A job posted on Amazon Mechanical Turk by three researchers asked workers to sort emails from the public Enron archive into four categories: spam, personal, work, or e-commerce. The pay was 25 US cents plus a bonus of one, two, or three US cents for every five emails sorted, and a worker could stop at any time. The screen around each email was randomly assigned to carry no ads, ads rated innocuous, or ads rated annoying. Each email sorted counted as one ad impression. The question was how many impressions each of the 1,223 workers would produce before quitting.1
Workers shown annoying ads sorted the fewest emails. Relative to that group, workers shown innocuous ads produced 19 percent more impressions, and workers shown no ads produced 25 percent more.1 The authors turned the gap into a compensating differential: the extra pay a person needs to produce the same number of impressions alongside annoying ads. The differential came to US$1.53 per thousand impressions against no ads, and US$1.15 per thousand against innocuous ads.1 A fuller version published the next year, with a mouse-tracking study of reading, concluded that in plausible scenarios running annoying ads can cost more money than it earns.2
Combining relevance with obtrusiveness does worse than either alone
That experiment measured how long readers stay. A field experiment reported in Marketing Science in 2011 measured purchase intent instead. Goldfarb and Tucker found that matching an ad to the page's content raised purchase intent, and so did making the ad more obtrusive. An ad that did both did worse than an ad that did only one.3 The authors traced the failure to privacy concerns: the penalty was strongest among people who declined to state their income and in categories where privacy matters most.3
AdBubbles matches a word already in a publisher's article to the advertiser who bought it for the month, which is targeting by content alone. The reader sees a light dotted underline, and a small bubble opens only when they hover or tap on the word. AdBubbles sets no cookie, stores no identifier, and collects no personal data from readers. We do not choose which people see a bubble, do not follow a reader between pages or sites, and do not promise any advertiser a result. The relevance comes from the page the reader chose to open, not from anything known about the reader.
Readers have ranked the ad experiences they find most annoying
Before the email task, the same researchers had 141 people rate a large set of ads taken from the web. The 21 most annoying ads were all animated, and the 24 least annoying were all static.1 The most frequent complaint was the animation itself, in words such as "too much movement". After that came complaints that the ad drew attention away from the content of the page, looked cheap, or looked like a scam.1
In March 2017 the Coalition for Better Ads reported that more than 25,000 consumers had rated 104 ad experiences while reading simulated content pages.4 On desktop, the experiences beneath its standard were pop-up ads, auto-play video ads with sound, prestitial ads with a countdown, and large sticky ads. The mobile list named eight: pop-up ads, prestitial ads, ad density above 30 percent, flashing animated ads, auto-play video ads with sound, poststitial ads with a countdown, full-screen scrollover ads, and large sticky ads.4 Chrome began filtering against that list on February 15, 2018; a site still carrying a listed experience 30 days after notice has its ads blocked by the browser.5
A Nielsen Norman Group survey of 452 adults in the United States adds one item. On desktop, four types tied for most disliked: modal ads, autoplay video ads, deceptive links, and intracontent ads that shuffle the page's content as they load.6 The least disliked types were static ads in the right rail and related links on desktop, and related links on mobile.6 Respondents' reasons were that an ad pops up, loads slowly, covers what they are trying to see, moves content around, occupies most of the page, or plays sound by itself.6
Readers who block ads read more of the site, not less
The Reuters Institute's 2017 Digital News Report, drawn from 70,000 respondents in 36 countries, found that 24 percent used an ad-blocker.7 In the United States, 52 percent of ad-blocker users had temporarily switched the blocker off for a news site.7 More than half of those who disabled a blocker said it was the only way to see the content.7
What those readers do next has been measured on one news site. A 2022 study in the Journal of Marketing Research followed 79,856 registered users across 3.1 million visits and measured how their reading changed after they adopted an ad blocker.8 Adoption raised the number of articles a user read by 21.0 to 43.2 percent, and the number of content categories by 13.4 to 29.1 percent, mostly through repeat visits.8
A bubble opens only when the reader acts on it
Each item on those lists is something done to a reader who did not ask for it. The ad appears, moves, or plays sound on its own, covers the article, or shifts the article as it loads. A bubble does none of those on its own. The embed, one script called adbubbles.js, loads after the page's DOM is ready and underlines the first occurrence of a bought word inside the publisher's chosen scope. A desktop reader opens the bubble by resting the pointer on the word for a moment and closes it by moving away or pressing Escape. Phone readers tap the word to open it and tap anywhere else to close it.
A bubble holds an "Ad" label, a headline of at most 40 characters, a body of at most 90 characters, the advertiser's domain, and one link. The bubble also honors the reader's reduced-motion setting. A publisher sets the maximum bubbles per page, the scope, and the advertiser categories to exclude in the configuration.
The "Ad" label follows the Federal Trade Commission's 2015 guide to native advertising. The guide says that knowing something is an ad affects whether consumers interact with it and the weight they give it.9 On placement, the guide says a disclosure should generally be made in the ad itself, shortly before the advertising message, and it lists "Ad" and "Advertisement" among terms likely to be understood.9 A reader who opens a bubble sees the label as the message arrives.
The cost of an annoying format is counted in readers
The consequence for a publisher is an accounting one. The payment for an annoying format appears on the remittance; the cost appears in readers who sorted fewer emails, installed a blocker, or read fewer articles. One experiment priced that cost at US$1.53 per thousand impressions, set against whatever the annoying ad paid.1 A format that waits for the reader's hover or tap makes the interruption the reader's own decision, and a reader who never hovers is never interrupted. Neither a publisher account nor the embed costs anything, and the crawl a new publisher can request shows which words on the site are currently sellable, before any bubble goes live.