For all the AI-in-ad-ops talk, plenty of publisher teams are still trapped in the grind of pulling GAM reports by hand and trying to reverse engineer why revenue dropped.
At Programmatic AI in Las Vegas, Jordan Cauley, who launched a publisher monetization tech consultancy after an eight-plus-year stint as a product lead at Mediavine, made a simple argument that large language models only help with the ad tech grind when they plug directly into systems publishers already use. For instance, Google Ad Manager (GAM), GitHub and revenue reconciliation feeds.
Cauley aimed his session squarely at publisher ad ops and product teams who want AI to take real work off their plates. His examples were all grounded in day-to-day tasks like diagnosing revenue dips, unpacking the impact of Prebid software updates and reconciling SSP discrepancies without losing days in GAM.
For Cauley, the benefit of adding LLMs is saving time. His clients are cutting revenue mysteries that used to take two weeks to solve down to three hours. He described using AI to solve how a recent Prebid update quietly cannibalized a publisher’s outstream video revenue months after launch. And he said his morning routine has gone from logging into three different platforms before coffee to looking at one synthesized view of GAM, GitHub and SSP revenue gaps.
The catch is that none of this comes out of the box. Every GAM instance is bespoke. LLMs still hallucinate. Most “agents” on the market are, in Cauley’s words, more Pinto than Ferrari.
For publishers that want AI to do real work in ad ops, the challenge is wiring models into the right systems and teaching them how the business actually runs.
From two‑week fire drills to three‑hour investigations
One of AI’s core ad ops use cases, Cauley said, is understanding a sudden revenue dip with no clear explanation. In a traditional workflow, a weird trend sends teams into a cycle of pulling niche reports, filing engineering tickets and waiting hours for someone to confirm when a change shipped, he said. That kind of investigation used to take ad ops teams two weeks to complete for a major issue.
Once Cauley wired Claude and ChatGPT directly into GAM and his clients’ ad stacks, the same kind of issue took about three hours to narrow down, he said. Instead of building one GAM report at a time, the model runs multiple queries in parallel across different slices of inventory or formats, then synthesizes the results into a single narrative.
He also explained how this AI integration can take the tedium out of ad ops teams dealing with GitHub and other change logs. The system pulls recent code commits – or snapshots of code prior to updates – tied to the ad stack and lines those up against GAM revenue and impression trends from the day before and after each software deployment.
Manually, that comparison is tedious enough that it rarely happens in depth, Cauley said. But AI can be wired to do it with a single prompt.
“Doing this manually is hard,” he said. “Doing this with Claude or ChatGPT takes minutes. This is life-changing stuff.”
But since publisher ad stacks are idiosyncratic, the LLMs have a bit of a learning curve.
“The data structures from GAM’s APIs are actually really solid,” he said. “The biggest problem is that every single publisher has set up GAM in a completely different way.”
Cauley’s fix is to teach the model how a given publisher’s world works before trusting the outputs. That starts with feeding internal docs that explain key values and revenue definitions, then running exploratory queries to map how GAM is structured. He also tells clients to explicitly instruct models to favor accuracy over speed and to cross-check AI-generated answers against raw GAM exports whenever something feels off.
Publishers, he argued, have an advantage since they can always go back to their own systems to see what actually happened. If an AI-generated analysis seems wrong, ad ops can always pull the raw report, reconcile it against SSP data or review the underlying changelog.
That safety net makes this a safer place to push AI than fuzzier creative tasks, as long as teams stay disciplined about verification, he said.
Agents are still in the fetus stage
Real impacts from using AI tools aside, Cauley said it’s still an unproven theory that agentic workflows based on the Ad Context Protocol are the next frontier.
Vendors are already pitching AI agents that will negotiate deals and optimize campaigns on a publisher’s behalf. But Cauley said he’s skeptical of how mature those systems really are.
He sees real potential in frameworks like AdCP for publishers that run many direct deals through GAM, but much less value for teams whose revenue is mostly SSP- and PMP-driven.
Still, his overall view of AI in ad ops is optimistic. The hard parts are no longer building the connectors or wrangling the APIs. The work now is wiring AI into the right data sources and teaching it how each business actually works.
“If you start with the work you already do every day and the projects no one will ever give you time for, AI stops being this abstract thing,” he said. “It just becomes another way to get to real answers faster.”
