No one would mistake Ben Cogan and Jesse Horwitz for "brogrammers", the jockish male coders swaggering across the tech landscape. Nor are they hustlers, the relentlessly outgoing types who quit their jobs to gamble on audacious ventures. They are two bookish friends, ages 27 and 29, who until recently lived across the street from each other in New York's Manhattan.
Horwitz worked for Columbia University's endowment fund; Cogan had a job analysing consumer behaviour. Their hobbies are quiet. Cogan dreams of earning a PhD in philosophy someday – "after all this is said and done," he says. Horwitz enjoys tracking various aspects of his life in Excel spreadsheets: restaurants visited, books read, jogs taken. Scrolling through those files, he says, fills him with a sort of data-based nostalgia. For years, the two men met for dinner every week or so, where talk often turned to business ideas.
One night in the northern summer of 2015, while dining at a Sichuan restaurant, Cogan asked Horwitz for advice about his latest notion: selling contact lenses online. The contacts business was dominated by a handful of multinational companies like Johnson & Johnson and Bausch & Lomb, which seemed to charge whatever they wanted – at least in Cogan's view, based on the price increases he noticed for his own lenses. Surely a low-cost competitor could tempt away customers?
Cogan pulled his laptop from his bag and opened it at the table in the middle of dinner, pushing aside plates of dumplings and spring onion pancakes. He had two plans to show Horwitz. They could sell a cheap disposable lens to doctors. Or they could mimic Cogan's employer, a wildly successful start-up called Harry's.
By late 2015, Harry's, which sold safety razors and shaving cream, was in the vanguard of upstart online retailers known as direct-to-consumer companies. The business model works like this: firms sell only their own products, only through their own websites. By cutting out retailers and distributors, they can charge less for their specialty products than entrenched competitors. Venture capitalists – convinced that consumers would increasingly patronise specialty online retailers as they grew more comfortable shopping online – were pouring money into direct-to-consumer start-ups, more than $US2 billion ($2.6 billion) over the past four years, according to CB Insights.
By February 2016, after many nights and weekends of emailing Asian manufacturers and reading up on Food and Drug Administration compliance (FDA, the US regulatory equivalent of Australia's Therapeutic Goods Administration), the vision of a viable business was coming into focus. The pair had found an FDA-approved manufacturer in Asia and figured out how to meet the necessary regulations. Still, Cogan was reluctant. He had been accepted to Wharton, the business school of the University of Pennsylvania, and had even put down a deposit. He believed that was the smarter option. At best, the contact-lens business would become a side project.
Before shelving their venture, they decided to try one more tack. They recruited two friends: Paul Rodgers, a buddy of Horwitz from Columbia who knew how to write computer code, and Dan Rosen, a friend of Cogan from the New York City selective high school, Bronx High School of Science. Together the four built what is known in the world of online retailing as a demand experiment. The technique, credited to Harry's founders (who give away its basic code), amounts to a two-page website. The first page explained the concept of a monthly subscription for disposable contact lenses and asked those who were interested to submit their email addresses. Visitors who did so were taken to a second page and were made an offer: share this referral code with friends, and if enough of them sign up, you'll get free contacts.
They posted a link to their site on the walls of about 40 Facebook friends. Within a few days, not only had their own friends signed up, but friends of friends of friends had, too – some 2000 people in all. Some of those distant connections were even evangelising the company on their own Facebook walls. "It went mini-viral," Cogan says.
He and Horwitz applied to tech incubators – organisations that invest in and coach young companies in exchange for minority stakes – using the demand experiment as one slide in their 16-page PowerPoint presentation. They pitched a few New York venture capitalists as well. They decided that if they were admitted to an incubator, they would work on the project full-time. If not, Cogan would go to Wharton. By April 2016, they had not only been called back for interviews with five incubators; venture funds were also offering to invest a total of $US3.5 million in their idea.
Cogan dropped his Wharton plans. He and Horwitz ordered 50,000 contact lenses and, with Rosen as creative director and Rodgers as chief technology officer, began working out of their investors' offices, stacking boxes and boxes of lenses along the walls by their desks. They eventually named their enterprise Hubble, after the orbiting telescope that can see into deep space.
Facebook helped them succeed with their demand test; now it would generate their first sales. During the middle of 2016, a friend of one of Hubble's prospective investors, a start-up veteran, recommended that the founders try a new type of Facebook advertising called Lead Ads. No outside website was needed: would-be customers simply clicked a button on the ad to submit their email addresses, directly from Facebook. Hubble directed these ads to postcodes in New York and Chicago, where they had already signed up optometrists willing to write prescriptions. After people clicked the ads, Horwitz emailed them to co-ordinate appointments and take their orders.
When Hubble's online store opened officially in November 2016, Cogan and Horwitz knew how to run a Facebook advertising campaign, and they were confident it would continue to generate sales. They planned to spend most of $US3.7 million on it.
Some implications of Facebook’s omnipresent, opaque, inhuman power are troubling. One investigation showed how its software gave advertisers the chance to target 'Jew Haters'.
In 2017, everyone seems to be wondering: is Facebook taking over the world? Most of us now realise that the social network has become far more than a repository for selfies and political rants of its more than two billion users. To ad sellers, Facebook is now a gluttonous monster, which, along with Google, is gobbling up the digital advertising business globally; according to Pivotal Research Group, the two companies controlled 70 per cent of the US market and most of the growth in 2016.
In considering Facebook's far-reaching influence, it's worth keeping in mind the perspective of the more than five million advertisers whose money is financing the social network's rampant growth. For them, Facebook and Instagram, which the company also owns, are the stuff of fantasy.
The process is easy, cheap and effective. With a few hundred dollars and a morning's effort, an entrepreneur can place his or her ads before social-media users that same afternoon. Companies unsure which ads are best can upload a handful of them and let Facebook's artificial-intelligence software test their efficacy. If they don't know who should see their ads, they can embed code on their websites that enables Facebook to monitor the traffic and then show ads to recent visitors.
Or companies can send the email addresses of their existing customers to Facebook, and it will locate their Facebook accounts and put ads in front of so-called Lookalikes, users who like and click on the same things that your proven fan base does. It's all about as straightforward as setting up an online dating profile. Steph Korey, a founder of Away, a luggage company based in New York that opened in 2015, says that when the company was starting, it made $5 for every $1 it spent on Facebook Lookalike ads.
One afternoon in March, I watched as Rosen selected three new ads from an extensive photo shoot the week before, his third in four months. Rosen resembled a sleep-deprived new parent: mussed hair, dull gaze. He spoke in a monotone. He attributed his fatigue, I would learn later, to Facebook's artificial-intelligence software that placed Hubble's ads. Rosen and his colleagues simply referred to it as "the algorithm".
The basic building block of Facebook advertising is an "ad set". It consists of the ads themselves and choices in three other categories: audience, goal and budget. That day, Rosen was designing a set to reach an audience of people on Instagram who had visited Hubble's website in the past 30 days. His goal was "conversions", or persuading users who had seen the company's ad to make a purchase. Finally, he set a budget of $US1000 per day. He uploaded the three images. Now they were ready to be tested, to see if any of them were winners in the eyes of users and the algorithm.
What happened at 8am the next morning, when the ad set became active, was complex – and far removed from human sight.
Just before Facebook places an advertisement in a user's feed, it holds a sort of instantaneous auction to determine which advertiser gets the space. The amount of each advertiser's bid is influenced by its budget size, of course, but the algorithm also weighs what it knows about the company, the ad and the individual Facebook user.
The process is never the same twice. The algorithm is constantly learning, using past results to inform how it weighs bids in the next auction. The intent, Facebook says, is to maximise value for everybody: to pair the advertiser with its likeliest customers, and to show ads that users want to see. And, of course, to make money for Facebook.
But from Rosen's perspective, nothing much had happened before he ambled into the office a little after 10am. Facebook had spent a grand total of $US1.86 (of his daily $1000 budget) on his ads. It had shown the first ad to 51 people, the second to 45 and the third to only two. The first ad had been clicked once. Rosen, unperturbed, poured himself a cup of coffee from the single-serve machine. The algorithm takes a little while to get warmed up, he said. "In an hour, it'll get exciting."
Twenty minutes later, Rosen refreshed his browser. The Ads Manager window displayed the latest numbers: Rosen could see only the results, not the process that produced them, but it seemed as if the click had inspired the algorithm to favour the first ad. As Rosen refreshed his browser, the sensation was like watching a seed sprout. The ad got more views. Some led to clicks. And eventually, sometime between 11:28am and 11:53am, one of those clicks led to the test's first sale. Commerce was in bloom.
Facebook's sales pitch – putting the right ad in front of the right person, thanks to the wonders of data technology – isn't exactly new. What sets Facebook and Google apart are scale and sophistication.
What also sets Facebook and Google apart from their direct-marketing forebears is that they give access to everyday advertisers. Anyone with a credit card can go online and test ads on Facebook's platform, one of the most sophisticated direct-marketing operations ever. But while average people can use the machine, there's still a lot of mystery about how it works. The methods and calculations of the algorithm – why it ends up pushing some ads and not others – are all hidden.
Almost as soon as they began, Rosen, Horwitz and the others at Hubble became determined to fathom the algorithm's secrets – to figure out why some ads succeeded and others didn't. Soon they were trading hypotheses with other entrepreneurs, cribbing ideas from other companies' ads and taking a formal approach to testing, rooted in the scientific method.
But even as the Hubble team gleaned more about what yielded successful Facebook ads, the algorithm could be unpredictable, almost moody. If you kept loading the same ads into the same ad set every day, they stopped performing as well. The founders figured at first that users were tiring of the same ads – but actual viewer numbers revealed that practically no individual user had seen any ad more than once. The algorithm itself seemed to grow bored.
At night, meanwhile, the algorithm spent lots of money and rarely found customers. The Hubble executives started shrinking the budgets at 11pm, which they called "putting the algorithm to bed". The algorithm could also be impulsive and streaky – some days it might go on a sudden jag, blowing a thousand dollars in a few hours with nothing to show for it. Rosen found himself checking the Ads Manager compulsively on his laptop and his iPhone. "It occupies my brain constantly," he says. "It's that feeling of, 'Did you leave the oven on?' "
One evening, we went to a stand-up-comedy night. (For the past four years, Rosen has moonlighted as a comedian.) But while the other comics were onstage, Rosen was on the Ads Manager the whole time. "Who cares about jokes?" he quipped afterwards.
Eager for help, he sought guidance from a former Facebook employee named Faheem Siddiqi, who now runs his own marketing agency. Hubble's sales representative at Facebook told him that Siddiqi had figured out the best ways to optimise Facebook advertising campaigns. But it turned out that Siddiqi and his employees checked the Ads Manager even more compulsively than Rosen.
Middlemen – creative agencies, media planners, publishers – have long ruled the advertising business. Yet until recently they have not been as omnipresent, opaque and inhuman as Facebook. The social giant now dictates, more fully and precisely than ever before, which ads we see and who sees which ads. Some of the implications of this are amusing, others troubling.
Recently ProPublica, the investigative-journalism non-profit, showed how bad actors can abuse this process: Facebook's software gave advertisers the option to target "Jew Haters", for instance. In a separate investigation, ProPublica found that Facebook made it possible to exclude specific "ethnic affinities" from seeing ads, noting that ads excluding people based on race are prohibited by American housing and employment laws.
This stereotyping isn't a glitch of Facebook's machine-learning process; it's how the software works. To formulate audiences, the algorithm scours profiles and analyses them for shared traits and correlations and self-identified interests and, it assumes, our preferences, grouping us into tribes that can be targeted. It's up to Facebook and advertisers to constrain this amoral process in ethical and lawful ways.
Facebook's artificial intelligence isn't operating unattended. Facebook points out that it makes efforts to prevent harmful advertising. Yet managing a platform this way – seeing what mischief the algorithm and its users gets up to, then responding with countermeasures – can be difficult to sustain. It makes Facebook, a company still largely controlled by a single man, Mark Zuckerberg, the ultimate arbiter of morality and taste for all two billion of its users. It also means the company has unilateral power to make or break companies when it tweaks its system.
As we delegate more control to artificial intelligence, both businesses as well as users are venturing into uncertain territory. In the meantime, more and more companies – start-ups, mum-and-dad stores, major corporations – are handing their dollars and their data to the social-networking giant. Facebook's Ads Manager is user-friendly. Sales are plentiful. And if you don't take advantage of it, your competitors will. How could you not go there?
By mid-March, the Hubble team no longer had 15 Facebook and Instagram ad sets. It had 40 – all pointed at discrete audiences, each with its own handful of ads. But Rosen looked more rested, less frazzled. He explained that he and Paul Rodgers had developed something they called "Robo-Dan", a few lines of code that checked the Ads Manager every hour, then adjusted the budget as Rosen would.
But by the end of June, a new problem arose: no matter what new ads they put in an ad set, the growth rate of sales declined and the cost per acquisition went up. They began to think it was an audience problem: had they found all the customers in those groups? As they searched for more and more audience descriptors, they landed upon a novel idea: they began trading their Lookalike groups with other online retailers, figuring that the kind of people who buy one product from social media will probably buy others. This sort of audience sharing is becoming more common on Facebook: there is even a company, TapFwd, that pools together Lookalike groups for various brands, helping them show ads to other groups.
Cogan and Horwitz have decided that they need to reduce their dependence on Facebook advertising, for the sake of their business and their own sanity. In May, they tested their first 15-second cable-television commercials. Even though the old medium provides them with less information than Facebook, in some ways the ignorance is bliss. "There's fewer levers; there's less to stress out about," Rodgers says. "You can push the button and get on with your life."
In August, the Hubble team finally handed over their domestic Facebook advertising work to an outside agency, Ampush, that charges them based on how many new customers sign up. "We ran their numbers – it's something we could beat," Rosen says, meaning Hubble could get more customers for less money if it did the ad buying in-house. "But it would destroy our lives."
Thanks largely to Facebook, Hubble is on track to finish its first full year in business having made $US20 million in revenue. In August, Hubble raised $US10 million, valuing the company at $US210 million. In January, Hubble will use those funds to expand its business to Europe. Its advertising strategy? Robo-Dan, with some help from Rosen. As Hubble advances into new territories, Facebook and the algorithm will be tagging along with them.
Edited version of an article first published in The New York Times Magazine.