Programmatic ad buying has become a critical part of the ad industry, accounting for more than $4 out of every $5 spent on digital display ads. Programmatic’s ascendance was the result of massive investment by brands, agencies and startups in software, hardware, and data science, all with the aim of optimizing campaign performance and bringing spending efficiency to an internet where audiences were increasingly spread out.

The technology took care of the efficiency piece, giving advertisers the scale they wanted without having to negotiate thousands of individual publisher-direct deals. True optimization has lagged though, due to one major hurdle: for all of the investment in technology and computing power, a human still steps in at the last mile to make critical decisions about targeting — decisions that dilute or possibly even negate the value of the technology-powered decision making that came before.

The issue stems from the primitive way that data segments are still bought and sold, more often than not by a human trader entering a keyword into a search box and choosing segments to target based on description alone. For programmatic to truly realize its potential — especially in a time of tightened budgets — now is the time to adopt tools which remove human intuition and let the machines do more of the thinking.

A last-mile issue

Today, the minority of programmatic buys are targeted, optimized and self-tuned in real-time via machine learning. The significant majority are targeted manually, with media traders logging into their platforms of choice, searching for their target audiences (surfers, soccer moms, auto intenders, etc.) and picking from 50 to 500 audiences that match the search.

They might pick an audience segment based on the data company’s brand name, or they may go with something that was effective for a past campaign from the same client. But in essence, they’re throwing darts to see if they can find the best audience for targeting the campaign because they don’t have access or insight into audience composition, or what’s truly going to work for their campaign.

Hundreds of millions of dollars have been invested in machine learning and state-of-the-art technology to execute targeted ad buys in a fraction of a second. Major advancements in processing power mean that billions of data points are collected and fed into the system to enhance future campaigns and find the best prospects. Once the first flight is up and running, machines should take over and continue executing the campaigns against the best audiences. Why then, are agencies turning everything over to a human being, who rely mostly on intuition?

The data optimization gap

The reason that so many programmatic buys are targeted by humans, rather than machines, is that the data purchasing process is still primitive, despite the sophistication of the technology used to buy and serve the ads.

Typing into a search bar and then clicking away is certainly easy, but the lack of sophistication makes this method feel like it’s a holdover from some 1.0 version of programmatic. There is very little visibility into the provenance, cleanliness or recency of the data selected. There’s merely a segment name, the company that built the segment, and the cost. The technology powering programmatic buying is ultra-sophisticated. Why then, does the audience selection at the last mile amount to an essentially random decision?

Most demand-side platforms provide tools that leverage machine learning to optimize audience targeting. These tools are constantly testing audience segments against campaign KPIs and shifting spends toward those that are delivering the best performance.  This is not an optimization that can be done at a human scale.

Trouble is, the data used for targeting is often an afterthought in the broader campaign planning process and is something left up to the media trader. Agencies work on tight time schedules, and their strategy and planning teams often hand their trading desks a very high-level demographic description of their target audience. Traders, then, attempt to find a fit for that targeting strategy within the sea of third-party segments in the data marketplace.

An optimized future

To be clear, nearly every player involved in programmatic advertising bears some of the blame here. Brands still think in terms of broad-based demos or personas, ignoring the technologies that allow them to optimize against the actual KPIs that drive business results. Media agencies propagate these legacy targeting strategies, which are the path of least resistance. And the buying platforms have focused primarily on driving usage of their technology, happy to integrate data partners but leaving the data sales component an afterthought.

The only path forward for brands and agencies is to advance the way data is bought and sold, evolving these primitive systems for the present and future. The IAB’s data transparency label effort is a major step toward understanding what goes into segments, but quality and accuracy remain outside the project’s scope. Many buying platforms realized in recent years that they had to clean up the inventory available, due to redundancy and low quality. This process improved experiences all around, and improved margins for the platforms and the agency trading desks. The same needs to be done for data. Data marketplaces are full of low-quality audience segments, with opaque methodology and questionable provenance, built using outdated observations. They need to be culled so that buyers aren’t spending money on stale junk.

Cleaner marketplaces and greater emphasis on data optimization would free traders up to apply that human judgment in a higher value way for their clients.  Rather than choosing audience segments from a search box, they should be deciding when to let computers optimize, what to optimize for, and how to know when that optimization is working. The best-case scenario is a human thoughtfully choosing KPIs that meet the brand’s objectives, then hitting the “optimize” button on their platform of choice.

This level of strategic thinking is harder than it sounds, and it’s much higher-value activity than the time spent combing through data segment search results. As the ecosystem changes amid the cookie phase-out, connecting objectives and KPIs to execution will only get harder. Letting machines perform the data selection will free traders to spend more time putting their minds toward critical matters of campaign strategy.