Knowing something about a customer and knowing what to charge that customer are not the same achievement. A business can hold a great deal of information and still make a poor prediction. It can make an accurate prediction and use it for a purpose the customer would reject.
To understand personalization, follow the decisions between those two points. What was observed? What was inferred? What action changed? What result was measured? These questions are more revealing than a general claim that an algorithm knows you.
Observations become guesses
A seller may have information a person provided directly, such as an account’s membership status. It may observe behavior within a session. It may obtain information from another source. These routes have different privacy implications and different levels of reliability. A direct answer can be wrong or stale; an inferred characteristic can be a good predictor without being true of an individual.
Suppose, as an invented example, a shop’s records show that visitors who return three times often buy. The shop could use repeated visits to choose a reminder. It could use them to offer a coupon. Or it could withhold a coupon because those visitors seem likely to buy anyway. The same observation supports very different commercial actions.
The crucial missing word is the objective. A system trained to maximize completed sales can make different choices from one trained to maximize immediate profit. A system rewarded for repeat custom can differ from one judged only on this week’s revenue. Saying that a system is accurate tells us little until we know what it was asked to predict and what action followed.
There is also a measurement problem. If a buyer accepts a price, we learn that the offer was acceptable under those circumstances. We do not learn the maximum they would have paid. If they leave, the reason might be price, delivery, distraction, distrust or something unrelated. Purchase records do not arrive with a complete explanation of human motives attached.
Experiments can improve the evidence. Randomly offering different prices to comparable prospects helps estimate how demand changes with price. That is a stronger basis for a demand model than treating every observed difference between customers as caused by price. It still measures the population, product and period studied. The design does not become universal just because a computer processes the results.
A study in which the buyers were employers
Dubé and Misra’s 2023 ZipRecruiter study combines experiments with a demand model. Its buyers were employers. Inputs came from voluntary business and job registration answers, rather than covert household browsing. Final article
The study matters here as an example of how researchers move from observations to a pricing rule. It also shows why we need the complete design rather than a headline about machine learning. A model-based prediction and a separately implemented trial are different parts of the evidence. Their prices and comparisons must not be exchanged midway through an argument.
We will return to the welfare result in the next chapter. For now, the lesson concerns interpretation: identifying the buyer, input and decision can change the meaning of a study before we reach its results. A small employer considering a recruitment subscription is not interchangeable with a household buying food. Their alternatives, timing and consequences of refusal can differ greatly.
The missing customer
A sales database can tell an unusually detailed story about people who buy. It may tell much less about those who leave. That difference is analytically important even before we know its size in any particular market.
Imagine an evaluation that reports a higher average payment among completed orders. This could follow from higher prices charged to the same buyers. It could also follow from lower-spending buyers dropping out, higher-quality products selling, changed product availability, or a mixture of those changes. The average alone cannot choose between these explanations.
The denominator is part of the claim. Are we measuring revenue per visitor, per offer, per customer or per transaction? A system that improves one measure can worsen another. A fair evaluation should include who was offered access, who accepted, who left and what each group received. Where the data cannot show that, the conclusion should narrow.
This is one reason I resist treating efficiency as a complete answer. We need to ask which task became more efficient. Finding a suitable product quickly is a benefit to a buyer. Finding the largest price that buyer will tolerate can primarily benefit the seller. Both might be described as better matching. Only the details tell us what happened.
Consent cannot be inferred from a purchase
A purchase proves that a transaction took place. It does not prove that the customer understood every use of information behind it. Nor does it prove that they would have rejected the transaction had they understood. Those are separate empirical questions.
We should avoid writing the answer for the consumer in either direction. Some people may welcome a relevant discount. Some may prefer a less personalized service. Others may feel they have too little time or too few alternatives to make the choice meaningful. The right research asks what people knew, what options they actually had and what changed when the information or options changed.
The policy implication I draw is that disclosure should arrive when it can affect a decision. A notice hidden in a large document may exist formally while contributing little to the buyer’s understanding. That is a design judgment, not a claim that every such notice violates a law. It can be tested: can people recognize the pricing practice, compare another offer and act on their preference without unreasonable effort?
And there is a further point. Even a clear notice cannot create a missing competitor. We will come back to that when considering fair prices. Information helps, but power also depends on alternatives.
A short record
I gave the shop an address,
so a parcel could find my door.
The form had space for what it needed.
I could not see what it would infer.
The parcel came on Tuesday.
The questions stayed with me:
which parts of being known had helped,
and who had chosen what they meant?
The poem gives an imagined buyer a question, not an answer about any company. The research task is to follow that question through the system: purpose, input, prediction, action and consequence. A precise account of those steps leaves us better able to recognize both useful service and an unfair use of advantage.