the platform gets the segment in that seat. Everyone else gets it on the open market.

  1. Report: A signals agent does not see the campaign. The ad is served on the platform. So the buyer side reports back how many impressions used the segment, and the agent records it per account, segment and period. Today we use these reports for billing. The next step is to use them for delivery as well: when a segment falls behind its goal we widen it, and when it runs ahead we narrow it, the way we already do for campaigns in our own platform. This part is planned, not built.

What happens behind the agent

The language model has one narrow job: it translates the brief into the structured segment our platform understands. It does not rank pages and it does not buy anything. Our agent checks the buyer's login, validates the drafts and runs the activation. The targeting work is done by the True Relevance platform, which ranks the pages of our content database for the segment and delivers the result to the destination.

The agent is public. It is listed in the AdCP registry at AgenticAdvertising.org, which checks every registered agent regularly against the protocol's compliance suite. Buyers sign in with OAuth.

Next: straight to the seller, without a DSP

Today every one of our destinations is a platform. AdCP allows a second kind of destination: the seller's own agent. A publisher or a sales house runs a sales agent, and a buyer agent books media there directly.

No DSP sits in between. For contextual data this is a natural fit. Our segment is a ranked set of pages, and a publisher knows its own pages. There is no user data to match and no ID to sync.

We have not built this yet. Our agent supports platforms as destinations and refuses anything else. What is missing is the delivery to a sales agent. We are in talks with the first companies that run sales agents, and we are actively looking for more to partner with.

We are looking for partners to test with

We would like to hear from:

  • Buyer-side teams, such as agencies, advertisers and companies building buyer agents, who want to create real campaigns with our signals and tell us where the flow breaks.
  • Publishers, sales houses and companies building sales agents, who want to receive contextual segments directly and work out the seller integration with us.
  • Anyone else interested in the technology, who can imagine using our agent in their own workflow.

If that is you, write to us at [email protected], or directly at [email protected].

Try it

Any AdCP or MCP client can connect to the agent, including other LLM tools such as ChatGPT. In Claude Code it is one command:

claude mcp add --scope user --transport http tr-adcp https://true-relevance-api.com/api/adcp/mcp

In ChatGPT you add the address https://true-relevance-api.com/api/adcp/mcp as a custom MCP server.

Describe the campaign you have in mind, and you can create open market segments right away for first tests. A login for segments in your own seat has to be set up on our side; write to us for that. You can also find the agent in the AdCP registry under Signals, as True Relevance Agent.

Florian Mühlhans