Somewhere in the taxonomies that online advertising runs on there is a species of label called the in-market segment. In-market for an SUV. In-market for a mattress. In-market for pet insurance. The label means the machine believes you are about to spend money on something specific, and it is among the most valuable things the machine can believe about you. It will form that belief about anyone, from whatever signals are lying around, for any product there is money in — including the ones nobody wants to buy.
Funeral services advertise. They always have — the local paper used to be full of them — and someone searching “funeral homes near me” at two in the morning is exactly the person a funeral director needs to reach. The auction that decides what that person sees runs at the same speed and with the same indifference as the one selling trainers. “In-market for a funeral” sounds like a bleak joke about this industry. It is closer to a line item.
That, though, is not the unsettling part. Advertising to the bereaved is old; every culture manages it with more or less grace. The unsettling part is what happens when the machine’s belief about you is wrong — or right, at precisely the wrong moment. Because the machine is wrong-but-confident by default, and it is built to act on its confidence instantly, at scale, with no capacity for tact.
The letter
In December 2018 Gillian Brockell, a journalist at The Washington Post, published an open letter to the tech companies whose ads had accompanied her pregnancy. “I know you knew I was pregnant,” it began. Of course they knew: she had searched what expectant parents search, clicked the maternity ads, used the hashtags. Then her son was stillborn. The pregnancy ads kept arriving, and when she tried to dismiss them, the systems drew the obvious statistical conclusion — this user is no longer pregnant, therefore the baby has been born — and moved her briskly along the conveyor to the next products.1
Rob Goldman, then Facebook’s vice-president of ads, responded publicly, and helpfully, with instructions for hiding certain ad topics in the settings. Which is true, and is also the whole problem in one gesture: the machine had assembled its intimate, largely accurate picture of her life for free, from ambient signals, without asking. When the picture curdled, correcting it became her job — a preferences menu, one toggle at a time, in the first week of grief.
The machine inferred the pregnancy for free. Un-inferring it was her job.
The genre is bigger than any one platform. In January 2014 a man in Illinois named Mike Seay received a marketing letter from OfficeMax addressed to “Mike Seay, Daughter Killed in Car Crash or Current Business”. His daughter had died in a crash the year before. OfficeMax blamed a third-party data broker, which is presumably accurate and entirely beside the point: somewhere upstream, a database held a field recording the worst fact of a stranger’s life, and nothing in the pipeline between that database and his letterbox checked what got printed.2
Wrong-but-confident by default
Be precise about what these systems do: “they know” is the wrong verb. Nothing in the stack knows anything. A model scores the probability that a given identifier belongs to someone about to buy prams, or protein powder, or probate services, from proxies: searches, dwell times, purchases, the behaviour of statistically similar people. Then the machinery acts as if the score were a fact, because acting is the only thing it can do with one. The famous story — the American retailer Target inferring a teenager’s pregnancy from her shopping basket before her father knew — comes from a 2012 New York Times Magazine piece and may be polished in the retelling; the capability it describes is by now mundane.3
Pregnancy is the canonical case because it is commercially enormous and sensitive in both directions at once. The inference can arrive before you are ready — before the twelve-week scan, before you have told your own mother — and it can persist after there is nothing left to advertise to. What the machine lacks is not data. It lacks the concept of an ending. A profile has no field for loss, only topics that were recently warm; and retargeting logic reads every non-purchase as not yet rather than never — keep trying, escalate the discount, extend the window.
Grief looks, to the model, like an unconverted lead.
The asymmetry
Why does the machine not simply hold back when it is unsure? Because of what an error costs, and who pays it.
$0.002
At open-web CPMs of a few dollars per thousand impressions, showing an ad to the wrong person costs a fraction of a cent. Failing to show it to the right person costs real revenue, invisibly. Every optimisation loop in the chain therefore learns the same lesson: when in doubt, serve. The whole apparatus is tuned, in classifier terms, for recall over precision — better to fire at a hundred people who are not pregnant than to miss the three who are.
The trouble is that the two costs fall on different parties. The advertiser pays for the impression; the platform is paid for it; the person it lands on pays for everything else. And there is no channel through which that payment registers. These systems have a signal for “this worked” — the click, the conversion — and no signal at all for “this wounded someone”. A grieving person closing the tab is, to the ledger, indistinguishable from a bored one.
The suppression tools, and where they leak
The platforms have not ignored any of this. In January 2022 Meta removed the detailed-targeting options that referenced health causes, sexual orientation, religious practice and political beliefs — advertisers can no longer aim at “chemotherapy” or “LGBT culture” by name.4 Google’s personalised-advertising policy restricts targeting built on sensitive interest categories such as health conditions and personal hardship — the policy’s own examples run to bereavement products —5 and its My Ad Center lets you ask for fewer ads about alcohol, dating, gambling, weight loss, and pregnancy and parenting.
These controls are real, and better than what preceded them. Their limits are structural rather than careless. First, they run on self-declaration: you must know the toggle exists, find it and flip it, at precisely the moment you are least equipped for settings menus — Brockell made exactly this point in reply to Facebook. Second, suppression operates on labels, and inference does not need labels. Delete “health causes” from the targeting menu, and a lookalike audience seeded from a clinic’s customer list will assemble much the same crowd without ever naming it. Prams find their way to people who buy folic acid. The category dies; the proxy carries on.
Third, and least visible: switching off “pregnancy and parenting” does not delete the inference. It mutes one use of it. The profile that concluded you were expecting persists, and feeds whatever downstream systems still consult it — which is part of how an opt-out can read as a birth.
A humane version of this machine is imaginable. It would treat certain inferences the way medicine treats certain test results: worth acting on only above a far higher threshold of confidence, with silence as the default and the cost of a false positive priced into the maths. Nothing in the current auction works that way, because the harm has no line in the ledger — and the ledger is the whole of the machine’s morality.
I write this for an ad-funded site, which is not lost on me; the machine is acting on its beliefs about you somewhere on this site right now. Most days its errors are merely absurd — the boots that follow you for weeks after you bought them. But the same cheap confidence, pointed at the wrong week of a person’s life, stops being absurd at all. You can, in fact, be in-market for a funeral. The machine may notice before your friends do. What it has never learned is when to look away.