Cookies are not “dead,” and AI does not remove the rules that govern them. Marketers in 2026 work with first-party records, cookies, pixels, device identifiers, clean rooms, contextual signals and modelled audiences. Each has a purpose, evidence requirement and legal boundary. Calling a system cookieless says little about whether it is fair or lawful.
This article, originally published in December 2025, is updated through 31 July 2026. It addresses UK data protection, electronic communications, consumer protection and advertising rules. Exact requirements depend on channel, audience, product, data, jurisdiction and whether the communication is business-to-consumer or business-to-business.
Map the decision before the data
Personalisation is a sequence of decisions: who is eligible, which content is considered, how it is ranked, what price or offer appears and which channel delivers it. Record each effect separately.
| Decision | Inputs that may be relevant | Control |
|---|---|---|
| Suppression | Valid opt-out and campaign purpose | Central preference service |
| Eligibility | Product, location and stated need | Deterministic policy rule |
| Ranking | Recent interaction and contextual relevance | Diversity and conflict checks |
| Creative variant | Approved claims and brand assets | Evidence-linked generation |
| Timing | Service context and channel permission | Frequency and quiet-period limits |
| Offer | Published terms and approved segment | Fairness and value review |
| Measurement | Exposure and outcome at useful grain | Consent, minimisation and retention |
Do not begin with every available field. Define what the intervention should improve for the consumer and the business, then identify the least information needed. Separate service communications from direct marketing and editorial recommendations from paid placement.
Our smart-retail AI guide covers marketplace and retailer-specific conflicts in more detail.
Replace the cookieless slogan with a technology register
The ICO’s final April 2026 storage and access technologies guidance covers cookies, tracking pixels, device fingerprinting and similar tools under PECR and, where relevant, UK GDPR. Technology-neutral rules mean avoiding a cookie does not avoid responsibility.
Maintain a register for every tag, SDK and identifier:
- owner and supplier;
- purpose;
- data stored or accessed;
- device and browser scope;
- recipient and destination;
- consent or exemption relied on;
- retention;
- trigger and pages;
- security and change process;
- deletion and withdrawal behaviour.
Scan the live site and app, because the consent platform configuration may not match what vendors actually load. Test before consent, after acceptance, after rejection and after withdrawal. Ensure an essential service still works where optional advertising technology is refused.
Server-side tracking and clean rooms still process information. Hashing an email address does not necessarily make it anonymous. Treat claims of anonymity as a technical and legal conclusion that needs evidence.
Apply direct-marketing rules by channel
The ICO’s direct marketing guidance was updated in April 2026. It sets out a design process: identify direct marketing, plan with a lawful basis, collect fairly and respect people’s preferences. People have an absolute right to object to direct marketing.
For email, text and similar stored messages, the ICO’s detailed electronic-mail marketing guidance explains consent, subscriber types and the limited soft opt-ins. A receipt address is not automatically marketing consent. A products-and-services soft opt-in has conditions; charities received a distinct soft opt-in through the Data (Use and Access) Act changes.
Build permissions as auditable events:
- identity or contact point;
- channel;
- exact statement shown;
- affirmative action;
- time and source;
- business or brand scope;
- products or purposes covered;
- withdrawal and suppression state.
The campaign service should check permission at send time, not rely on a list exported weeks earlier. Keep suppression data sufficient to honour the objection even after deleting wider profile data.
Do not hide marketing in an account, security or delivery message. If a service message also promotes a product, assess the whole communication.
Use profiling without inventing a person
Models can infer interests, price sensitivity, health concerns or life events from weak signals. The inference is personal data when linked to an identifiable person and may be sensitive even when wrong.
For each feature, ask:
- Did the person provide it, did the organisation observe it or did a model infer it?
- Would the person reasonably expect this use?
- Could it reveal special-category or highly sensitive information?
- What harm follows from an incorrect inference?
- Can the campaign work with a broader, less intrusive segment?
- Can the person understand and change the relevant preference?
Do not target or exclude based on inferred distress, illness, ethnicity, religion or vulnerability without a clearly lawful, fair and defensible basis. Avoid proxy features that reproduce those effects.
Data minimisation applies to model training and evaluation as well as activation. Remove obsolete features, set time windows and prevent customer-service notes from becoming advertising features by default.
Our UK AI and data-privacy guide explains the wider controller, DPIA and rights framework.
Generate creative inside an evidence envelope
A generative system should assemble from approved claims, prices, terms, visual assets and disclosure rules. It should not browse freely for a stronger assertion.
Create a claim registry containing:
- exact approved wording and permitted variations;
- evidence owner and source;
- products, markets and audiences;
- effective and expiry dates;
- mandatory qualifications;
- prohibited comparisons or implications;
- required legal and brand review.
The ASA’s June 2026 guidance on AI and deepfakes for advertisers makes the key point that the CAP Code remains media-neutral. AI-generated imagery, testimonials and endorsements are assessed under existing rules.
Review the overall impression, not only literal text. A synthetic doctor, local storefront, customer quote or celebrity likeness can imply expertise, location, experience or endorsement that does not exist. Label AI content where necessary for clarity, but do not expect a label to cure a misleading claim.
Keep human approval for new objective claims, regulated products, sensitive audiences, comparative ads and high-reach campaigns. Small spelling or layout variants may use bounded automation once tested.
Protect consumer choice
The CMA’s guidance on unfair commercial practices explains the Digital Markets, Competition and Consumers Act 2024 provisions applying to commercial practices from 6 April 2025. They cover misleading actions and omissions, aggressive practices and banned practices including fake reviews and certain pressure selling.
Personalisation must not turn those practices into individual experiments. Prohibit:
- invented scarcity or countdowns;
- material information hidden for a predicted low-attention user;
- a harder cancellation path for a retention-risk segment;
- fake reviews, testimonials or endorsements;
- paid ranking presented as neutral advice;
- a higher mandatory charge revealed only at checkout;
- repeated pressure after a person declines;
- claims selected because the model predicts they will not be challenged.
The CMA’s online choice architecture collection shows how ranking, defaults, information and pressure can help or harm choice. Test the complete journey, including rejection and exit, rather than optimising the accept button.
If personalisation affects a displayed price or fee, use the CMA’s current price-transparency guidance and retain the basis for any segment-specific offer. The total price and mandatory charges must not become a model-generated guess.
Experiment with guardrails
An A/B test is an intervention on people, not just a dashboard. Pre-register the purpose, primary metric, harm metrics, population, duration and stop rules. Avoid running multiple overlapping experiments that make the effect uninterpretable.
Use outcome measures beyond conversion:
- cancellations and returns;
- complaint causes;
- unwanted-contact rate;
- unsubscribe and objection completion;
- price or term comprehension;
- accessibility failures;
- adverse differences between relevant groups;
- long-term retention without lock-in;
- content correction and takedown time.
Hold out a stable comparison group where proportionate and lawful. Do not optimise continuously until every person receives the most aggressive variant. Review whether short-term engagement conflicts with value, trust or consumer understanding.
Frequency caps should work across email, app, web and paid media. A customer who says no in one channel should not be pursued through five others because identities were not reconciled.
Secure the marketing supply chain
Campaigns cross customer-data platforms, model APIs, agencies, ad platforms, analytics and content-management systems. Inventory data movement and administrator authority.
Controls should include:
- role-based access and strong authentication;
- separate production and test audiences;
- synthetic data for prompt development;
- allow-listed data exports and destinations;
- approval for bulk activation or spend changes;
- prompt-injection and malicious-asset testing;
- provenance for generated text and images;
- supplier change and incident notification;
- deletion and suppression propagation;
- tested pause, rollback and provider exit.
Never paste a customer list into an unapproved consumer model. Strip hidden instructions and metadata from uploaded briefs. A creative agent should not be able to alter budget, audience and claims in one unreviewed action.
Use a 90-day bounded programme
Days 1–30: inventory and choose. Map tags, identifiers, permissions, suppliers, audiences and claims. Select one low-consequence use, such as ordering already-approved help content for consenting customers. Define consumer benefit, lawful basis, PECR position, baseline and prohibited features.
Days 31–60: shadow and test. Generate rankings or creative without live activation. Review profiles for sensitive inference and stale data. Test consent, rejection and withdrawal. Evaluate claims and imagery across audience groups. Red-team prompts and supplier access. Agree metrics and stop rules.
Days 61–90: release gradually. Use a limited audience, channel and approved content set. Keep price and sensitive targeting out of scope unless separately governed. Review objections, complaints, differences in outcomes, technical drift and security weekly. Reconcile every live tag against the register.
At day 90, expand only if the team can explain which data changed which experience, why it was lawful and fair, and how a person can opt out without losing the service.
Define marketing pause gates
Pause activation or generation when:
- consent, lawful basis or direct-marketing status cannot be demonstrated;
- a rejected or withdrawn tracker continues loading;
- suppression fails or objections are not honoured promptly;
- special-category or vulnerability inference enters an audience unexpectedly;
- generated content makes an unsupported claim or false endorsement;
- total price, eligibility or key terms are wrong or hidden;
- complaints, unwanted contact or group differences cross thresholds;
- a supplier, model or tag changes outside review;
- audience data is exported to an unauthorised destination;
- rollback cannot remove the live variant across channels.
The safe state uses a non-personalised experience, approved static creative and central suppression. Preserve evidence, correct affected campaigns and tell people where a material error reached them.
Good personalisation reduces irrelevant effort while keeping choice and meaning intact. The mature marketing stack is not “cookieless” or autonomous. It is legible: every signal has a purpose, every claim has evidence, every channel respects preferences and every experiment can be stopped.



