An online travel firm let two marketing researchers, Anja Lambrecht and Catherine Tucker, run a field experiment on its retargeting. Consumers who had viewed products on the firm's site were later shown its ads on other websites. Some saw a generic brand ad. Others saw a dynamic ad with images of those same products. The detailed ads did worse: dynamic retargeted ads were "on average, less effective than their generic equivalents."1
Timing was the exception. When a consumer's browsing showed that "their product preferences have evolved," the dynamic ads "no longer underperform." A consumer beginning a search, the authors suggest, holds preferences at a high level and responds to high-level information; detail helps only once those preferences narrow.1 The firm knew which products each consumer had viewed, and on average the ad built from that knowledge lost to the ad built without it.
Timing, not detail, decided the result in two field experiments
Another experiment, at a financial tools provider, points the same way. Paul Hoban and Randolph Bucklin randomized cookies between the firm's display ads and charity ads, and sorted users by stage, from nonvisitor to converted customer. The ads raised visits to the firm's site at most stages. They had no positive effect on people who had already visited without creating an account.2 The ads were identical; only the stage differed.
Both experiments relied on a record of the individual, and AdBubbles keeps none. An advertiser buys a word, and the embed matches that word, wherever an article contains it, to that advertiser for the month. We do not pick which people see the bubble, do not follow a reader from one page or site to another, and do not promise a result. AdBubbles sets no cookie, stores no identifier, and collects no personal data from readers. Precision targeting usually means precision about the person; here it means the word and the moment.
A reader's eye rests on each word for about a quarter-second
Keith Rayner and four colleagues summarized the reading research in 2016. The eye stops in fixations "which last approximately 250 ms for experienced adult readers," then jumps forward about seven letters. Readers skip the next word "about 30% of the time." The eye moves, the review says, so that the part of the retina that sees in detail sits over the words the reader wants to process.3 A word is seen by being fixated on, and the fixation is the reader's own act.
A bubble opens on one further act by the same reader: a hover on the underlined word after a short delay, a tap, or Enter. Before that act, the page shows only a light dotted underline. A bubble that stays visible for one second is counted as an open.
Viewing time is concentrated in a page's first paragraphs
Where on the page a word sits changes how often that moment comes. Nielsen Norman Group recorded more than 130,000 fixations from 120 participants on a 1920 by 1080 desktop screen. Users spent about 57 percent of their viewing time above the fold and 74 percent within the first two screenfuls. Readers "tend to look more thoroughly at the text placed close to the top of the page (the first few paragraphs of text)" and give fewer fixations to what appears lower.4
From a log of 59,573 page views, Jakob Nielsen estimated how much of a page gets read. On an average page, he concluded, "users have time to read at most 28% of the words during an average visit; 20% is more likely."5 The embed underlines the first occurrence of a bought word, so a word that first appears in the opening paragraph sits where those studies found the most viewing time.
Kara Pernice's 2017 account of the F-shaped pattern adds a condition. Readers scan in an F, across the top and then down the left edge, when three things hold. The text has little formatting, the reader is trying to be efficient, and the reader is "not so committed or interested that he is willing to read every word." Highly motivated readers show a commitment pattern, "fixating on almost everything on the page."6 Suppose two sites carry one keyword, and its first occurrence falls in the opening paragraph on one and in the eighth on the other. The scanning data put the first placement where fixations are densest and the second where an F-shaped scan may never arrive.
The page reveals what is being read and not why
Andrei Broder argued in 2002 that the words of a query are not "the need behind the query," and sorted those needs into navigational, informational, and transactional. In his user survey, 24.5 percent of queries were navigational, an estimated 39 percent informational, and 36 percent transactional; his query log analysis gave 20, 48, and 30 percent.7 Whichever need a query served, its words were all the search engine had.
The distinction applies to a page as it does to a query. An article containing the phrase "tankless water heater" tells an advertiser the subject the reader chose, and the surrounding sentence tells how the subject arose. Neither says whether the reader is comparing models, repairing one, or writing about them.
A bubble's contents are closer to Lambrecht and Tucker's generic brand ad than to their dynamic product ad.1 A bubble holds an "Ad" label, a headline of at most 40 characters, a body of at most 90 characters, the advertiser's display domain, and one link. That is room for a short statement and none for a product image. Whether that level suits a given reader depends on where the reader stands, which we do not know.
The one second of an open starts at the reader's action
The Media Rating Council counts a display ad as viewable when at least 50 percent of its pixels sit on an in-focus tab for one continuous second, "post ad render." An "opportunity to see," the guideline says, exists with a viewable impression and may or may not exist with a served one.8 An open uses the same length of time with a different start. The council's second begins when the ad renders; ours begins when a bubble the reader opened, by hover, tap, or Enter, becomes visible.
Neither count measures reading. A viewable impression may sit where the reader never looks, and an open records that a bubble stayed open, not that its 130 characters were read. The advertiser dashboard reports views, opens, clicks, open rate, and click rate per keyword per site per day. No column reports the stage of the reader who opened the bubble.
Building a campaign costs nothing. Before paying, an advertiser can pick words, see how many sites carry each one and their estimated monthly page views, and choose the whole network, one category, or named sites. The advertiser page describes a campaign, pricing gives the packages, and how it works lists the seven steps. None of the studies above placed an ad inside the fixated word. Whether a reader who opens a bubble from a word in an article is the consumer whose preferences have narrowed, or the one who has only begun, is the question the research has not settled.