The feed knows: 52,000 data points, Cambridge Analytica and the systems behind attention

Systems architecture · November 14, 2026 · 9 min read

Every scroll is an exchange. Attention goes in, behavioral data comes out, and a ranking system decides what appears next. Most people sense this. Few have looked at what the public record actually shows about how deep the exchange runs.

What platforms know

In 2016, ProPublica built a browser tool that let Facebook users see the categories the platform had assigned them for ad targeting. The investigation found more than 52,000 unique attributes in use, from hobbies and life events to inferred political leanings and “ethnic affinity.” It also found that Facebook supplemented its own data with information bought from commercial data brokers covering offline life: income bands, homeownership, the kinds of stores people shop at. Users could see some of these categories in their ad preferences. They could not see the broker data.

The number matters less than the structure behind it. A platform with billions of users does not need to understand any one person. It needs enough signals to predict which of several options a person will engage with, and then it needs to learn from the result. Every interaction refines the next prediction. The system is not a file on an individual so much as a constantly retrained model of what holds attention.

Likes as a window into personality

Academic work showed how much these signals reveal. A 2015 study in the Proceedings of the National Academy of Sciences found that a model trained on Facebook likes predicted personality more accurately than friends or family did. With about 300 likes, the model’s judgments matched a person’s own self-assessment more closely than a spouse’s did.

The personality model in question was the Big Five, or OCEAN: openness, conscientiousness, extraversion, agreeableness and neuroticism. It is the most widely validated framework in personality psychology, and it is exactly the kind of structure a marketer can map messages onto.

A 2017 follow-up in the same journal tested whether that mapping changes behavior. Across campaigns reaching about 3.5 million people, ads matched to a person’s inferred extraversion or openness produced up to 40 percent more clicks and up to 50 percent more purchases than mismatched ads. Psychological targeting works. The effect is real, measurable and modest in absolute terms: better odds, not mind control.

Cambridge Analytica: the scandal and the limits

In 2014, a personality quiz app collected data not only from the people who installed it but from their friends, under platform rules that allowed it at the time. Facebook later estimated that up to 87 million people’s data was improperly shared with Cambridge Analytica, a political consulting firm that marketed “psychographic” targeting to campaigns.

The consequences were large. The UK Information Commissioner’s Office fined Facebook £500,000, the maximum under the law then in force. In the United States, the Federal Trade Commission imposed a $5 billion penalty in 2019 and new privacy oversight requirements. Cambridge Analytica shut down in 2018.

What the scandal proved is narrower than its reputation. It proved that personal data moved without meaningful consent and that the rules governing it were weak. Whether the firm’s psychographic models changed election outcomes is far less clear. The ICO’s final report in 2020 concluded that its methods largely relied on commonly available technology and well-established techniques. Both the political left and right have used the episode to tell a story that suits them. The evidence supports a duller conclusion: the data practices were the real problem, and the persuasion claims were partly marketing of their own.

The systems behind the feed

The more consequential systems are the ones running every day without a scandal attached. Feed ranking, recommendation engines and ad auctions decide what billions of people see, and their internal logic is proprietary. Platforms publish principles and some transparency reports; they do not publish the models.

The public got an unusual look inside in 2014, when Facebook researchers published an experiment that adjusted the emotional tone of the News Feed for 689,003 users for one week. People shown fewer positive posts wrote slightly fewer positive posts themselves, and the reverse held for negative posts. The measured effect was small. The finding that a ranking change could shift mood across a population at all, without the participants knowing, was not.

Ad delivery adds another layer. Advertisers choose a target audience, but the platform’s delivery system decides who inside that audience sees the ad, optimizing for the outcome the advertiser selected. Two advertisers with identical targeting can reach different people because the algorithm predicts different responses. Much of what is called targeting today is really the platform’s own model at work.

Why this matters for marketing

None of this makes persuasion illegitimate. Direct response marketing has always matched messages to audiences and measured what worked. What has changed is scale, opacity and the distance between the person and the decision about what they see.

Three conclusions follow for any business that wants growth without compromising trust:

Earn data rather than extract it. First-party data from people who chose to engage, such as subscribers, customers and survey respondents, is more durable than borrowed signals. Privacy law and platform changes, like Apple’s App Tracking Transparency in 2021, keep shrinking the borrowed kind.
Use psychology to fit, not to exploit. Matching a message to how a segment thinks is good communication. Engineering urgency that is not real, or targeting vulnerability, is not. The line is whether the buyer would endorse the tactic if they saw it.
Understand the algorithm as a partner with its own goals. Ranking and delivery systems optimize for engagement and auction revenue. Content and offers that genuinely serve the audience tend to align with those goals over time; tricks tend to be patched out.

The systems behind attention will keep getting more capable. The advantage goes to marketers who understand them well enough to work with them honestly, and who build assets, like audiences and reputation, that no ranking change can take away.

Sources

All articles