The Target Pregnancy Story, Revisited: The Ethics Line in Personalisation

The story you have probably heard

I want to start with the most famous story in the whole of personalisation, because I think almost everyone tells it to make the wrong point. You have probably heard some version of it already. A large American retailer got good enough at reading buying patterns that it could work out a customer was likely pregnant from what she was quietly putting in her basket, and it started sending her baby related offers. As the story is usually told, a parent found out about the pregnancy from the coupons landing on the doormat before the family had been told a word. It is a brilliant story and it gets wheeled out at every marketing conference as proof of how clever data has become.

I think that is exactly the wrong lesson to take from it. And before we go any further, one honest caveat, because it matters for how we use the story. This is a tale that has been retold so many times, in so many articles and keynote slides and conference stages, that the tidy version we all repeat is almost certainly smoother than whatever actually happened in real life. Details get sharpened a little every time it is passed on, the way any good story does. So I am not going to lean on any private detail as fact, and I am not going to pretend I know exactly what one company did on one day years ago. I am going to treat it as what it really is, a widely reported piece of marketing folklore, and use it as a mirror we can all look into. Because the useful part is not whether every beat is literally true. The useful part is that it makes an uncomfortable feeling very easy to point at, and once you can point at a feeling you can build a rule around it.

Can we and should we are two different meetings

The interesting thing about the pregnancy story is not that the maths worked. Predicting a life event from a basket is genuinely impressive, and by now it is not even rare. Half the tools on the market will happily score your customers for intent, churn, life stage and a dozen other things before lunch. The interesting thing is that working something out and acting on it were two completely separate decisions, and a business can get the first one brilliantly right and the second one badly wrong.

Most data teams only ever hold the first meeting. Can we predict this. Is the model accurate. Does the segment convert. How big is the uplift. Those are good questions and they are the ones everybody is trained to ask, because they are the ones with numbers attached. But there is a second meeting that almost nobody schedules, and it is the one that quietly decides whether you keep your customer's trust or burn it. That meeting asks a completely different question. Not can we, but should we, and would the person on the other end be comfortable if they could see exactly how we knew.

The aha: the accuracy of a prediction and the wisdom of acting on it live in two different rooms, and success in the first room tells you nothing about the second. A model can be right in a way that makes a human recoil. The test was never the maths. The test is how the person feels the moment they realise you made the call.

The filter I actually use

So here is the filter I run on any segmentation or targeting idea, mine or a client's, before it goes anywhere near a live campaign. It is three questions, in this order, and the order genuinely matters because each gate assumes the one before it has been passed.

First, is it accurate. If the prediction is wrong you have a data quality problem and nothing else is worth discussing yet. Fix that first, because acting on a confident wrong guess is its own special kind of damage. Telling a grieving customer about a milestone they never reached, or congratulating someone on a thing that did not happen, is worse than saying nothing at all. So accuracy is the floor, not the finish line.

Second, is it welcome. Would the customer actually want this message, this offer, this nudge, at this moment, in this channel. Useful to you is not the same as wanted by them. You can be completely right about what someone needs and still land as an interruption because the timing is wrong or the moment is private. A reminder that you are running low on coffee is welcome. The same mechanism pointed at something sensitive is not.

Third, and this is the one people skip straight past, would they be comfortable seeing why. Imagine the customer could click a little link sitting right next to your message that said, in plain words, here is exactly how we decided to send you this. If that explanation makes them feel looked after, you are fine. If it makes them feel watched, you have crossed the line, no matter how clever the model was or how well the segment converts in a test.

Run those three in order and most creepy ideas die at the third gate while still looking perfectly healthy at the first two. The pregnancy story sails through accurate, probably passes welcome for a lot of expectant parents who genuinely do want the offers, and then walks straight into a wall at the third question. Because the honest explanation, we noticed the specific and rather personal things you bought and inferred a private life event you had not told us about, is not a sentence anyone wants to read on their own doormat.

Draw it as a decision

Here is the same filter as a diagram, because I find it much easier to argue with when it is up on the wall in front of the team.

Notice that only one path out of the whole diagram ends in act on it. Three of the four exits are a full stop. That ratio is roughly right for how this feels in practice. Most predictions you can make are not predictions you should send, and a healthy personalisation programme throws away far more ideas than it ships. The teams that get into trouble are the ones where every box that lights up green on accuracy gets pushed straight to the customer, because nobody in the room was ever given the job of standing at the third gate.

Inference is the tell

If you want one word for where the line sits, it is inference. Helpful personalisation works with what the customer knowingly handed you, for the obvious reason they handed it over. Creepy personalisation reaches past that, into what they did not tell you, and then acts on the guess in a way that shows its working.

Here is the same idea as a plain word formula, because it is the single most useful thing in this whole piece.

Helpful use = data they gave you, used for the reason they gave it
Creepy use  = that same data, plus inference they never authorised, shown out loud

The two halves of that second line are what tip a message over. The reaching past, and the showing out loud. You can often do one without the other and stay on the right side. Quietly using an inference to rank a homepage is very different from writing the inference into the subject line where the customer reads it back. The moment your copy effectively says we know this private thing about you, you have handed the customer the exact sentence that turns service into surveillance in their head.

Helpful and creepy, side by side

The line is easier to feel than to define, so here are a few pairs. Same underlying data, same customer, and the only thing that changes is how visibly you use what you know.

Signal you hold	Helpful use	Creepy use
Browsing history on your own shop	Show the items still in the basket next time they visit	Email them about a product they looked at once, minutes later, unprompted
Purchase of a life event item	Offer a genuinely useful reorder reminder when supplies run low	Infer a private circumstance and market the whole life stage back at them
Location data	Tell them which nearby branch has the thing in stock	Message them the moment they walk past a shop they never gave you
Time between orders	Send a helpful running low nudge on a consumable	Comment, in the copy, on how often they buy something private
A quiz or profile they filled in	Use the preferences they told you to tailor what you show	Combine it with things they did not tell you to guess something they hid

Read down the two columns and the pattern jumps out. The helpful column only ever uses information the customer knowingly handed over, for the obvious purpose they handed it over for. The creepy column uses inference to reach past what they told you into what they did not, and then shows its working out loud. That is the tell. Personalisation stops feeling like service and starts feeling like surveillance at the exact moment the customer realises you know something they never chose to share.

The line moves with context

Here is the wrinkle that catches even careful teams. The line is not fixed to a data type. It moves with context, with channel, with timing, and with how sensitive the topic is. The same signal can be a lovely bit of service in one setting and a violation in another, which is precisely why a fixed rulebook of allowed and banned data never quite works, and why the comfort question does.

Same signal	Feels like service when	Feels like surveillance when
You know their birthday	A warm note and a small treat on the day	You reference their exact age in a nudge about a sensitive product
You know they abandoned a basket	A gentle reminder that their items are saved	Three escalating emails in an hour implying you are watching them decide
You know their health or wellbeing interests	They opted into that topic and you keep to it	You infer a condition from unrelated purchases and market to it
You know a device or channel they use	You format the message to work well there	You reveal, in the copy, that you know where and when they read
You know they are a lapsed customer	An honest we have missed you with a reason to return	A message that recites their old behaviour back to prove how closely you tracked it

The right hand column is not a different dataset from the left. It is the same knowledge, tipped over by context. Sensitivity of topic, aggressiveness of timing, and how loudly you narrate what you know are the three dials that move a message across the line. Turn all three down and almost anything reads as care. Turn all three up and even innocent data starts to feel like a stakeout.

A worked example

Let me make this concrete with a small, invented example, so nothing here rests on anyone's real data. Say I am working with a homeware shop. Their analytics quietly flags a cluster of customers whose recent baskets, cots, blackout blinds, a particular kind of nightlight, a nursing chair, point fairly strongly at a new baby on the way. The model is confident. The commercial temptation is obvious. There is a whole nursery and baby category sitting right there, ready to be recommended.

So we run the three gates.

Is it accurate. Probably, for most of the cluster. Passes.

Is it welcome. For a good chunk of new parents, honestly, yes. People setting up a nursery often do want relevant, useful things. So far the numbers all say go.

Would they be comfortable seeing why. Here it falls apart. If the message reads congratulations on the baby, here is everything for the nursery, and the customer knows they never told this shop anything about a baby, the explanation behind that message is we inferred your pregnancy from your basket. That is the doormat moment. And it is worse for the slice of the cluster where the model is wrong, or where the pregnancy is early, private, complicated, or ended. For those people the message is not just creepy, it can be genuinely painful.

So what do we actually ship. We do not throw the insight away, we defang it. We make the nursery and baby range easy to find and beautifully merchandised for anyone who goes looking, so the customer who wants it walks straight to it. We let the customer tell us, with a gentle and clearly optional setting up a nursery preference they can switch on, at which point everything becomes fair game because they said so. And we keep the automated outreach to the boring, safe, explainable stuff, your blinds are on the way, here is how to fit them. Same underlying data, a fraction of the risk, and the only thing we gave up was the one move that would have felt like surveillance. The insight still earns its keep. It just stops narrating itself at the customer.

The aha: you almost never have to choose between the insight and the trust. You have to choose whether the insight speaks. Let the customer walk toward the prediction rather than pushing the prediction at the customer, and the same knowledge becomes an asset instead of a liability.

The trust battery

Here is the part I most want you to take away, and it is the reason I bother with any of this beyond staying on the right side of the law.

When you cross the creepiness line, the cost is not a complaint. A complaint you can answer, and honestly a complaint is a gift because at least you found out. The cost is trust, and trust is the entire asset in a customer relationship. It is the only thing that makes someone open your email instead of deleting it, click your recommendation instead of ignoring it, hand you their card instead of your competitor's. Every clever thing personalisation can do for you runs on that one battery, and the battery does not recharge on demand. Once a customer decides you are the shop that watches them, that feeling quietly colours every message you ever send afterwards, including the genuinely helpful ones. You do not get to explain your way back. You just become slightly creepy in their mind, more or less permanently, and you will never see it in a dashboard, because they do not complain, they simply trust you a little less and buy a little less and drift a little further and you never once find out why.

That asymmetry is brutal and it is the whole argument. A well judged bit of personalisation earns you a small amount of trust. A creepy one does not cost you a small amount, it can cost you the relationship, quietly, with no receipt and no chance to appeal. You are not playing a game with symmetrical stakes. You are protecting the one thing that took years to build and cannot be rebought with a discount code.

And here is the bit that should genuinely cheer you up, because this is not really a lecture about restraint. The businesses that win at personalisation over the long run are not the ones with the most invasive models. They are the ones customers trust enough to hand over more information voluntarily, which makes the personalisation better, which earns more trust, which brings more information still. It is a flywheel, and consent respecting marketing is what keeps it turning.

Consent respecting marketing is not the polite tax you pay on growth. It is the growth. The creepy shortcut does not just risk the fine, it snaps the very wheel you need turning, and a snapped flywheel does not get you there faster, it stops.

The common mistakes I see

Most teams do not cross the line on purpose. They walk over it while looking at a chart. Here are the patterns I run into most.

Optimising to the click and nothing else. If your only measure is short term conversion, creepy will often win the test, because a message that says we know this private thing about you does get attention. It just spends trust to buy that attention, and your metric cannot see the spend. The uplift is real and the cost is invisible, which is the worst possible combination.

Confusing consent to collect with consent to infer. A customer ticking a box to receive offers is not the same as a customer agreeing to have their private circumstances deduced from unrelated behaviour. The first is permission. The second is a leap they never signed off on, and dressing it up in a lawful basis does not make it feel any less like a leap.

Treating a legal yes as a comfort yes. You can be entirely within the letter of the rules and still make someone's skin crawl. Compliance is the floor you must not go below, not the ceiling you are aiming for. The comfort test lives well above the legal line, and it is the one customers actually feel.

Narrating the inference in the copy. This is the self inflicted one. Even a fair, useful bit of targeting turns creepy the instant your words tell the customer how closely you have been watching. Use the insight to decide what to show. Do not read the insight back to them like a case file.

No owner at the third gate. In most teams, everyone owns accuracy and nobody owns comfort. Whoever wants to ship has revenue on their side, and whoever feels uneasy has only a feeling, which loses every argument it is made to have. If comfort has no named champion with permission to say no, it will lose by default, quietly, every time.

Renting inference instead of earning it. Buying scored audiences and third party inferences feels efficient, but you are building your customer relationships on data they never gave you and cannot see, which is the definition of the creepy side. Owned, volunteered data is slower to gather and worth vastly more, because nobody can ever feel violated by you using what they chose to tell you.

How to build consent respecting personalisation into how you work

None of this needs a philosophy department. It needs a few habits wired into the way campaigns actually get built, so the right thing happens by default rather than by heroics.

Ask rather than infer, wherever you reasonably can. A preference someone typed into a form is worth more than one you guessed, not only because it is more accurate, but because they gave it to you, so using it can never feel like a violation. The whole point of a proper first party relationship, a real account, a profile the customer controls, is that they are choosing to tell you things. Give people easy, obvious ways to declare what they want, then honour it. That owned, consent first data foundation is exactly the sort of thing I keep steering clients towards instead of renting inferences from a black box, and it is a big part of what my data and personalisation work is built around.

Put the third question into the actual process. Before any segment ships, someone in the room says out loud, if this customer could see precisely why they are getting this, how would they feel. Make it a required box on the brief, not an optional conscience. Give one named person permission to say this one is over the line without needing to win an argument about revenue. That single ritual catches almost everything, and it costs nothing but a minute of honesty.

Make it explainable on purpose. If you would be uncomfortable showing the customer the reason behind a message, that discomfort is the signal, not an inconvenience to route around. A good working test is whether you could put the why in the footer of the email itself. Personalisation you would happily explain to someone's face builds trust. Personalisation you would need to hide is quietly borrowing against it.

Let people walk toward the prediction, not get pushed by it. When you hold an insight that is too sensitive to act on directly, make the relevant thing easy to find for anyone who wants it, and let the customer opt in to more. You keep the commercial value and drop the ambush.

Measure trust, not just clicks. Watch the slow signals, unsubscribes, quiet disengagement, the customers who open less over time, alongside the fast ones. If a clever campaign lifts conversion this week but bleeds engagement next month, your comfort line just showed up in the numbers, and you want to be the team that reads it.

A quick self audit

If you want something you can run this week, here it is. Take your three most personalised, most automated messages, the ones you are quietly proud of, and put each one through this.

Question to ask of the message	If the honest answer is no
Is the prediction behind it actually accurate	Stop and fix the data before anything else
Did the customer knowingly give us the data it uses	You are inferring, so tread very carefully
Would we happily print the reason in the footer	Rewrite it so you would, or do not send it
Are we narrating what we know back at them	Take the insight out of the copy, keep it in the logic
Does anyone own the comfort call on this	Give one person the job before it ships again

If a message you love fails two or more of those, it is almost certainly earning short term clicks by spending long term trust. That is a bad trade even when the dashboard smiles at you, because the dashboard only ever sees one side of it.

The one question to leave with

Forget every framework for a second, including mine, and keep just this. The next time your data throws up something clever about a customer, and it will, do not ask whether the model is right. Assume it is. Ask instead whether that customer would smile or flinch if they could see exactly how you knew. If they would smile, you have a lovely piece of service and you should ship it today. If they would flinch, you have found the line, and the clever thing to do is walk back from it, because the prediction was never the asset. The trust was.

That is the real lesson of the story everyone tells for the wrong reason. It was never a story about how clever data has become. It was a story about the gap between can we and should we, and about the quiet, unrecoverable cost of getting that gap wrong. If you want a hand building personalisation that stays firmly on the smile side of the line, and a targeting approach your customers would happily see the reasoning behind, that is exactly the conversation I like having. Come and talk it through with me and we will pressure test your best ideas against the comfort line together.