Targeted marketing works when it makes a message more useful, not when it proves how much a company knows. Customers are likely to reject targeting that feels intrusive, inaccurate or impossible to control. The practical answer is not to abandon relevance. It is to use fewer, better signals, explain the exchange clearly and give people meaningful choices.
Why targeting can feel uncomfortable
Marketers often see a segment, score or audience. The person receiving the message sees a company making an assumption about them. Discomfort grows when the assumption concerns a sensitive subject, follows someone across many websites, appears immediately after a private conversation or cannot be corrected.
Even accurate targeting can feel wrong when the value is one-sided. A brand may gain a higher click probability while the customer receives an advertisement they did not ask for. Relevance is therefore not just statistical accuracy. It is a judgement about context, expectations and benefit.
Begin with a legitimate customer need
Before choosing data or technology, define the decision you want to improve for the customer. Do you want to prevent irrelevant product recommendations, remind someone about an unfinished task or adapt information to a professional role? If the only answer is “increase conversion,” the customer benefit is still unclear.
Use a simple test: would you be comfortable explaining the targeting rule on the page where the message appears? “We show this guide because you selected finance as your role” is easy to defend. “We inferred a sensitive situation from behaviour across unrelated sites” is much harder.
Use the least data necessary
More data does not automatically produce better marketing. Old, purchased or inferred attributes introduce errors and can create surprising experiences. Start with first-party information that people intentionally provide, such as selected interests, saved products, account settings and direct feedback.
Collect only what is needed for a defined purpose, set retention periods and remove attributes that no longer improve decisions. Separate operational data from marketing permissions. A company may need an email address to deliver a service without automatically having permission to use it for every promotion.
The UK Information Commissioner’s Office advises organisations to plan direct marketing carefully, respect preferences and make it easy for people to change their mind. Its direct marketing checklist is also a useful operational review. Local law differs, so obtain legal advice for the markets and channels you use.
Make consent understandable
Consent is not a paragraph of legal language next to a preselected box. A useful choice states who will communicate, through which channel, about what and how often. It should be as easy to withdraw as it was to give.
A preference centre can go beyond yes or no. Let people choose topics, frequency and channels. Preserve the decision across systems so that an unsubscribe in one place is not reversed elsewhere. The ICO’s guidance on respecting people’s preferences stresses that people can withdraw consent or opt out.
Choose segments that help rather than label
Good segments reflect a current need or task. Weak segments turn demographic or behavioural clues into a fixed identity. Someone who reads one article about redundancy is not necessarily unemployed. A customer who buys a gift should not receive months of messages based on the recipient’s interests.
Use observable context with a short shelf life where possible: page category, selected use case, declared role or stage in a known process. Allow people to change preferences and provide a route back to a general experience. Our comparison of persona versus person explains why a model should never replace attention to the individual.
Control frequency and sensitive moments
A relevant message becomes irritating when repeated too often. Cap frequency across campaigns and channels, not just within one advertising platform. Suppress promotions after a complaint, cancellation, bereavement notice or unresolved service problem. Coordinate service and marketing data so that automation does not ignore the relationship’s current reality.
Treat sensitive categories, location, health, finances and children with extra caution. Some uses require explicit consent or may be inappropriate even when technically possible. The standard should be more demanding than “the platform allowed it.”
Measure customer value, not only response
A targeted campaign can improve click-through while damaging trust. Add unsubscribe, complaint, frequency, preference changes, suppression errors and long-term retention to the measurement plan. Compare targeted experiences with a useful general alternative. Incrementality matters: a conversion that would have happened anyway is not proof that more surveillance was worthwhile.
A seven-question targeting review
- What customer decision becomes easier?
- Which minimum signals are required?
- Would the source and use of those signals be expected?
- Is the message safe if the inference is wrong?
- Can the person see, change or stop the use?
- Are sensitive moments and excessive frequency suppressed?
- Do we measure trust signals as well as conversions?
Use the review before launch and during regular audits. Privacy, relevance and brand value are connected: a short-term response gain is expensive if it teaches customers to distrust future communication. Strong brand value is protected by behaving consistently when data makes more actions possible.
Frequently asked questions
Is personalisation always better than a general message?
No. A clear general message can outperform weak or inaccurate personalisation. Personalise only when the available signal changes what is genuinely useful.
Is first-party data automatically safe to use?
No. Its use still needs a defined purpose, suitable legal basis, security and respect for expectations and channel rules.
What is the safest place to start?
Start with declared preferences and contextual relevance. They are easier to explain, maintain and correct than opaque inferences from third-party behaviour.
