ADINT: advertising data as reconnaissance¶

ADINT is the use of the advertising industry’s own plumbing as an intelligence source. Researchers at the University of Washington proposed the term in 2017, and the market that supplies it now sells finished interfaces with a login and a map rather than raw files. The structural appeal is that the material is bought rather than probed. Nothing about acquiring it touches the estate it describes.

The unit is not a person. It is a device identifier with a timeline attached: the advertising identifier Apple and Google issue, a coordinate, a timestamp, and no name. Everything useful comes from reading the timeline rather than the record.

Reading a file¶

An identifier that rests at one address overnight and spends weekdays inside a fence has stated a home and a workplace without either being written down. Repetition turns that into rhythm: shift boundaries, who is present at three in the morning, which days a site runs thin, when a building empties. Two identifiers that keep appearing in the same places at the same times imply a relationship, though whether that is a car share, a team or a household is a guess until something else confirms it. Journeys expose satellite offices, suppliers, and the distance a person will travel to keep a habit.

None of this is asserted by the data. It is inferred from repetition, which is why a read strengthens the longer a file runs and decays as the file ages.

Feeding the graph¶

Pattern of life arrives before access and fills in parts of the identity map that normally wait for a foothold: which people exist, where they are, and when they are somewhere else. Candidate selection improves in the same move, because a file distinguishes the engineer who works nights from the manager who never does, and the contractor who attends three sites from the employee who attends one.

The pattern worth looking for is a device carrying two organisations’ rhythms, an integrator’s office and a plant, a supplier and a client. Third-party access is usually invisible from outside until somebody’s schedule betrays it, and here it betrays itself. For pretext work the yield is different again: specific, checkable detail about a person’s week, which is what lets a call survive its first awkward question.

Failure modes¶

Coverage is uneven and largely accidental. It depends on which apps a person carries and how those apps behave, so an estate can be well represented in one file and nearly absent from the next.

Files are snapshots rather than feeds. Samples in circulation have covered a single day, or a couple of months, which supports rhythm analysis and not live tracking. Accuracy varies inside one file, from records vague by kilometres to records good to the metre, and the vague ones look identical to the precise ones until plotted.

Volume flatters quality. In the German file the reporting noted that some timestamps and identifiers were themselves erroneous, which inflates any count of distinct devices.

A device is not a person. Handsets get shared, replaced, and left at home, identifiers get reset, and opt-outs zero them entirely, so continuity breaks without announcing itself. Absence proves nothing: where an organisation takes phones at the door, the data stops at the door while the person carries on inside.

The join is where identification happens and where it goes wrong. In the Norwegian case reported in May 2025, a dating app’s coordinates were too coarse to place a man while a messaging app carrying the same identifier led to his front door. The identification came from the second dataset, and so would a misidentification.

Price, and where the trace sits¶

The market is cheap by intelligence standards. A German file of 3.6 billion location points reached journalists as a free preview, against a continuous stream advertised at around 14,000 dollars, and the tooling above it is a product: one vendor’s interface reads across up to 500 million devices, with an American immigration authority and government bodies in Hungary and Austria among its known users.

The purchase leaves no trace on the target estate, which is the whole attraction, but it is not traceless. The record sits on the buyer’s side, in contracts, invoices and vendor logs, which is how most known purchases came to light. The legal ground is contested rather than settled: none of Germany’s sixteen state data protection authorities could name a legal basis for police use of commercial location data. For an authorised engagement, buying brokered data about a client’s staff normally sits outside scope, and that is worth settling in writing rather than in a debrief.

Read alongside public-record correlation, the difference is where the work happens. Tenders and job advertisements have to be combined before they say anything. A location file has already been joined by the market, and the reading starts at the timeline.