On June 6, 2026, Gregory Shevchenko spent the day at Profound HQ in New York City as a selected participant and solo builder at the Marketing Engineering Hackathon. The event challenged attendees to find marketing processes too inhuman in scope or scale for manual execution and ship a system that runs them. His DEV.to essay breaks down what he learned—and it's a sharp critique of how most teams approach marketing AI.

The Core Insight

Shevchenko's thesis cuts through the hype: marketing agents don't become useful because they produce better paragraphs. They become useful when they run bounded workflows with clear inputs, preserved evidence, explicit review gates, and a measurement loop. That's less flashy than "autonomous marketing agent," but it's what actually matters. The problem he identifies is architectural, not prompting-related.

Why Most Marketing Agents Are Just Demos

The common failure mode isn't lack of AI tooling—it's unclear ownership of the workflow. One team owns content. Another owns SEO. Someone else owns analytics. A founder owns positioning. Then someone drops an AI tool into the middle and asks it to "make it faster." The result is more output with the same weak source graph. Shevchenko argues that until you define what question the workflow answers, which sources it's allowed to use, and where human approval sits, you're not running a system. You're running a demo with nice-looking slides.

AEO/GEO Is a Workflow Problem

For those working on AI Search visibility—getting cited by ChatGPT Search, Perplexity, Google AI Overviews, Gemini, Claude, Copilot—the tactical checklist approach (answer questions clearly, mention the right entities, add schema) helps but doesn't scale. Shevchenko reframes AEO/GEO as an evidence pipeline: Which buyer prompts matter? Which engines are being checked? Which answers cite competitors instead of you? Which claims need more proof? That's not one content task—it's a repeatable workflow with state and accountability.

The Packet as the Unit of Work

Shevchenko proposes a different mental model. Instead of "an AI-written article," think in packets containing: the prompt set, current answer snapshots, cited-source analysis, competitor gaps, claim inventory, approved source pack, canonical page or distribution asset, human review decision, and follow-up measurement window. The article is just one output inside that packet. This framing makes agents easier to design because each step has inputs, outputs, failure states, and a review point—boring in the best possible way.

Draft Is Not Publish

One rule Shevchenko keeps returning to: draft is not publish. A published page with weak sources can become a liability. A fabricated claim attached to a brand damages the source graph you're trying to build. Marketing agents need permission boundaries that should not be one level: read-only inspection, draft generation, internal review, human approval, external publish, post-publish measurement. The agent should know which state it's in. The human should know what's being approved. The system should keep the evidence trail.

What This Means for Builders

If you're building marketing agents or workflow systems around AI search pipelines, Shevchenko's framework suggests starting with a small measurement loop: pick ten buyer prompts, run them across one or two answer engines, save snapshots, record brand mentions separately from citations, classify source gaps, then close the most critical gap. That shifts the question from "What should we write next?" to "Which source gap is blocking a useful answer, and what evidence-backed asset would close it?" That's where AEO/GEO starts feeling like engineering instead of content marketing with better tooling.

The Bottom Line

The best version of marketing engineering isn't old growth work with a new title—it's a change in the unit of work from artifacts to systems. Marketing agents become useful not as chatbots or content mills, but as bounded systems that help teams inspect, decide, produce, publish, and measure with an evidence trail humans can trust. The fix isn't better prompts. It's clearer workflow boundaries.