THE ALGORITHMIC ADDICTION EMPIRE
An advertising-funded platform is often described as selling your attention. That is close enough to be misleading. What is actually sold is the predictability of your attention: an advertiser is buying a forecast that a specific person will be present at a specific moment in a specific state of mind. Predictability is not produced by making something enjoyable. Enjoyment is erratic. Predictability is produced by habit.
This matters because it means nobody in the building has to want you to be compulsive. Compulsion is simply what an objective function selects for when the objective function is session frequency and the search space is interface design. The mechanism does not require an intention, which is precisely why it is durable, and why the usual argument about whether anyone meant it is the least interesting question available.
The mechanism
Revenue for this business decomposes into four multiplied terms:
Revenue = users x sessions per user per day x ad slots per session x price per impression
Three of those four terms are product decisions rather than market conditions. User count is bounded by population and competition. Ad slots per session are bounded by tolerance, and raising them visibly degrades the product. Price per impression is set by the advertising auction, which the platform influences but does not control.
Sessions per user per day is the term with the most headroom and the least visibility. It is not capped by population, it is not capped by taste, and unlike ad load it can be increased without the user perceiving that anything was done to them. That is where optimisation goes, and it goes there for structural reasons, not cultural ones.
The techniques are not secret and have never been secret. Variable-ratio reinforcement, in which reward arrives on an unpredictable schedule, is one of the most replicated findings in behavioural psychology and long predates the internet; it produces higher response rates and greater resistance to extinction than fixed schedules. A feed that sometimes contains something excellent and usually does not is a variable-ratio schedule with a scroll gesture as the lever. Infinite scroll, autoplay, streak counters, batched notification timing and unread-count badges are all documented in publicly available product and design material. None of this is concealed. It is taught.
The arithmetic
Model a platform with 100,000,000 users in one market. Assume 8 sessions per user per day, 6 ad slots per session, and a price of 9 EUR per thousand impressions, which is 0.009 EUR per impression.
- Daily impressions: 100,000,000 x 8 x 6 = 4,800,000,000
- Daily revenue: 4,800,000,000 x 0.009 = 43,200,000 EUR
- Annual revenue: 43,200,000 x 365 = 15,768,000,000 EUR
Now change one term. Raise sessions per user per day from 8 to 9, an increase of 12.5%. Because the terms multiply, revenue rises by the same 12.5%.
| Quantity | At 8 sessions/day | At 9 sessions/day | Change |
|---|---|---|---|
| Daily impressions | 4.80 bn | 5.40 bn | +0.60 bn |
| Annual revenue | 15.768 bn EUR | 17.739 bn EUR | +1.971 bn EUR |
| Annual revenue per user | 157.68 EUR | 177.39 EUR | +19.71 EUR |
One additional visit per user per day is worth 1.971 bn EUR per year on these assumptions, or 19.71 EUR per user per year.
Now price the intervention that produces it. Assume one extra daily session is prompted by one additional notification, and assume the marginal delivery cost of a notification is 0.00002 EUR.
- Notifications per year: 100,000,000 x 365 = 36,500,000,000
- Annual cost: 36,500,000,000 x 0.00002 = 730,000 EUR
- Return ratio: 1,971,000,000 / 730,000 = 2,700 to 1
The return on interrupting a person is roughly 2,700 to 1. At that ratio, no conspiracy is required and no villain is necessary. Any organisation that measures anything will find this, and any organisation that finds it will do it, and every quarter in which it does it will look like a quarter of ordinary competent product work.
What the same numbers look like from the other side
The four-term equation has a mirror image, because impressions are made of minutes.
Assume an average session length of 4.5 minutes. Then 8 sessions per day is 36 minutes per user per day, which is 36 x 365 = 13,140 minutes per year, or 219 hours per user per year.
Across the modelled user base: 100,000,000 x 219 = 21,900,000,000 hours per year.
A waking year, at 16 waking hours per day, is 16 x 365 = 5,840 hours. So the platform’s annual attention intake is:
21,900,000,000 / 5,840 = 3,750,000 waking human years per year
Set that against the revenue computed above:
15,768,000,000 / 3,750,000 = 4,204.80 EUR per waking human year
That is the clearing price. It is not an accusation, it is a quotient, and it can be recalculated by anyone who disagrees with the inputs. The interesting property of the figure is not that it is low. It is that it is a price at all, arrived at with no negotiation, by a party who was not told they were selling.
The strongest counter-argument
Three objections deserve to be taken seriously.
The first is revealed preference. People use these products because they like them, and treating voluntary use as capture is condescending. Time spent is genuine evidence of value received, and the same 219 hours could describe a beloved hobby. A model that cannot distinguish a compulsion from an enthusiasm has not proved anything about either.
The second is that there is no addiction dial. The ranking system that surfaces compulsive content is the same system that surfaces a message from a relative, a job opening, or a functioning small business. It is one optimisation, and it cannot be tuned to keep the good half. Demands to “remove the addictive part” frequently amount to demands to make the product worse at the thing users came for.
The third is displacement. Attention released from one medium does not return to the reader; it migrates. Broadcast television absorbed comparable per-capita hours for decades under an advertising model with the same four-term equation and a far cruder ranking function. If the objection is the equation, the objection is older than the industry accused of it.
These are strong arguments. They do not disturb the arithmetic, and they are not meant to. They bear on what should follow from it.
Consequence, and a falsifiable prediction
Because the four terms multiply, a regulation’s bite equals the term it touches.
Rules about data collection and targeting act on price per impression, one term among four, and only partially: a less precisely targeted impression is worth less, but it is still sold. Rules about interface mechanics act on sessions per day, the term with the largest headroom and the direct multiplier on revenue. A rule that removes one session per user per day removes 12.5% of revenue in the model above, immediately and without substitution.
That yields a prediction that can be checked rather than argued: advertising-funded platforms will settle privacy and data-governance rules comparatively quickly, and will litigate interface and session-shaping rules, default-off notifications, mandatory session breaks, and age-gated autoplay, considerably harder and longer, even where the privacy penalties are nominally larger. If the opposite pattern is observed over several years and jurisdictions, the model in this note is wrong and should be discarded.
The empire in the title is not a metaphor for scale. It is a description of a revenue model that treats a finite, non-renewable, individually owned resource as an input, buys it without a transaction, and reports the result as engagement growth.