The exact LLM prompts media buyers are using right now

Media buyers are no longer simply experimenting with large language models (LLMs); they’re embedding them into live campaign workflows, turning to AI to increase planning precision, speed up campaign activation and extract better insights from the deluge of reporting data.

Prompt design itself has become a competitive advantage. The strongest prompts include extensive context and request outputs that can be validated and iterated on, said Christopher Robinson, head of client insights at Tinuiti.

“Usually this means including relevant datasets as attachments with the prompt and always asking for code output that we can review and run directly, outside of the chat,” he said.

But it’s mission critical to have human-verified inputs and outputs, said Rob Auger, executive vice president at Digitas. “We train our media teams to effectively build prompts that extract maximum value from AI agents while maintaining accuracy and client relevance,” he said. This can also look like structuring prompts as step-by-instructions to “maintain human oversight and input at each phase.”

That shift shows up across all three stages of the campaign process.

Stage one: Planning

At the planning stage, buyers are using LLMs for intelligent recommendations designed to maximize campaign performance.

For example, one might ask: “Please look at [specific advertiser account]. Use your tools and knowledge of programmatic advertising to find creative ways to increase the performance of these campaigns,” explained Jeff Ellin, vice president of product architecture at FreeWheel.

At PMG, AI Product Lead Brent Aydon said buyers might prompt a model to profile audiences with instructions like: “Assume the role of a social media manager, market analyst and brand strategist. Describe three target audience segments for an online retail brand specializing in eco-friendly products, considering demographics, interests and online behavior.”

Newton Research, a marketing-analytics AI company, has designed and implemented specialized buy-side agents that can measure, optimize, report and plan. An actual recent client prompt said: “I need help profiling these four audiences as it relates to their media habits. Please start to flesh out the four segments and show how each one is unique from their gender counterpart — female vs. male financial service clients and financial service advisers,” said John Hoctor, the CEO and co-founder of Newton Research. He noted that these prompts rely on knowledge inherent to Newton’s specialized agents and would not work on a typical general-purpose LLM.

But beyond audience development, buyers are increasingly using LLMs as “scenario engines,” according to Christina Kochenderfer, vice president of social and digital investment at independent ad agency Monks.

Instead of asking for a single “best” plan, they’ll prompt something like: “Given a $2 million, CTV-heavy budget and an awareness goal, show me three distinct strategies and the trade-offs of each.” The value isn’t the output itself, but how quickly assumptions can be pressure tested before a planning spreadsheet even exists.

More advanced teams layer custom inputs such as historic performance, client briefs and benchmarks. This allows LLMs to blend industry-wide trends with first-party client data, producing plans that are “as tailored and realistic as possible.”

Stage two: Activation

During activation, LLMs act less like execution tools than operational copilots. “Buyers still launch campaigns inside DSPs, so tools like ChatGPT or Claude aren’t replacing hands-on activation,” Kochenderfer explained. “Instead, teams are using LLMs to speed up and strengthen the work that happens around the build.”

In practice, that means buyers increasingly rely on LLMs to convert a media plan into a recommended campaign structure, with queries like: “Based on this CTV strategy and budget, how should campaigns and line items be organized to manage pacing and frequency?” Kochenderfer said.

“In this scenario, teams use LLMs to support audience setup with prompts like, ‘Given this target audience and KPI, make me a list of relevant keywords and interests to target,’” Kochenderfer said. Post-build, teams rely on LLMs for quality assurance (QA) by asking, “Review this exported build — compare the campaign and line-item names with this taxonomy source of truth. Flag what does not align.”

“The value in LLMs isn’t in automating the actual activation,” Kochenderfer emphasized. “[It’s] in making the surrounding steps faster, more consistent and less error-prone before dollars go live.”

While buyers aren’t launching campaigns directly from LLM interfaces yet, the industry is moving closer, Kochenderfer said. “In early workflows, LLMs generate activation logic or recommended configurations, humans review and approve those decisions, and scripts or APIs push the final setup into platforms.”

Stage 3: Measurement

This is where AI adoption is “accelerating fastest,” Kochenderfer said.

Buyers are using LLMs to make sense of fragmented data sources — surfacing what changed, why it likely happened and where to look next. Teams will prompt models with questions like, ‘Based on these delivery and performance shifts, what are the most likely drivers — pacing, frequency, supply availability, creative rotation or audience saturation? Rank them by likelihood.’”

On the sell side, media buyers can also prompt AI agents to analyze auction dynamics and bidding efficiency.At FreeWheel, Ellin described prompts such as: “Identify line items where we’re winning > 40% of auctions, where we’re likely overbidding. Model bid reductions — 10%, 20%, 30% decreases — and estimate cost savings for each scenario. Recommend the optimal bidding strategy.”

What’s next

As 2026 unfolds, Kochenderfer argues competitive edge won’t come from chasing the newest tools, but from codifying decision-making processes and teaching AI how teams think.

“AI won’t replace media buyers; but it will replace the parts of the job that never truly required human judgment.”