Hyper-Personalization Without Being Creepy: Where the Line Sits

For most of the history of marketing, the limit on personalization was cost. Writing a different email for every customer segment took people, and people are expensive. Generative AI has removed much of that limit. A small team can now produce tailored copy, offers and images for dozens of micro-segments, and in some cases for individuals, in the time it used to take to write one campaign.

That shifts the question. It is no longer whether you can personalize, but how far you should. The same capability that makes a customer feel understood can make another feel watched. This guide sets out where the line sits, using research rather than gut feel, and turns it into rules a small marketing team can apply.

Customers want personalization, on their terms

The demand side is not in doubt. In McKinsey’s Next in Personalization research (2021), 71% of consumers said they expect companies to deliver personalized interactions, and 76% said they get frustrated when that does not happen. McKinsey (2021) also found personalization most often drives a 10 to 15% revenue lift, with a company-specific range of 5 to 25% depending on sector and execution (McKinsey).

Look at what those consumers described as good personalization, though, and it is modest. They associated it with being recognized and made to feel special: a check-in after a purchase, a how-to video, a request for a review. None of that requires knowing anything the customer did not hand over willingly.

Trust is the other half of the picture, and it is thin. In a Pew Research Center survey of 5,101 US adults in May 2023, 67% said they understand little to nothing about what companies do with their personal data, and 73% felt they have little or no control over it. Among those who had heard of AI, 81% expected companies’ use of it to lead to personal information being used in ways they would not be comfortable with (Pew Research Center).

So the starting position is a customer who wants relevance and does not trust you with their data. Every personalization decision sits inside that tension.

Where the creepy line actually sits

The most useful evidence on this comes from a set of studies by Tami Kim, Kate Barasz and Leslie John, published in the Journal of Consumer Research in 2019 and summarized for practitioners in Harvard Business Review (Kim, Barasz and John). They asked people how they felt about different ways their data could be used to target ads, and found two clear dividing lines.

1. Where the data came from. information gathered on the same website where the ad appears was seen as acceptable. Information gathered elsewhere, for example by tracking someone across other sites, was seen as less acceptable.
2. How you came to know it. information the customer stated was acceptable. Information the company inferred about them was less acceptable, even when the inference was accurate.

The finding that matters commercially is what happened next. When the researchers revealed an unacceptable data flow alongside an ad, the ad became less effective, because it raised people’s concern for their privacy above their interest in the relevance. When the platform was trusted and the data flow was acceptable, revealing it made the ad more effective. Two field studies on a loyalty program website, a trusted platform, showed the same benefit.

Put simply, the problem is rarely personalization itself. It is personalization that exposes a data practice the customer did not agree to. Transparency works for you when your practice is clean, and against you when it is not.

The best-known cautionary tale fits this pattern. In a 2012 New York Times Magazine investigation, Charles Duhigg reported how Target’s marketers asked a statistician whether they could work out if a customer was pregnant, even if she did not want them to know (The New York Times). Inferred, highly sensitive, and never stated: it crosses both lines at once.

Why AI raises the stakes

AI does not create the creepy line, but it moves you toward it faster. McKinsey (2025), in its work on personalized marketing, describes two shifts: AI-driven targeting of promotions to micro-communities, and generative AI that produces tailored copy and imagery at a volume that used to be cost-prohibitive. It also warns that companies need ways to validate and govern AI-created content against bias, toxicity and hallucinations (McKinsey).

Three things change in practice when AI is doing the personalizing:

• Inference becomes the default. models are built to find patterns, so they will happily infer life events, income bands or health conditions from behavior. The system does not know which of those inferences are off-limits unless you tell it.
• Specificity rises. generated copy can name the exact product someone browsed at 11pm, or reference their neighborhood. Specific is not the same as relevant, and past a point it reads as surveillance.
• Review gets harder. when one campaign becomes four hundred variants, nobody reads them all. A mistake that a human would catch in one email ships in hundreds.

There is also an older warning worth keeping in mind. Back in 2019, Gartner reported that 27% of marketers saw data as the key obstacle to personalization, and its analyst Charles Golvin noted that the quest had in some cases backfired as consumers rejected brands’ overtures (Gartner). Better models have not removed that risk. They have made it cheaper to repeat at scale.

The rules that keep personalization on the right side

These rules follow directly from the research and the regulation. They are deliberately simple, so a team can apply them without a lawyer in the room for every campaign.

• Build on stated, first-party data. preferences the customer set, products they bought from you, content they chose to read on your site. This is the data the research shows people accept.
• Ask instead of guessing. a short preference center or a one-question survey often gets you more useful data than a model, and it turns an inference into a statement. That moves the same personalization from the creepy column to the acceptable one. Do not extend this to sensitive topics: asking about health or beliefs needs a clear lawful basis, and disclosure should never be a condition of service.
• Keep personalization on your own channels. your email, your app, your site and your loyalty program. Following someone across the web with what you know is the flow customers rate least acceptable.
• Set a hard list of things you never infer. health, religion, political views, sexual orientation, ethnicity, financial distress and relationship problems. Write it down and put it in the prompt and the audience rules, not just a policy document.
• Personalize the benefit, not the knowledge. say “new arrivals in your size” rather than “we noticed you looked at this three times.” The first serves the customer. The second tells them how closely they are being watched.
• Show your reasoning. a short “because you bought X” line works in your favor when the data flow is clean, which is exactly what the transparency studies found on trusted platforms.

What the rules and platforms already require

The “never infer” list is not only an ethical preference. Regulators and ad platforms already draw much of it for you.

Under UK GDPR, special category data includes racial or ethnic origin, political opinions, religious or philosophical beliefs, trade union membership, genetic and biometric data, health, sex life and sexual orientation. The UK Information Commissioner’s Office is explicit that an inference counts: if you infer one of these traits to treat someone differently, for example to target ads, you are processing special category data regardless of how confident you are that the inference is correct (ICO).

Google Ads enforces a similar list in its personalized advertising policy. Advertisers promoting in sensitive interest categories, which include health, religious beliefs, sexual orientation, negative financial status and relationship hardships, cannot use their own curated audiences such as Customer Match. Users under 18 are not eligible for personalized ads of any kind (Google Ads policy).

If you operate across markets, check the local data law as well. The principle travels well, though: the more sensitive the trait, the more an inference about it is treated as if the customer had disclosed it.

Examples: helpful versus creepy

The same data can produce either outcome. The difference is usually in the execution.

• Replenishment. helpful: “Your coffee order from March usually runs out around now.” Creepy: a reminder that references a medical product the customer bought, sent to a shared family inbox.
• Browsing follow-up. helpful: the item stays in the cart on your site, with a note when the price drops. Creepy: the same item follows them onto news sites for three weeks.
• Life events. helpful: a customer tells you they are moving and gets a checklist and a delivery-address prompt. Creepy: a model infers a move, a pregnancy or a divorce from behavior and adjusts the offers.
• Location. helpful: store hours for the branch they selected as their favorite. Creepy: a push notification that greets them by name as they walk past a shop they never told you they visit.
• Generated copy. helpful: product recommendations that match stated sizes and past purchases. Creepy: AI copy that mentions details the customer never gave you, because the model filled the gap with a guess.

A practical checklist before you ship

Run every new AI personalization use case through these questions:

1. Source. did the customer give us this data, on our channel?
2. Stated or inferred. if inferred, is it on our never-infer list?
3. Explain-it test. would we be comfortable showing the customer exactly why they see this?
4. Benefit. does the message help the customer, or only prove how much we know?
5. Review. has a human sampled the generated variants, including the edge cases?
6. Measure. are we tracking unsubscribes, complaints and opt-outs, not just clicks?

If you are setting up wider guardrails for AI use across the team, the same approach applies beyond marketing (see our guide to AI governance for small teams).

How to measure whether you crossed the line

Click-through rate will not tell you. A slightly creepy message can still get clicks out of curiosity or alarm. Watch the signals that show trust wearing down instead.

• Unsubscribe and opt-out rate by variant. a spike on one personalized version is an early warning.
• Spam complaints. they hurt deliverability for every future email, not just this campaign.
• Preference center activity. people turning personalization off is direct feedback.
• Customer service mentions. a few “how did you know that?” tickets are worth more attention than a dashboard.

Test personalized against generic versions with a holdout group, and judge the result on revenue per recipient over several weeks rather than the first open. Our guide to AI for marketing analytics covers how to set that measurement up (AI for marketing analytics).

AI has made hyper-personalization cheap. It has not made trust cheaper. The brands that get the gains McKinsey describes will be the ones that personalize with what customers chose to share, keep it on their own channels, and never let a model guess at what a person would rather keep private.

If you want help designing a personalization program that grows revenue without spending customer trust, that is the kind of work we do at Gaveau Strategy.