The Transit of a Person

At 7:14 in the evening someone opens a weather app, and in that instant a record comes into being. It's a record of the opening of the webpage, and it carries the time, accurate to the second. It carries the identifier that the device's operating system keeps for advertising, a string of thirty-six characters that doesn't spell out a name but belongs to that one device and no other. It carries the device's model and the version of its software, the network address the phone is using at that moment, and the name and version of the app that produced it. It also carries a session marker, a note that this web-page opening is one action in a continuous run of actions by the same device.

What the record doesn't carry is worth stating plainly. It doesn't carry the person's name, and it doesn't name the city they were checking the weather against. What it carries instead is smaller and more particular: that at 7:14 in the evening this person, on this device, reached for something. It records the reaching in the terms a machine uses for events (when it happened, which device, which network, which app) but the reaching was theirs, and the record created is a fragment of their evening before it's anything else.

That's the whole of it at the start. Everything that later comes to be said about this fragment of data is said by something other than the fragment itself.

The app doesn't keep the data fragment to itself, and the reason why is built into how the app was made. An app on a phone or computer isn't a single piece of software from a single company; it's an assembly, in which the code that draws the forecast for a user's data is surrounded by smaller components supplied by other firms. These components are software development kits, or SDKs, included on the device so the app's maker can measure how the app is used, see where its users came from, and then sell the advertising space inside it. The SDKs run inside the app without belonging to it, and they were already present when the user first turned the device on.

So, the data fragment doesn't travel to just one place. Within the first second, copies of it (or particular slices that each component of the device is built to read) leave the phone for the servers of companies that the user has no relationship with: the analytics and advertising arms of Google and Meta, and attribution services such as AppsFlyer and Adjust, whose job is to log which advertisement someone saw BEFORE installing an app. The app's own maker receives a copy, and so does each of the other firms.

This is the first division in the fragment's custody. Until this point, it had rested in one place, held by the party the person chose to open. Afterwards, the same data fragment of that person's evening scroll exists in several places at once, held by parties the person didn't choose to be involved with and will never see. None of them asked the person permission, and none of them was required to, because the transfer is simply a property of how the app was assembled.

By the time the fragment reaches this stage, it still holds only what it held at creation. There's a device identifier, a location worked out from the network address, a timestamp, and the technical description of the phone. There's no name, no income, no religion, no medical history, and no statement of intent. Everything of that kind is about to be added, and the firms that add it describe their work in measured language.

In their own words, the companies in this part of the data chain present themselves as identity and audience specialists. They describe their role as resolving scattered signals into a single, privacy-safe view of a customer; matching records through hashed and pseudonymized identifiers so that no raw personal detail is exposed; and assembling what the industry calls 'audience segments.' This is so that a marketer can reach the right audience. The location data, they say, is precise, and the supply is consented. Described this way, the work sounds like ordinary infrastructure: loose observations, carefully joined, and responsibly handled.

The record of what the work involves in practice is held by the Federal Trade Commission (FCC). In a 2024 action, the agency found that the broker Mobilewalla built and sold audience segments (assembled out of raw location data) that included pregnant women, Hispanic churchgoers, and members of the LGBTQIA+ community. In a parallel action, the FCC found that Gravy Analytics, and its subsidiary Venntel, used geofencing to identify and sell lists of people who had visited healthcare facilities and places of worship, and then sold further lists of data tying named individuals to sensitive characteristics such as health or medical decisions, political activities, and religious viewpoints, all inferred from where a phone had been. The same company, the FCC found, continued using the location data after learning that the people it came from had not given informed consent. Its position data was advertised as accurate to roughly one meter.

None of this required a person's name to begin with. A related trade in the same area had already shown how little it costs: in 2022, reporters and members of Congress documented a broker selling data on visits to more than six hundred Planned Parenthood locations over a single week for just over a hundred and sixty dollars, and intelligence agencies were later found to be buying the same category of location data from brokers without consent. The buyer didn't need to know who anyone was. The pattern of a device returning each night to one address, and appearing one afternoon at a clinic, was enough.

This is the seam the rest of the page turns on, so it's worth stating without ornament: The record observed where a person's phone had been; It did not say she was pregnant; It did not say she was Catholic, or gay, or in treatment, or undecided about a pregnancy. Other companies decided those things about her, from the places her phone had gone, and sold these inferred decisions under her own identifier as though she had disclosed them herself. The observation and the interpretation are different things with different authors. The system files the observation and the verdict in the same row, and sells them together.

The data fragment is now in motion in a second way, and this is the part that happens fastest and is seen least.

When the weather app loads, the space for an advertisement inside it isn't sold in advance. It's sold in the moment the screen draws, by auction, while the person waits the half-second for it to fill. The app's supply-side platform packages the opportunity into a bid request and broadcasts it; not to one buyer, but to many at once. Into that bid request goes what the data fragment has become: the advertising identifier, the device and its software, the network address, the location worked out from it, the kind of content being viewed, and, increasingly, whichever segments have already been attached to that identifier by the firms in the previous stage. The request is an offer to sell access to this person, described in these terms, to whoever bids highest.

The buyers are demand-side platforms acting for advertisers (The Trade Desk, Google's buying platform, Amazon's, and others) and they don't deliberate, because there's no time to deliberate. In roughly the time it takes the page to finish loading, each one reads the segments, matches them against what its advertisers are trying to reach, and returns a price. The exchange takes the highest bid, and the winner earns the right to place something in front of the person. The person sees an advertisement appear. They don't see that an auction occurred, who was invited to it, what was said about them in the bid request, or what they sold for.

What they sold for has a name, and underneath the name is a calculation the buyer is making about the person. The price is quoted as CPM (cost per mille, the cost to reach a thousand people like them) and the buyer arrives at what it will pay by weighing what the advertiser gains if the person acts against how likely this particular person is to act, spread across the thousand impressions the quote is priced over.

The first term, 'the value of the action', is what the advertiser stands to gain if the person does the thing: buys the car, takes the loan, books the treatment, fills the prescription. The second, 'the chance the person takes it', is the buyer's estimate of how likely this person is to do it, read off the segments attached to their identifier. A car sale is worth more than a glance at the news, so a person that the segments mark as shopping for a car is worth more to bid on; not worth more as a person, but more likely, in the buyer's estimate, to produce an action the buyer can bank. This is why the reaching described above in stage zero matters here: the price is a number a stranger puts on a prediction about a human being.

And the auction has a residue the person is never told about. The bid request was broadcast to many buyers, and only one won, but the others received it too; the same identifier, the same location, and the same segments. In its 2024 action against Mobilewalla, the FCC found exactly this: the company sat in the bidding exchanges and kept the information out of bid requests it didn't win, building a store of more than five hundred million advertising identifiers paired with precise location, harvested from auctions it lost. It was the first time the agency had treated collecting from a losing bid as an unlawful act. Losing the auction didn't mean not receiving the person. It meant receiving the person for free. The thing being auctioned was a prediction about someone's next act; the thing that quietly accumulated, in the files of everyone who showed up to bid, was the someone.

The auction settles a single impression in a fraction of a second, but the price it settles isn't invented fresh each time. Behind the moment of bidding sits a slower, steadier judgment: what a person is generally worth, held as a number and updated as the segments attached to them change. The auction reads that number. The stage before it built that number. This is where the page has to be careful, because the number is real, the firms that maintain it are precise about it, and the precision is the point.

The number isn't uniform, and the firms don't pretend it is. Reaching a thousand people who were reading the news costs a buyer a few dollars. Reaching a thousand people the segments mark as in the market for a particular thing costs many times that, and the multiple tracks the thing. The categories that command the most are the ones where a single action is worth the most to the buyer: a person tagged as shopping for a vehicle, as carrying or seeking debt, as managing a health condition, as approaching a major purchase. The estimates are published, compared, and bid against. A person is not priced once. They are priced continuously, by category, and the price moves as the system decides they have moved closer to or further from an action someone is willing to pay to influence.

It's worth being exact about what that means for the person, because the system is exact about it and the exactness is easy to miss. The person reading the news is cheap not because they matter less, but because nothing the system expects them to do next is worth much to a buyer. The person managing a diagnosis is expensive not because they matter more, but because what they may do next (fill a prescription, choose a treatment, switch a plan) is worth a great deal to whoever sells those things. The value the market assigns isn't a measure of the person, it's a measure of the action the market predicts the person is about to take, and the money that action would move. The person never sees this number, never agrees to it, and cannot contest it. It's set entirely by parties estimating what they can get from the person's next move.

And the number compounds, which is the part the person would least expect. A single observation is cheap; a history is not. The longer the identifier persists and the more places the device goes, the more confidently the segments can be drawn, and the more confidently they are drawn, the more the person is worth — because confidence is exactly what the buyer is paying for. The buyer is paying to be less wrong about what the person will do. So the person's own accumulated life is the thing that raises their price. Every additional day observed makes them a more valuable thing to sell, and none of that added value returns to them. It accrues to the firms holding the history, and to the buyers willing to pay more for a surer prediction. The person is the source of the appreciation and the only party in the chain who does not share in it.

The data fragment's working life as an auction signal is brief, but the fragment itself doesn't expire when the auction closes. It's retained. The previous stage showed it landing in the files of every buyer who received the bid request, winner and losers alike. This stage is what those files become: not a record of one evening, but an inventory: held, copied, combined, and offered again.

This is the part of the chain most people don't know exists, and it's also the least dramatic, which is part of why it persists. The fragment is now a line item. It's matched against other fragments carrying the same identifier (the same person, observed elsewhere, on other days) and the matched set is worth more than any single piece, because a pattern sells for more than a moment. The combined set is licensed to other firms, which merge it with what they already hold and license the result onward. One broker sells a segment built from a slice of location history; another buys the segment and joins it to purchase records; a third scores the joined file for the likelihood of some action and sells the score. At each step the person is further from the transaction, and the thing being sold is more refined, more confident, and more completely a stand-in for them.

None of these transfers requires the person, and none of them notifies the person. The identifier moves between companies the way a part number moves between warehouses. The same federal action that documented the harvesting documented the resale: Mobilewalla, the Federal Trade Commission found, sold access to this raw, un-anonymized information to third parties including advertisers, other data brokers, and analytics firms. The fragment didn't rest; it circulated. A data fragment created in one evening can be bought and sold for years, in rooms the person will never enter, describing a person who may since have changed, forgotten, or moved on.

Eventually the fragment, and the larger thing it has become part of, is used. This is the stage the whole chain exists to reach, and it's the one the person is most likely to mistake for the entire system, because it's the only part they can see: the advertisement that appears, the offer that arrives, the price that is quoted to them.

But the use is wider than advertising, and this is the point at which the page has to widen with it. The assembled file (the segments, the scores, and the history) is read in order to make decisions about the person. It sets which advertisement they see and what is bid to reach them, but it also informs which offers they are shown and which they are not, what interest rate or premium a model proposes for them, which content a system places in front of them to hold their attention, and which risks an institution attaches to their name. Advertising is the visible surface. Underneath it's a broader practice of deciding things about people from the representations built out of their behavior (pricing, ranking, screening, predicting) and acting on the decision before the person is aware a decision was available to be made.

The person experiences only the output. They see the advertisement, or the price, or the recommendation, and not the file that produced it, the segments inside the file, the firms that assembled them, or the auction that placed it. The representation acts in the world on the person's behalf without the person's involvement, sometimes to the person's cost, and the person has no access to it: cannot read it, correct it, or refuse it. It speaks for them in rooms they're not in, and they're bound by what it says.

The moment that created the fragment ended long ago. The person closed the weather app within seconds and forgot they had opened it. The fragment didn't end.

This is the property that separates the representation from the person it represents: the person is bounded in time and the representation is not. The person forgets; the file remembers. The person changes (moves, recovers, reconsiders, becomes someone the old segments no longer describe) and the file persists in its earlier description, copied into systems that never receive the correction, replayed in decisions long after the behavior that produced it stopped being true. A person who's no longer shopping for a car is still, in some file, a car-shopper. A person who has recovered is still, in some file, managing the condition. The representation drifts free of the person and keeps acting, and the person cannot call it back, because they don't hold it and often don't know it exists.

This is also where the full shape of the thing becomes visible, if it hasn't already. What's been moving through this chain was never a single record of a single evening. It was the front edge of an accumulation; every fragment carrying the same identifier, gathered and matched and modeled over years into a persistent, predictive, acting version of the person, held entirely by others. The system recorded the pieces as separate data and treated them as detachable from the person who produced them. They are not separate, and they are not detachable. Assembled, they're a continuation of the person into places the person cannot enter, and it's this continuation (not any single fragment) that the chain has been pricing, selling, and acting upon from the start.

PDD prices per browsing event. The founding baseline is one cent per event, anchored to what the audience-targeted programmatic market pays for high-intent categories — roughly $10 per thousand impressions, observed May 2026. Ten dollars per thousand is one cent per event. A person's price never falls below the floor:

Of every cleared contract, 60% goes to the person, 30% to PDD for clearing, settlement, and enforcement, and 10% to buyer-side processing.

Nine categories carry reference rates, frozen as observed May 2026, shown per thousand impressions: Managing money, $10. Health plans, $10. Looking for work, $10. Big purchases, $8. Planning, $7. Learning, $6. News, $4. Streaming & TV, $4. Food & cooking, $4.

The references are comparison; the floor is the mechanism — and the floor only rises as real PDD sales clear.