A buyer might learn about your company from a sales email, ask an AI agent whether your product fits their needs, and return to the same conversation weeks later to work through objections. The agent’s judgment can shape the sale at each of these moments. That makes agents an audience your company needs to understand and persuade.
Agent marketing is the practice of changing AI’s behavior and beliefs to support your business goals. At Unusual, we act as an agent marketing team for clients and build tools to support this work. In the process, we’ve learned a lot about what it means to run this function. Three responsibilities account for most of it.
Understanding how agents perceive your company
Any good marketing effort starts with an understanding of the audience. With agents, we can investigate that audience in considerable detail: vary the buyer’s circumstances, ask the question again, and observe how the recommendation changes.
The starting point is deciding what you need agents to believe or do. Say your company is moving upmarket. You might need agents to include you in enterprise shortlists, understand your security capabilities, and prefer you to a competitor when those capabilities matter. These goals give us something specific to measure.
We generally begin with awareness, perception, and preference, building recurring surveys around the same divisions the company uses to understand its market—personas, use cases, product lines, and competitive sets. Each survey consists of a panel of prompts, ranging from simulated buyer questions to direct inquiries about the brand. Responses are graded by an LLM judge using a rubric we develop with the client.
Together, the prompts and rubric give us a view of how agents position the company across the situations we’ve tested. Repeating the survey lets us see where that position is changing. The scope and frequency depend on the budget, and on how much sampling we need to distinguish a meaningful change from ordinary variation.
Researching what would change their judgment
A problem in the survey gives us a place to look more closely. If agents recommend your product to small businesses but exclude it from enterprise evaluations, we can vary the requirements, competitors, and wording to find the circumstances in which that judgment holds. Looking at the answers alongside the sources agents retrieve helps us develop an explanation for what’s happening.
From there, an intervention has two parts: the information we think would change the agent’s mind, and a way to get that information in front of it.
One research technique we like is a counterfactual experiment. Suppose an agent dismisses your product because it believes a required integration is missing. We can supply evidence that the integration exists, disable further web search, and ask whether that changes its answer. A changed recommendation gives us a hypothesis worth pursuing. We still need to find out whether the evidence has the same effect when the agent encounters it during its own research.
That brings us to delivery. Mapping the searches and sources behind the problematic answers can reveal where the information needs to live: a documentation page, an FAQ, a comparison, or an industry publication. AEO and GEO techniques are useful here, helping us make the argument discoverable.
Finally, we test the intervention against prompts that reproduce the problem but weren’t used to develop the fix. Did the agent encounter the new evidence? Did its recommendation change? The answers determine what to do next. Evidence that goes unread calls for work on distribution; evidence that is read and discounted sends us back to the argument. Repeated measurements under comparable conditions help us assess the result, while accounting for other changes that could have affected it.
Building experiences for agents
Understanding an audience also means understanding how it wants to engage with your business. Agents evaluating a product may need detailed documentation, explicit limitations, pricing, and evidence they can inspect. A recommendation may depend on a small technical detail that a human visitor would happily leave to a sales call.
Serving this audience can also mean building new ways to interact. A brand-hosted agent could answer detailed questions about the product; a demo or trial endpoint could let another agent try the software on its user’s behalf. The agent marketer’s job is to identify where the experience breaks down and work with product, engineering, and web teams to make it useful.
Where the agent marketer fits
Spiritually, the agent marketer’s closest partner is product marketing. Both have a mandate across the buying journey. As product marketing defines who the product is for and why they should choose it, the agent marketer investigates how agents understand that argument, where they disagree, and what would change their judgment.
The relationship runs both ways. Research into agent perception can expose a missing proof point, a weakness in positioning, or a use case the company hasn’t considered. Sometimes the agent has misunderstood the product. Sometimes it has found a limitation the company needs to address. Understanding the difference is part of the job.