Averaged across the data brokers tested in three field studies, the gender recorded in a web user's purchased profile was correct 42.3 percent of the time.1 Individual brokers ranged from 25.7 percent to 62.7 percent. The studies, by Nico Neumann, Catherine Tucker, and Timothy Whitfield, covered more than 90 purchased audiences from 19 data brokers. Random classification would be right about 50 percent of the time, and the paper says that using browsing profiles for gender "appears on average less efficient than using nothing." Age was worse. Accuracy for the 18 to 24 band averaged 10.7 percent, for 25 to 34 it averaged 25.7 percent, and for 35 to 44 it averaged 32 percent.
A purchased audience is a set of inferences the buyer cannot inspect
The Federal Trade Commission described how such segments are built in a 2014 report on nine data brokers. "Data brokers infer consumer interests from the data that they collect," the report says. "They use those interests, along with other information, to place consumers in categories."2 The categories are then sold under names such as "Urban Scramble" and "Mobile Mixers." An advertiser who buys one is buying that inference, and the Neumann paper notes that little is known about how reliable such profiles are.1
The field tests linked each broker's cookie to one survey panel member, whose own answers were the check on the label.1 On its own, a profile raised the share of ads reaching a person with the wanted attribute by 0 to 77 percent over random selection. With optimization software the average gain was 123 percent, but the added cost was 151 percent. The authors conclude that third-party audiences are "often economically unattractive, except for higher-priced media placements."
Interest segments scored higher because they were built from pages people visited
The same study found interest segments more accurate than demographic ones. Users labeled as interested in sports had told the panel so 87.4 percent of the time; for fitness the figure was 82.1 percent and for travel 72.8 percent.1 Part of that is the base rate, because the paper puts sports interest among Australians at 67 percent. Against those base rates the authors calculate that interest segments improved on random selection by 30 percent for travel, 30 percent for sports, and 71 percent for fitness. "Overall," they write, "we find higher hit rates (accuracy) for our tested audience-interests than for our previously tested demographic attributes."
An interest segment is a browsing record turned into a label, attached to a cookie, and sold as a predefined audience.1 The label follows the cookie to whatever page the person opens next. Each step between the reading and the ad is a place for the label to be wrong. A bubble skips those steps. AdBubbles matches a word on the page the reader is looking at to the advertiser who bought it, with no stored guess about the reader.
Readers of a specific page have already shown what they are looking for
Europe's Privacy Directive gave Avi Goldfarb and Catherine Tucker a different test of the same point. Their data came from 3.3 million survey-takers who had each been shown one of 9,596 banner campaigns. Once the Directive restricted data use and tracking, effectiveness fell by around 65 percent.3 General-content sites such as news services lost more than specific-content sites such as travel and parenting sites. "Customers at travel and parenting websites have already identified themselves as being in a particular target market," the authors write, so those sites had less need of browsing data to target their ads.
Kaifu Zhang and Zsolt Katona define contextual advertising as showing ads "based on the content that consumers view, exploiting the potential that consumers' content preferences are indicative of their product preferences."4 A parenting site is context at the scale of a site; a word inside an article is context at the scale of a sentence. The reader who has reached the paragraph where "convertible car seat" appears has narrowed the audience further than any site-level label could, without anyone recording who they are.
A business buyer is identified by the article, not by a vendor's record
The Neumann study tested consumer attributes: gender, age, and leisure interests. It did not test the job titles and company sizes that business audiences are sold by, and we have not found a published field test that does. The reader of a trade article needs no profile at all. Someone reading about coolant filtration for machining centers has identified their trade and their problem more precisely than a job-title field could, by opening the page.
Anja Lambrecht and Catherine Tucker found a related pattern in retargeting, the practice of showing a person the exact product they viewed earlier. In their field experiment, "dynamic retargeted ads are on average less effective than their generic equivalent."5 The specific ad stopped underperforming only when later browsing, such as visiting review sites, suggested the person's preferences had changed. The reading was the signal; the stored product view was not.
Suppose a firm sells dust collection systems to small woodshops. On AdBubbles it would buy the words "dust collector" and "shop air quality" on the sites it names. For that calendar month it would be the only advertiser on those words on those sites. Its bubble would open only when a reader hovered over or tapped one of those words in an article.
Most readers assume they are tracked and few feel in control
Pew Research Center asked 4,272 U.S. adults in June 2019 how much tracking they assumed. It found that 72 percent felt all, almost all, or most of what they do online or on their cellphone is tracked by advertisers, technology firms, or other companies.6 By May 2023, in a survey of 5,101 adults, 73 percent said they had little or no control over what companies do with their data.7
An advertiser reaching a niche reader through the words on the page has no part in that. AdBubbles sets no cookie, stores no identifier, and collects no personal data from readers. The advertiser's dashboard shows page views, bubble opens, and clicks per keyword per site per day, and never who. The privacy policy says the same.
Where the page says nothing about the reader, a profile is the only signal
Demographic targeting still does something a page cannot, and the two field studies show where. Goldfarb and Tucker's largest losses were on general news sites, which they call places "where non-data-driven targeting is particularly hard to do."3 Neumann and colleagues exempt higher-priced placements from their conclusion that purchased audiences do not pay, and their interest segments outperformed random selection.1 A product that no one reads about before buying, or a buyer defined by a life stage that no article names, gives a word-based ad nothing to attach to.
A word that appears on no page in the network cannot be bought on it. AdBubbles does not choose which people see an ad, does not follow a reader from one page or site to another, and does not promise a result, as the seven steps describe. Building a campaign costs nothing, and before paying an advertiser sees for each word how many sites carry it and the estimated monthly page views. That is enough to learn whether a business's buyers already read about it.