When the Feed Becomes a Sales Funnel

At 11:47 p.m., you open a video app to watch one repair clip. Ten minutes later, you are angry at a stranger, worried about retirement, and looking at an ad for a trading course. Nothing supernatural happened. A recommendation system measured what you paused on, replayed, shared, or skipped, then used those signals to choose the next item. The feed felt like a stream of entertainment, but it was also a small experiment with your attention.

The internet did not become dangerous because every website suddenly turned dishonest. It changed shape. The early web was a collection of destinations: a personal page, a forum, a blog, a shop. You chose where to go. The modern web is increasingly a prediction machine that decides what to place in front of you before you have formed a clear intention. That distinction matters because a system optimized for keeping you present can learn to prefer whatever reliably produces a reaction, even when the reaction is dread.

The feed is a feedback loop

A recommendation system is software that predicts which piece of content you are most likely to watch, click, or otherwise engage with. A ranking model is the part that gives each candidate a score. It does not understand a video the way a person does; it estimates behavior from patterns.

An illustrative model might look like this:

score = (
 0.42 * predicted_watch_time
 + 0.23 * predicted_return
 + 0.18 * predicted_share
 + 0.12 * predicted_click
 - 0.05 * predicted_hide
)

Those weights are invented, not a description of any real platform. The point is the direction of travel. Here, predicted_return means the chance that you come back to the service, not a financial return. If a frightening clip keeps people watching, sharing, and returning, its score can rise even if the system has no concept of fear or truth. An A/B test—an experiment that compares two versions with different groups of users—then helps the service learn which arrangement produces more of the desired behavior.

TikTok publicly describes its For You recommendations in similar behavioral terms: likes, shares, follows, comments, video details, and watch behavior all provide signals, with completing a longer video treated as a stronger sign of interest than a weak contextual match. That design can help a new creator find an audience. It can also reward material that holds the eye by provoking alarm or envy. (newsroom.tiktok.com)

No engineer has to write a rule saying make people miserable. The pressure comes from the objective function, meaning the measurable target the system is trained or tuned to improve. When attention is the target, emotional intensity becomes a useful shortcut. Calm accuracy asks for patience; outrage offers an immediate click.

Why the checkout button keeps appearing

A sales funnel is a sequence that moves someone from first exposure to a purchase or signup. Online platforms are unusually good at building funnels because discovery, targeting, messaging, payment, and retargeting can happen in one connected system. The ad or post does not need to convince everyone. It only needs to find the person whose insecurity, loneliness, health worry, or money problem makes the pitch land today.

Why do online scams spread so well on social media? The answer is partly the same machinery that makes legitimate recommendations convenient: profiles reveal interests, interactions reveal attention, and cheap distribution lets a small operator test thousands of messages. The Federal Trade Commission reported that people in the United States reported $2.1 billion in losses from scams that started on social media during 2025, about eight times the 2020 figure. Investment scams accounted for $1.1 billion of those reported social-media losses. These figures are complaints, not a census of every victim, so they describe a floor rather than the whole room. (ftc.gov)

The handoff can be almost invisible. A person sees a post about financial freedom, joins a chat group, meets a friendly guide, watches a dashboard showing fictional gains, and is told that a larger deposit will unlock the real opportunity. The platform may not be running the fake investment site, but it helped assemble the audience and supplied the first moment of trust.

The customer becomes the distribution layer

This is where ordinary affiliate marketing needs a careful distinction. Affiliate marketing is a legitimate arrangement in which someone earns a commission for referring a customer to a real product or service, usually with a disclosure. The trouble begins when the product is mostly a promise of status, access, or future income—and the easiest way to earn is to recruit another promoter.

That creates a recursive business: a course teaches people to sell the course, a community sells access to the community, and a trading guru sells the dream of becoming a trading guru. People who have already paid may keep promoting because of sunk cost, the psychological pressure to defend money and effort already spent. The promoter is not always a confident predator. Sometimes the promoter is trying to make the story true after buying it.

The result is a strange collapse of roles. The viewer is a customer, a lead, a testimonial, and sometimes unpaid marketing staff. A referral code sits where a personal recommendation used to be. Friendship becomes a distribution channel.

Crypto, synthetic media, and the cheaper lie

Cryptocurrency did not invent online fraud, and not every crypto project is fraudulent. It did, however, provide a particularly convenient package for persuasion: technical language that is hard to verify, prices that move dramatically, transfers that can be difficult to reverse, and a culture in which public promotion is part of holding the asset.

The FBI’s 2025 Internet Crime Report recorded 61,559 complaints about cryptocurrency investment fraud and about $7.2 billion in reported losses. The common pattern is social contact, a private messaging group, a fake investment interface, visible profits that encourage larger deposits, and fees demanded when the victim tries to withdraw. (ic3.gov)

Artificial intelligence adds another layer. Synthetic content means computer-generated or altered text, voices, images, or video that imitate real people or events. The FBI said its 2025 data included 22,364 AI-related complaints with adjusted losses above $893 million. The important change is not that machines suddenly learned how to deceive; it is that convincing, personalized deception can now be produced at much lower cost. That makes the one-to-one scam feel industrial.

Designing an exit from the loop

The answer is not to ban recommendation systems. They help people find obscure music, repair advice, communities, and useful products. The better question is what the system is rewarded for, what it is allowed to infer, and where it must slow down.

A healthier design would:

• measure successful tasks and long-term satisfaction, not only minutes watched;
• add friction before high-stakes purchases, investment transfers, or invitations to private groups;
• make sponsorships, affiliate relationships, and synthetic media hard to miss;
• offer a chronological feed, meaning posts ordered by time rather than a profile of your behavior;
• track scam exposure and repeat targeting as failures, not as acceptable side effects.

Regulation is beginning to push in this direction. Under the European Union’s Digital Services Act, designated very large platforms and search engines—those with more than 45 million monthly users in the EU—must assess systemic risks, provide more information about recommender systems, and give users more control over personalized recommendations. That does not make a feed wise or kind, but it treats algorithmic design as a public responsibility rather than a private detail. (digital-strategy.ec.europa.eu)

The old web felt human partly because it was made of places. The new web is made of predictions, and predictions can be tuned toward care or extraction. The technical challenge is not removing persuasion from the internet; persuasion is part of communication. It is refusing to build a machine that notices your fear, finds the highest-bidding seller of that fear, and calls the transaction engagement.