Effective cost per thousand impressions, or eCPM, has a one-line definition on Google's AdMob help page: total earnings divided by impressions, multiplied by 1,000.1 The page calls the result "an estimate of the revenue you receive for every thousand ad impressions" and says it "is highly dependent on the market and platform." Suppose a blog's ad units are shown 40,000 times in a month and the account earns US$120. The eCPM is US$3.00 whether the blog explains routers, mortgages, or sourdough. The subject enters the fraction only through the numerator, and the numerator is money other people decided to bid.
The numerator is bid by advertisers and the denominator is counted on the page
Google's AdSense help page on revenue share says that AdSense "pays publishers an effective cost per mille (eCPM) for advertiser bids."2 Publishers keep 80 percent after the advertiser platform's fee, the page adds, or about 68 percent when the buyer came through Google Ads. Google Ads tells advertisers that a winner pays only enough to "clear the Ad Rank thresholds and beat the Ad Rank of the competitor immediately below."3 With no competitor below, the winner pays the reserve price. The cost per click (CPC) on a word is set by the next buyer down, or by the floor when no other buyer exists. A page supplies the impressions, and the advertisers who want its words supply the price.
The same word costs more where more buyers need it
Avi Goldfarb and Catherine Tucker measured that rule in one profession. Their paper in Management Science in March 2011 used the prices lawyers paid for 139 Google search terms in 195 locations.4 Where state regulation stopped lawyers from contacting clients by mail, the price per click for the same terms was 5 to 7 percent higher. The substitution toward online advertising was strongest in markets with fewer customers.
Anindya Ghose and Sha Yang, in the same journal in 2009, modeled six months of data on several hundred keywords bought on Google by a large nationwide retailer.5 Their abstract reports that keyword characteristics, ad position, and the advertiser's landing page quality score all bear on CPC. A higher landing page score goes with a lower CPC. The publisher whose page carries the word never sees that landing page.
On that rule, the spread between niches comes from who else is bidding for the same words, and how much. Suppose two blogs each draw 10,000 views a month, one explaining mortgage refinancing and the other sourdough starters. If four lenders want the first blog's words and one supplier wants the second's, the second blog's clicks sell at the reserve. The first blog's clicks sell at enough to beat the next lender. Tech, finance, and lifestyle are labels for pages; the price attaches to the words on them, one word at a time.
A market that sells the word settles the amount before the month begins
What an advertiser buys on AdBubbles is a word, not an audience. One advertiser holds a keyword on a site for one calendar month. Packages are prepaid by the month. The pricing page gives the split: publishers keep 70 percent of what advertisers pay for the campaigns that run on their pages, and AdBubbles keeps 30 percent. Earnings are shared among the sites a campaign ran on in proportion to bubble opens, accrued daily. The amount to be shared is therefore fixed when the campaign is paid for, and a site's portion follows its share of the opens.
Setting up a publisher account and the embed costs nothing, and a publisher who adds a site can request a crawl listing which of its words are currently sellable. That crawl is the nearest thing to a niche table the network offers. A bought word on a page gets a light dotted underline at its first occurrence. The bubble is the small panel that opens from that underline, and an open is a bubble that stayed visible for one second. The crawl lists the sellable words, the dashboard counts the opens, and the publisher's share depends on the opens.
The published cases changed what advertisers knew, not what the pages said
Three published cases moved publisher revenue by a measured amount, and none of the three is sorted by niche. Each varied what the advertiser could know about the reader and held the page constant. Deepak Ravichandran and Nitish Korula of Google ran the largest. Over 96 days from May to August 2019, they disabled third-party cookies on a small fraction of traffic at the top 500 global publishers.6 Average publisher revenue in the treatment group fell 52 percent, and the median per-publisher decline was 64 percent. Publishers in the News vertical lost 62 percent on average and 60 percent at the median, the one vertical the study reports separately.
Veronica Marotta, Vibhanshu Abhishek, and Alessandro Acquisti examined millions of ad transactions across the websites of one large media company during a week in May 2016.7 Their 2019 preliminary draft found that a publisher's revenue increased by only about 4 percent when the user's cookie was available. Zhengrong Gu, Garrett Johnson, and Shunto Kobayashi evaluated more than 200 million impressions across more than 5,000 publishers.8 Their 2026 paper in the Proceedings of the National Academy of Sciences reports that removing third-party cookies reduced publisher revenue by 29.1 percent. Privacy Sandbox, Google's potential replacement for the cookie, preserved 4.2 percent of the loss.
Those three results disagree on size and agree on what was varied.678 In every case the page's words were the same with and without the cookie; the money that moved was what advertisers paid for information about the reader. A niche eCPM table measures that market at one date, with an unknown share of that information in the figure. "Maximizing eCPM," the field's usual phrase, means raising the numerator, and the words themselves are the part of the fraction none of the three experiments touched.
The first figures will come from the network's own first sites
We have published no open rate, no click rate, and no earnings figure. The first publishers on the network are the owner's own sites, and the first advertisers are the owner's own businesses. The first figures will come from those sites, and any scenario before then is a supposition. AdBubbles sets no cookie, stores no identifier, and collects no personal data from readers. AdBubbles pairs each bought word on a page with the advertiser who paid for it. The network does not choose which people see a bubble, does not follow a reader between pages or sites, and does not promise any result.
The dashboard shows, per site per day, page views with bubbles, opens, clicks, and earnings accrued. Suppose a site's rows for one month sum to 30,000 page views with bubbles, 900 opens, and US$45 accrued. Earnings divided by views, times 1,000, is US$1.50 per thousand views; earnings divided by opens, times 1,000, is US$50 per thousand opens. The first figure sits beside a display report's eCPM for the same month, because both have the page's own traffic in the denominator. The second counts only bubbles a reader chose to open, and both figures come from the same rows on the first day the rows exist.