What Sites Work Best for Ads?

For many businesses, deciding where to advertise is a shot in the dark. Because they don’t really know which websites will reach the right customers, they try different websites, eventually settling on what works. In the process, they lose time and money, as their expectations are often off target.

“They are typically four times too optimistic about how many people will click on their ads, meaning they spend too much money on sites that don't work for them,” said Carl Mela, the T. Austin Finch Foundation Professor Emeritus at Duke University’s Fuqua School of Business.

Imagine a flower shop trying to build awareness for a new store. The owner has dozens, maybe hundreds of websites to choose from for ad placements. Testing just one site by running a campaign costs roughly $800, Mela’s research shows, and advertisers often need to test many more before figuring out the best choice.

“This learning by doing process is very expensive,” he said.

In the paper Advertiser Learning in Direct Advertising Markets, published in Marketing Science, Mela and co-authors (Jason M. T. Roos, Rotterdam School of Management, Fuqua PhD, and Tulio Sousa, Indiana University, Duke Econ PhD) propose a smart matchmaking service. Rather than learning through trial and error, an intermediary could use historical data from thousands of past campaigns to recommend where a new ad is most likely to perform well. 

Think of an “ad network,” a company that sits between advertisers and publishers such as newspapers, blogs and websites.

By pooling data on which kinds of ads worked on which sites, the network can direct advertisers toward better matches, leading to better performance and consequently more spending on ads, all the while increasing publisher profits.

“It’s a win-win service that helps advertisers find the right audience and publishers increase their revenue as a result,” Mela said.

The researchers’ model uses AI tools to scan (or “tag”) the ad’s design—such as whether it contains flowers, cars, people, or logos—and compare it with similar ads that have already run in the past. If visually similar ads performed well on specific websites, the network recommends those publishers to new advertisers.

While the research focused on “direct buy” ad markets, this data-pooling approach can improve performance in several other advertising contexts, Mela said.

The direct-buy, display market

In direct buy ad markets, advertisers purchase ad space directly from publishers—such as news sites, blogs, or magazines—often at a fixed price for a week or longer. Rather than buying one impression at a time (for example, in ad exchanges), advertisers buy large bundles of impressions before knowing how well the campaign will perform. 

Publishers often sell their inventory through ad networks, which bring together advertising space from many websites into a single marketplace. Instead of negotiating separately with hundreds of publishers, advertisers can browse ad inventory through a single platform.

But that convenience comes with a challenge, Mela said. Because direct-buy campaigns are purchased in bulk—say, a banner for one week, estimated to generate thousands of impressions—advertisers receive feedback only after a campaign has finished. Unlike programmatic advertising, where algorithms continuously adjust bidding as new information arrives, direct-buy advertisers have few opportunities to learn as they go. 

That trial-and-error approach is expensive. The researchers found that an advertiser typically spent about $800 to test a single publisher, receiving roughly 820,000 ad impressions. Many advertisers eventually abandoned the sites they initially chose, suggesting that they had overestimated how well those placements would perform.

A smarter matchmaker

Mela and his co-authors wondered how to improve advertisers’ information.

An ad network already has something individual advertisers lack: a record of thousands of past campaigns showing which kinds of ads performed well on which websites. By pooling that information across advertisers, the network could recommend publishers that are likely to be a good match before an advertiser spends money testing them.

To make those recommendations, the researchers turned to Google's Cloud Vision API, an artificial intelligence service that analyzes images. The tool identifies an ad’s visual characteristics—recognizing concepts such as flowers, cars, logos, people, or other objects. Ads with similar visual features tend to perform similarly on the same websites, the researchers hypothesized, allowing the system to compare a new advertisement with thousands of previous ads and estimate where it is most likely to succeed. This approach solved a problem that had delayed the research for years until machine-learning technology became sufficiently mature, Mela said.

The researchers tested the approach using data from a direct-buy ad network covering roughly 8,000 advertisers and 3,200 publishers between 2006 and 2009. The results suggested that better information benefits everyone involved. The new matchmaking model improved advertisers’ ROI and allowed publishers to attract advertisers that fit their audiences more closely. The ad network itself also stood to gain from higher overall advertising spending: the researchers’ model projected an additional $7.5 million in revenue over six months.

While the model is facilitated by the presence of an ad network, it does not require it, Mela said. “For example, an ad agency could use this approach. It’s just easier with a network,” he said.

A win-win solution

In theory, publishers might be reluctant to share information about past advertising performance, Mela said.

If advertisers are overly optimistic about an ad’s performance on a website, giving them better information could discourage them from buying ads there. “Why would publishers help advertisers spend less?” Mela said.

The researchers found that this view overlooks a more powerful effect.

Better information would help advertisers discover websites where their ads are much more effective. Instead of wasting money testing one publisher after another, advertisers can invest more confidently where they are likely to generate stronger results.

And, importantly, as ads become more effective, advertisers spend more overall.

In the researchers' simulations, that increase in spending more than offset the revenue publishers lost from correcting advertisers' overly optimistic expectations. The model projected that the median publisher's revenue would increase by 77% over six months, driven by advertisers investing more in websites that better matched their audiences.

"The advertisers they lose are more than offset by the advertisers they gain," Mela said. "By sorting advertisers and publishers and matching them better, everybody is better off."

Beyond direct-buy advertising

According to the paper, direct-buy advertising accounts for roughly 75% of the $130 billion display advertising industry.  Beyond direct-buy, similar challenges are common in many other advertising formats, Mela said.

“Choosing the right publisher, audience, bid, or the visual design of the ad has become so complex that many businesses are handing those decisions to platforms such as Google, Meta, and TikTok,” he said. “Rather than selecting every advertising setting themselves, advertisers increasingly define a goal and overall budget. The rest: figure it out for me," he said. 

For these reasons, the research suggests that platforms could instead pool data from thousands of past campaigns to recommend better matches from the start.

The idea also extends beyond ad networks. Large publishers such as big newspapers or retail media platforms like Amazon and Walmart, as well as streaming services and traditional media such as television and radio, all face versions of the same learning problem: helping advertisers find the audiences most likely to respond before they spend money testing every option. 

"The scale of these problems—the number of participants, the competition, the economics, the statistics—is just so complex. It's almost impossible for a single company to realistically optimize every decision on its own," Mela said.