Why Marketing's AI ROI Crisis Is Unfixable in the Agentic Era

Today, June 9, 2026, Gartner's Marketing Symposium/Xpo is hosting a panel on "Marketing's AI-Era ROI Crisis." Industry leaders will debate data management challenges, attribution models, and proof of impact. They'll walk out with framework slides and vendor partnerships. They'll miss the actual problem.

The ROI crisis isn't a measurement gap. It's a causality collapse.

When autonomous AI agents make purchasing decisions, shopping agents, recommendation engines, chatbot checkout flows, there is no clean causal path backward to a marketing decision, a campaign, or a creative asset. A customer bought something. An agent facilitated it. Your campaign may have touched the consideration funnel six months ago. Or it may be completely irrelevant. There's no way to know. And no amount of better attribution modeling fixes that.

Most marketing teams still operate under attribution assumptions from the Google Analytics era: touchpoints, funnels, paths-to-conversion. These models assume human decision-making at each step. A person sees an ad, gets curious, clicks a link, reads a review, adds to cart, checks out.

With agentic AI, the human part is gone. Or it's invisible. A person browses, an agent surfaces recommendations, the person buys. Or: a person offloads the entire purchase decision to an agent, which negotiates price, compares reviews, and completes the transaction. The person never consciously touched your campaign.

Current solutions pretend this doesn't matter. Upgrade your CDP. Layer in probabilistic modeling. Use incrementality tests. Attribution vendors are selling you sophistication as a substitute for causality. It doesn't work.

Why? Because causality requires a decision-maker. If the decision-maker is an autonomous agent, your campaign influenced the agent's training data or retrieval results, not the purchasing decision directly. You can't measure influence on a black-box model's outputs.

Attribution lives on one assumption: you can trace a customer action backward to a marketing touchpoint that influenced it.

Humans break this assumption in human ways. They use ad blockers. They switch devices. They forget where they heard about you. But they're still the decision-maker.

Autonomous agents break this assumption fundamentally. An AI shopping agent doesn't just consider your campaign. It scans real-time inventory, pricing, reviews, competitor availability, customer purchase history, and contextual preferences (shipping speed, eco-friendly options, budget range). Your campaign is one signal in a 50-signal model. When the customer buys, you have no idea if your campaign was the deciding factor or background noise.

Worse: agents change behavior continuously. If OpenClaw routes to Claude this week and Gemini next week, the model's reasoning logic shifts. A customer's agent-assisted purchase path from today isn't reproducible next month. Your historical ROI doesn't predict future ROI.

Even worse: if the agent uses reasoning models (like o1 or future variants), it's doing multi-step problem-solving that's partially opaque even to the agent builder. The agent itself doesn't know why it recommended X over Y. It just did. And now you need to attribute a purchase back through that reasoning.

Some teams lean on incrementality testing to escape attribution hell: run an experiment, hold out a cohort, measure the difference, prove impact. This works great for campaigns. Run two creative versions, measure lift, done.

This breaks for agents. Why? Because agents compound across interactions. If you pull a campaign for a test cohort, the agent still has access to all your other touchpoints, content, and brand signals. The agent still recommends your product. You measure 5% lower conversion. But is that because the campaign created the demand, or because the agent would have recommended it anyway?

Second problem: agents learn in real time. By the time you finish a 2-week incrementality test, the agent's model has updated, the customer base has shifted, and the test results are partially obsolete. You prove impact on a version of the system that no longer exists.

Third: attribution isolation is impossible. You can't measure "only the impact of this campaign" because the agent pulls from everything, your website content, SEO, past customer reviews, competitor signals. Separate the campaign and the agent still wins with the background noise.

The panel will likely suggest: first-party data strategies (you own the signals), advanced identity resolution (know which customer it really is), and LLM-assisted attribution (use AI to guess). These are necessary. They're not sufficient.

First-party data is crucial, but it doesn't solve agent opacity. Even if you know a customer's full journey, you still don't know which signals the agent actually weighted.

Identity resolution helps you track the same person across channels, but it doesn't help you attribute to the right channel if agents are doing the heavy lifting.

LLM-assisted attribution sounds clever, "use AI to understand AI", but it's just another layer of guessing. You're asking a model to explain why another model did something. The explanation is plausible. It's not accurate.

Marketing teams are behaving like the problem is fixable through better tools. It's not. The problem is structural. In an agentic world, the causal chain from marketing to sale is broken. You can't fix a broken causal chain with better attribution. You need a different business model.

Here's what shifts: First, shift from ROI on campaigns to ROI on brand infrastructure. Instead of measuring "what did this ad spend create," measure "what market position does our brand occupy." Did agents rank you first in reliability? Second in price? Third in innovation? Those are the outcomes that matter when agents are the funnel.

Second, shift from campaign attribution to agent influence measurement. Work with your agent provider (OpenClaw, Claude, whatever orchestrates for your customer) to see when and how often your brand/product was surfaced, ranked, or chosen by the agent. It's not a customer conversion. It's an agent recommendation. These are different metrics.

Third, accept that some ROI is unmeasurable now. A customer used an agent, the agent was trained on your content, the content influenced the model. But you can't prove it. Get comfortable with that. Measure what you can (agent impressions, agent recommendations, market share signals) and stop pretending you can trace individual conversions.

Fourth, invest in becoming the default choice for agents to recommend. That means incredible product, incredible reviews, incredible customer experience. Agents optimize for customer satisfaction, not marketing spend. Become too good to ignore.

The ROI crisis Gartner is hosting a panel about today is real. But it's not solvable with measurement. Marketing's job is changing. You're not running campaigns anymore. You're building brand infrastructure that agents want to recommend. That requires different skills, different metrics, and a completely different definition of success.

The teams that figure this out fast will have a massive advantage. The teams that spend 2026 trying to perfect attribution models will spend 2027 wondering why they lost market share to brands that agents prefer.

The measurement crisis is unfixable because the problem isn't measurement. It's causality. And causality broke when purchasing decisions moved from humans to machines.