Chapter 6 of 11 · 10 min read
The Other Half
From The Market for Your Mind by Wrotebook
On a Tuesday evening, a person in Norwich taps a link to a local newspaper's website, to read a story about a planning dispute over a car park. What follows is an illustrative account, with the companies left unnamed and the details simplified, but the sequence is the one that runs, many billions of times a day, behind almost every page on the open web that carries advertising. It takes roughly a tenth of a second, often less and sometimes more. The reader experiences it as a short pause before the page settles.
In the first few thousandths of that pause, the page notices that it has spaces for advertisements and nothing to put in them. It sends a message to an advertising exchange, and the exchange composes a short description of the opportunity. A reader is here, it says, on this page about a car park, in roughly this part of Norfolk, on this kind of phone, at this time of the evening, and here is what the identifiers in the reader's browser have to say about where else this reader has been. The description goes out to a large number of prospective buyers at once, each of them a piece of software acting for advertisers. Each consults its own records, which may include what this reader looked at last week and did not buy, and works out what the opportunity is worth to its clients. Some decline. Some bid. The highest bidder wins and pays roughly what it needed to beat the next one, the winning advertisement is fetched and dropped into the space, and the reader, who has been described, appraised, auctioned and sold, scrolls down to find out what the council has decided, past an advertisement for a pair of walking boots they had looked at on Saturday.
Between the advertiser who wanted to sell the boots and the newspaper that wanted to be paid for the car park, a surprising number of businesses touched the money on its way through. There was software acting for the advertiser, choosing which auctions to enter and how much to bid, and charging for the service. There was software acting for the newspaper, offering its spaces to as many exchanges as possible, and charging for that. There was the exchange in the middle, which charged for holding the auction. There were companies supplying data about the reader, to help the buyers decide what the reader was worth, and charging for the data. And there were companies paid to verify that the advertisement had actually appeared, on a real page, in front of a real person, for long enough to count, which charged for the reassurance. Each of these was a reasonable business doing a recognisable job. Each took its share whether the boots were bought or not, and almost all of them were paid out of the same pound, which set off from the advertiser at the top of the chain and arrived at the newspaper, eventually, rather lighter.
Nobody in the history of the attention market had ever done this before. Television sold a crowd, in half-hours, to advertisers who could not see it. The banner sold a space at the top of a page. Search sold the moment someone asked for something. Real-time bidding, which grew up around 2009 and 2010 and acquired a shared technical language when the OpenRTB standard was begun in November 2010 and adopted by the industry's trade body in January 2012, sold individual people, one page-load at a time, to whichever buyer valued them most in that tenth of a second. It is a remarkable piece of engineering. It is also, from the point of view of the person being sold, entirely invisible, with one exception, which appears on the screen every day and which almost nobody reads.
The cookie banner is the one place where the machinery asks permission. It is usually attributed to GDPR, the European data protection law that applied from 25 May 2018, but the requirement to ask before placing non-essential cookies on someone's device is older and comes from elsewhere: the European ePrivacy rules as amended in 2009, which took effect across the EU by May 2011, and their British counterpart, the Privacy and Electronic Communications Regulations, amended the same month. What GDPR added was a standard for what counts as consent, which is that it must be real. The banner was the industry's response to being told that it had to ask, and it was designed, with considerable care, by people who did not want the answer to be no.
Accepting usually took one click and refusing took several, buried a layer or two down behind a button marked something like Manage Options. On 6 January 2022 France's data protection regulator, the CNIL, fined Google €150 million and Facebook €60 million because, on their sites, refusing cookies was more trouble than accepting them. Britain, for its part, loosened the rules a little from 5 February 2026, when the Data (Use and Access) Act 2025 allowed some cookies, for things like analytics, to be set without asking, while keeping the requirement for advertising cookies. The banner remains, therefore, the closest thing the open web has to a receipt, presented before the purchase rather than after it and signed by most people without a glance. Written honestly, it might read:
This website would like to sell you. While this page loads, we will describe you to several hundred companies you have never heard of, let them bid for your attention, and show you whichever advertisement wins. Some of them will remember you afterwards. [ Sell me ] [ Sell me, but let me feel I was consulted ] [ Read 40 pages of settings, then sell me anyway ]
For a while it seemed that the identifiers behind all this might disappear of their own accord. Google, which makes the world's most widely used web browser and also runs the world's largest advertising business, announced a plan to remove third-party cookies from Chrome and spent several years developing replacements under the name Privacy Sandbox. In July 2024 it abandoned the plan to remove the cookies, and in October 2025 it retired most of the replacements, citing low adoption. Chrome still accepts third-party cookies. The company with the most to gain from the arrangement and the company with the power to end it had held a long and thoughtful consultation with itself and concluded that the arrangement should stay.
So much for the people being sold. The more surprising discovery, when it came, concerned the people doing the buying.
The web had been sold to advertisers, from the very beginning, as the end of a century of guesswork. The complaint usually attributed to John Wanamaker, that half of the money spent on advertising was wasted and nobody knew which half, was supposed to be answered at last by machines that counted everything. Every pound could be followed from the advertiser's account to the screen where the advertisement appeared. In the 2020s the advertisers finally tried to follow it, and the trail went cold in places nobody had expected.
In May 2020 ISBA, the trade body for British advertisers, published a study it had commissioned from the accountants PwC, which followed programmatic advertising money through the supply chain from some of Britain's largest advertisers to some of its most reputable publishers. It found that the publishers received, on average, about 51 per cent of what the advertisers had spent. Much of the rest went on fees to the intermediaries in between: the exchanges, the platforms acting for buyers and sellers, the data and verification companies. But about 15 per cent, roughly a third of the supply chain's costs, could not be attributed to anyone at all. The researchers called it the unknown delta, which is accountants' English for money that has left the building without saying goodbye. They were working, it should be said, from a system so opaque that only about 12 per cent of the impressions they started with could be matched from one end of the chain to the other, which made the figures best estimates and the opacity itself the most striking finding.
In the United States, the Association of National Advertisers ran a study of its own. Its first look, in June 2023, estimated that open-web programmatic advertising in America was an $88 billion market, and that about 21 per cent of the impressions bought in it, and 15 per cent of the spending, went to what the industry calls made-for-advertising sites: pages built chiefly to be counted, crammed with advertisements, filled with content of no particular interest to anyone and visited largely by people who had been sent there by other advertisements. Anyone who has followed a headline promising twenty astonishing facts about a celebrity's childhood and found themselves on a page of one sentence, three images and eleven advertisements, with a button marked Next leading to another page exactly like it, has visited one. They are, in their way, a perfect product: built to the specifications of the measurement, delivering exactly what the metrics count and nothing they do not, and they were being bought, in volume, by the most sophisticated advertisers in the world. The ANA put the waste at as much as $20 billion. The full report, in December 2023, estimated the efficiency gap at $22 billion and concluded that only about 36 per cent of each dollar entering the platforms that buy advertising on advertisers' behalf effectively reached the consumer. The study rested on the records of 21 volunteer advertisers, scaled up to the market, and the advertisers who volunteered were presumably the ones who thought they were doing rather well.
Some of the missing money was not wasted so much as taken. In December 2016 the fraud-detection firm White Ops described an operation it called Methbot, which ran a large network of computers pretending to be people watching videos on websites pretending to be real publishers, and which the firm estimated was taking between $3 million and $5 million a day from advertisers. That figure was the firm's own extrapolation. When one of the men behind it, Aleksandr Zhukov, was convicted by a jury in Brooklyn in May 2021, and sentenced that November to ten years in prison, the losses prosecutors could actually prove came to a little over $7 million. The distance between the estimate and the proof was itself an exact description of the problem. Nobody could say with confidence how much had been stolen, because nobody could say with confidence how much had been real.
The platforms that measured their own audiences, meanwhile, were measuring them in their own favour. In September 2016 Facebook acknowledged that a figure it reported to advertisers, the average time people spent watching their videos, had been inflated, because it left out views of less than three seconds. Early reports put the overstatement at 60 to 80 per cent; advertisers who later sued alleged a good deal more. Facebook settled their class action for $40 million in October 2019, without admitting any wrongdoing, which is the customary way of concluding that a number was wrong without anybody having made it so. The advertisers had been buying audiences counted by the seller, inside the seller's walls, on the seller's definitions, and checking the count by asking the seller.
The largest seller of all turned out to be sitting on several sides of the table at once. On 17 April 2025 a federal judge in Virginia, Leonie Brinkema, ruled that Google had illegally monopolised two of the markets at the heart of the machinery that sold the reader in Norwich: the software publishers use to sell their advertising space, and the exchange where that space is auctioned. She also found that Google had unlawfully tied the two together. A company that ran the auction, acted for many of the sellers and many of the buyers, and supplied much of the information about who had won and at what price was, in short, the auctioneer, the scoreboard and several of the bidders. In a remedies opinion of 2 September 2026 the judge declined to order Google to sell its exchange or its publisher software, and imposed conduct rules overseen by a monitor instead. The European Commission, which had reached a similar view of the facts, had already fined Google €2.95 billion in September 2025 for favouring its own exchange. Google is appealing.
Put those findings side by side and the shape of the market changes. Until now, the story has been about media learning to sell their audiences to advertisers, with the advertisers cast, by default, as the buyers who knew what they were getting. They did not. They were paying for attention measured by people who were also selling it, through chains of intermediaries who took their cut whether or not the attention was real, on platforms that marked their own homework and in auctions run by a company that was also bidding. The advertisers were not only the customers in this market. They were marks too, and the people who did well whatever happened to the advertising, whether it worked or failed, reached people or bots, appeared on a newspaper or a page built to be counted, were the people who owned the measurement.
That is the answer, as far as it can be given, to the old question of which half. It was not lost in the sense of being spent on advertisements that people ignored, although plenty of it was. A good deal of it was never spent on advertisements at all. It went to the machinery, and to the people who ran the machinery, and to the people who counted what the machinery did and reported, with great precision, that it had worked. And some of it went nowhere that anybody, including the accountants hired to find it, could identify, which is how the most measured advertising system ever built came to be summarised, in its most careful audit, by a figure of about 15 per cent, labelled unknown.