The CPM Era: Renting Eyeballs

October 27, 1994. A rainbow-lettered rectangle, 468 by 60 pixels, goes live at the top of HotWired asking whether the reader has ever clicked a mouse right here, and promising that they will. AT&T paid $30,000 for three months of that space. The click led to a virtual tour of seven art museums. Nearly half the people who saw it clicked, because nobody on earth had ever seen such a thing before.

The 44% gets repeated constantly and it deserves less respect than it gets. It came from staff recollection, not an audited system, since there were no ad servers yet and counting meant combing server log files by hand looking for hits on the ad’s .jpg — as one Modem Media account man put it, the first web analytics tool was a highlighter pen. At least one contemporaneous account says 42%. And the “first” claim is soft too, since GNN was running banners two or three weeks earlier, a Silicon Valley law firm bought a clickable ad around the same window, Club Med and Zima and Volvo were all in the mix, and Prodigy had been sticking banners at the bottom of the screen years before that. AT&T on HotWired is the famous first. Famous and first are different words.

The magazine model, ported wholesale

The web’s first commercial instinct was to behave like a magazine. Space sold by the thousand impressions, hard-coded into the page, transacted on insertion orders, priced high because the inventory was scarce and the format was new and nobody had a way to argue with the number. CPM rents attention. It pays for exposure and stops there, which means the only lever a publisher can pull is more pages in front of more eyeballs, and that incentive ran straight through everything built in this period.

DoubleClick industrialized it. The thing traces to a Poppe Tyson agency division in April 1995 and got spun out and branded as DoubleClick in early 1996 under Kevin O’Connor and Dwight Merriman, headquartered in New York’s Silicon Alley (the 1995-versus-1996 confusion in most write-ups comes from collapsing those two events into one). By the end of 1996 they’d unveiled DART, for Dynamic Advertising, Reporting, and Targeting, which served and tracked and targeted ads across an entire network of sites rather than one page at a time. Roughly $6.5 million in revenue that year across about 250 clients. A database of around ten million user profiles inside the first year. Yahoo offered $95 million in 1996 when O’Connor wanted a hundred, and they wouldn’t budge, and DoubleClick went on to raise $62.5 million in its 1998 IPO and eventually sell to Google for $3.1 billion.

Summer 1996: selling out

Open Text, a leading engine at the time, started selling paid placements inside its search results. Preferred Listings. Clearly labeled, plainly disclosed, and the backlash on mailing lists and newsgroups was severe enough that the program died within a few weeks. Danny Sullivan’s framing is that the web was still so new, and commercialization still so novel, that it read as selling out.

Two years later GoTo did the same thing and built an empire on it. The difference was partly timing and partly that GoTo arrived as a fresh brand with no reputation to taint, which is worth holding onto going into Part 3. The idea was not rejected. The messenger was.

Portals, and a television show about them

Lycos, Excite, AltaVista, Infoseek, Yahoo. All monetized on banner CPM, all therefore incentivized toward page views rather than answers, all steadily piling on cruft — news, weather, horoscopes, stock tickers, free email, chat — because every additional click was another impression to sell. A search engine that sent someone away quickly was a search engine losing money. That is the entire reason Google’s blank white page felt like a revelation four years later.

Halt and Catch Fire built its final season on exactly this fight, and got it righter than most documentaries. Season four runs through 1993 and 1994 and pits Comet, a hand-curated web directory, against Rover, an algorithmic crawler, which is Yahoo against roughly AltaVista with the serial numbers filed off. Comet is warm and human and beautifully made and it loses, because Yahoo lands on the Netscape toolbar and distribution eats craft for breakfast. (The show is set in Dallas–Fort Worth, in the old Silicon Prairie, which is where I sit typing this, and I’ve never quite gotten over how thoroughly that fact goes unmentioned in the histories.) The finale ends in 1994 without ever showing what settles the argument, which is correct, because the thing that settled it hadn’t been built yet.

The decay

Rates above five percent were unremarkable early on. The nineties averaged somewhere around three percent, sliding to a range of 2.4% down to 0.4% by 2002, and modern display sits between roughly 0.05% and 0.46% depending on whose benchmark gets used. The 44% was never a benchmark. It was a novelty premium being paid out in a medium with no clutter in it, and novelty premiums amortize to zero.

Then it got a name. Jan Panero Benway and David Lane at Rice University published work in 1998 titled “Banner Blindness: Web Searchers Often Miss ‘Obvious’ Links,” demonstrating under controlled conditions that people hunting for specific information skipped banner-shaped elements even when the banner contained the exact thing they were hunting for. Not annoyance. Not irritation. Filtering that happens before the ad ever reaches awareness, triggered by shape and position alone.

Any format that works because it’s unfamiliar is a wasting asset. Novelty is not a strategy. It’s a countdown. — The History of PPC, Part 2

Which is the whole lesson of the era, and it’s a brutal one for anyone whose plan depends on a format being new. Attention is not created by buying it, impressions are infinite and getting cheaper by the month, and a pricing model that charges for exposure will always reward the publisher who manufactures more exposure regardless of whether anything happens downstream.

Carryover: the vocabulary that survived

Terms established in this era and the instruments they became.

CPM · 1994

Price paid per thousand ad impressions delivered, regardless of outcome. Still the buying unit for display, video, and YouTube; target CPM and viewable CPM bidding.

Impression · 1994

One instance of an ad being served onto a page. Impression share, search impression share lost to budget and rank, viewability standards.

Click-through rate · 1994

Clicks divided by impressions, expressed as a percentage. Still a Quality Score input and the primary creative diagnostic in every search account.

Ad server · DART, 1996

Software that delivers, rotates, counts, and reports on ads across sites rather than hard-coding them into pages. Google Ad Manager, DV360, third-party ad tags.

Ad network · DoubleClick, 1996

Aggregated inventory from many publishers sold to advertisers as one buy. Google Display Network, partner networks, open programmatic exchanges.

Insertion order · 1994

The signed contract committing an advertiser to a fixed volume of inventory at a fixed rate. Programmatic guaranteed and reserved buys; still the paperwork for anything not bought at auction.

User profile database · DoubleClick, 1996

Stored behavioral records used to decide which ad a given visitor sees. Audience segments, remarketing lists, customer match.

Preferred listing / paid placement · Open Text, 1996

Purchased position inside results that would otherwise rank on merit. The “Sponsored” label and every disclosure rule attached to it.

Standard banner dimensions · 468×60

Fixed creative sizes so inventory could be bought and swapped interchangeably. IAB standard sizes, responsive display assets, asset-level reporting.

Banner blindness · Benway and Lane, 1998

Learned filtering of anything shaped or positioned like an ad, before conscious attention engages. Creative fatigue metrics, ad schema avoidance, and the permanent performance gap between display and search.