You search for a backpack, read two reviews, and later see ads for hiking gear on another website. The coincidence may seem random, but it may also be the result of behavioral advertising: ads selected using signals about previous activities, interests, or behavior.
The basic idea is straightforward. Rather than considering only the content a person is viewing at that moment, a system may use previously observed information or inferences drawn from it. That does not necessarily mean it knows the person’s name, nor does every personalized ad work in the same way.
Understanding the distinction makes for more useful questions: what was observed, what was inferred, who was involved, and what real options exist to accept, reject, or change the use of that information?
What is behavioral advertising?
Behavioral advertising is the selection or tailoring of ads using data about a person’s, browser’s, or device’s prior activity. Depending on the system, those signals may include pages visited, searches, clicks, interactions, timing, approximate location, or use of an app.
Its defining feature is not whether the ad is accurate, nor simply that it appears online. What matters is that the selection relies on known behavior or inferred interests, not only on the page being viewed at that moment. This distinction is consistent with explanations from bodies such as the UK ICO on online advertising and with the terminology collected by TERMCAT.
Behavioral advertising should not be confused with behavioral economics, which studies how people make decisions, or with every form of personalization. A retailer that displays the language selected by a user is personalizing an experience, but it is not necessarily choosing advertising from a history of behavior.
Key idea: an ad is behavioral because of the information used to select it, not because it is persuasive, timely, or unsettlingly precise.
From a signal to an ad
There is no single architecture, but the process can be understood in four stages.
First come observable signals: a visit, search, click, or interaction. Those signals may be recorded through cookies, pixels, link parameters, device identifiers, or other techniques. Cookies matter, but they are not the only possible mechanism; restricting them does not necessarily eliminate every form of tracking.
Next, one or more actors may organize the signals. Here it is important to separate observed data from inference. Visiting several pages about mountain trails is observed behavior. Classifying a visitor as an “outdoor enthusiast” is a probabilistic inference. The European Data Protection Board’s guidance on targeting expressly distinguishes between data that are provided, observed, and inferred.
Third, the system assigns the browser, device, or account to an audience available to a campaign. Finally, when an advertising opportunity appears, it selects and delivers an ad according to different criteria, including that profile.
A profile need not be a file containing a full name. It may consist of an identifier and several estimated categories. But the absence of a civil name does not automatically make data anonymous: a system may still distinguish an activity, group it, and make decisions about it.
An example: from backpacks to hiking
Imagine that Clara compares backpacks in an online store. She then looks up mountain trails and watches a video about waterproof gear. Later, she sees an ad for hiking boots while reading the news.
Several layers are at work in this sequence. The visit to the store and the searches are observed signals. The assumed interest in hiking may be an inference. The boot ad could be based on a broad category such as “outdoor activities.” If the ad displays the exact backpack Clara left without buying, it is probably a more specific form of retargeting or remarketing.
The example alone cannot tell us which company assembled the information, what technology it used, or whether a coincidence was actually caused by tracking. Answering those questions would require examining the specific notices, controls, and participants. The visible experience is only the last step in a chain that may involve the site visited, a platform, technology providers, and advertisers.
Useful distinction: recording that someone visited a page describes behavior; assigning that person an interest turns the signal into an inference.
Behavioral, contextual, targeting, and retargeting
These terms often appear together, but they are not interchangeable:
- Contextual advertising: selects an ad based on current content or context. An ad for cameras in an article about photography may be contextual even when the reader’s history is not used.
- Behavioral advertising: uses prior or inferred signals about a person, account, browser, or device. Current content may matter, but it is not the only basis.
- Targeting: is the broader decision about which audience will receive an ad. It can draw on context, general location, declared characteristics, behavior, or other criteria.
- Retargeting: seeks to reach someone again after they have visited or interacted with a site, app, product, or business. It can be part of behavioral advertising, but it does not exhaust it.
The practical difference lies in where the information comes from. Two people reading the same page might receive different ads because of their past activity; that suggests a behavioral component. If both see ads related to the article’s subject without prior history being used, the criterion is contextual.
Potential usefulness and information costs
Personalization can reduce irrelevant ads, help people discover a useful offer, and fund services the public uses without direct payment. It can also facilitate voluntary exchanges by connecting interests and offers. These are reasonable possibilities, not guarantees: an ad’s apparent relevance does not by itself show that it is effective or beneficial.
The cost arises when selection depends on observation a person does not understand or cannot control. The asymmetry is not merely that an organization holds data; it may combine them, infer categories, and act on them while the user sees little more than an ad. For more context, it helps to distinguish digital privacy from digital security: the two are related, but a practice can be technically secure and still be opaque.
From the standpoint of individual liberty, the question is not whether all personalization should be accepted or prohibited. What matters is whether participation is understandable and voluntary: what information is collected, for what purpose, for how long, with whom it is shared, and how a person can change their mind. Control over digital property strengthens autonomy only when choices are accessible and have real effects.
Consent, moreover, is not reduced to clicking a button. Its legal validity depends on the jurisdiction and the circumstances, so a general explanation is not a substitute for legal advice. As a practical standard, an interface that pressures, confuses, or hides alternatives weakens informed consumer choice, even when it formally presents an option.
Evaluation standard: the relevance of an ad does not by itself make up for a lack of transparency; people need information and controls proportionate to the scope of tracking.
How to recognize and assess it
Seeing a related ad after a search is an indication, not conclusive proof. To assess the practice, it is more useful to review the service’s explanation, its ad preferences, and browser or device controls. The essential questions are concrete: does the ad respond to current content or past activity? Were the data declared, observed, or inferred? Can that use be declined without navigating a maze?
Behavioral advertising turns activity traces into criteria for deciding what to show. It may offer more relevant ads and help sustain services, but it can also shift power toward those who observe and classify without making themselves visible. The sensible response is not to guess how much an ad knows, but to check how much the system can know and how much control a person retains over that information.
Share