Generative AI has made it inexpensive to create a different visual for every audience segment. The harder question is no longer whether a brand can personalize an ad. It is how much the ad should reveal that it knows about the person seeing it.
A preprint released on September 11, 2026 offers a useful, carefully bounded signal. In a preregistered, within-subject study, 100 US-based native-English participants rated AI-generated ads for four products at three personalization levels. Light personalization improved attitudes toward the ad and the product versus a generic version. It did not produce the same clear improvement in purchase intention. The strongest personalization did not add value and was rated as substantially more creepy.
This was a study of static images, not video. It involved a limited sample and product set, and it did not measure long-term sales. The responsible conclusion is therefore not that personalization has stopped working. It is that brands need to choose the resolution of personalization before they commission dozens—or thousands—of generated versions.
Relevance and surveillance can rise together
Personalization activates two forces at the same time.
The first is relevance. A familiar use case, problem, or product configuration reduces the work a viewer must do to understand why the product matters.
The second is surveillance anxiety. When an ad exposes highly specific details about someone’s appearance, family, finances, health, or private experiences, the viewer starts asking where that information came from and why it is being dramatized. The 2026 study also identified a production problem: as more attributes are pushed into an image prompt, the model has more opportunities to omit, distort, or combine them incorrectly. An ad can feel invasive even when it is accurate—and absurd when it claims to be personal but gets the person wrong.
This is also a policy issue. Google’s current personalized-advertising policy restricts the use of sensitive interest categories, including health, negative financial status, political affiliation, race and ethnicity, and relationship hardship. It also limits personalized advertising to minors, the connection of personally identifiable information with anonymous or pseudonymous data, and overly narrow audiences. The US Federal Trade Commission has separately warned about designs that manipulate consumer choices, obscure material terms, or induce people to share more data.
Technical access is not creative permission. A data field that can be read should not automatically become a visual cue.
The ONCE Personalization Resolution Ladder
The research compared three experimental levels. For production, ONCE uses a more practical five-level ladder. It is not a relabeling of the paper’s conditions; it is a commissioning tool for video and creative teams.
| Level | What changes | Example | Default decision |
|---|---|---|---|
| 0. Universal product truth | Structure, operation, materials, limits and real use that must not vary by viewer | Ports, dimensions, installation, included accessories | Lock it. Film it or control it strictly. |
| 1. Public context | Region, season, placement, time of day or public cultural context | Rainy commuting, holiday gifting, vertical social placement | Usually suitable for scaled variants. |
| 2. Declared need state | A problem the shopper explicitly states or reveals through the current task | Reducing desk cables, packing lighter for travel | Useful when wording avoids implying hidden surveillance. |
| 3. Product configuration | The SKU, color, bundle or feature that matches the need | Plug type, capacity, accessories, local-market version | Connect to catalog and inventory; lock product facts. |
| 4. Identity or sensitive attributes | Name, appearance, family structure, health, finances or private experience | A generated character made to look like the viewer or an ad naming a sensitive condition | High risk; exclude from automated production by default. |
For most ecommerce brands, the productive zone is Levels 1–3. An ad can be more useful because it shows the right situation, answers the declared task, and presents the correct SKU. It does not need to recreate the viewer.
Commission a Personalization Production Contract
When the marginal cost of another version approaches zero, “deliver 50 videos” becomes a weak brief. A brand should require five auditable deliverables.
1. Product truth master
Capture the product details that no version is allowed to change: physical form, operation, scale, packaging, limits and correct model. AI may replace the setting; it must not invent the product. ONCE’s earlier framework for separating stable footage from dynamic video variables is useful here.
2. Variable map
List every element that may change: context, relationship, occasion, opening problem, SKU and call to action. Attach an allowed data source to each variable. Many effective variations need only public context or current-session intent—not a persistent personal profile.
3. Data provenance and consent sheet
For every field, answer: Did it come from public context, an explicit user input, consented first-party data, or model inference? Is the creative use compatible with the original purpose of collection? How long may the field be retained? Can the production vendor see it? If the team cannot answer, the field should not enter the prompt.
4. Exclusion matrix
Write down prohibited sensitive attributes, misleading inferences, minor data, excessively narrow segments, and identity features that may not be generated. Platform rules are a floor. The brand still needs its own reputation standard.
5. Render and approval log
Keep the product master, variables, prompt or workflow version, human approver, intended placement, and withdrawal condition for every final render. Pairing this with an AI asset provenance record prevents responsibility from disappearing as volume rises.
Start with three resolutions, not fifty variants
A low-risk pilot needs one product, one market, the same offer, and the same placement. Create only three groups:
- A generic version built from product truth alone.
- A context version adapted to a public situation or season.
- A need-state version responding to an explicitly stated shopping task without identity or sensitive attributes.
Conduct a human preflight before launch. Check product accuracy, natural human representation, catalog alignment, and whether the visual implies knowledge the brand should not have.
Then monitor two classes of signal. Relevance measures might include completion rate, click-through, and product-page behavior. Risk measures should include ad hides, negative comments, complaints, policy rejections, wrong-SKU renders, and generation-error rates. If a small survey is possible, ask “This ad is relevant to me” and “This ad knows too much about me” as separate questions. A click caused by surprise or discomfort is not automatically persuasion.
Define stop conditions before scaling: a sensitive trait appears, an unauthorized face or identity feature is generated, a product fact changes, a platform warning is triggered, or the higher-resolution version produces a meaningful rise in aversion. There is no universal threshold; it should reflect the category, market and risk tolerance.
The production decision for brands connected to China
For brands that manufacture in China, the most robust workflow is to build a high-quality product-truth master and controlled scene components close to the supply chain. Local-market teams can then combine those assets with public context, declared needs, and the correct SKU.
This concentrates high-value production where product access, engineers, samples and variant inventory are available. It also prevents identity simulation and sensitive inference from becoming the default definition of “personalization.”
AI makes one-to-one creative technically possible. It does not make knowing everything about everyone necessary. Mature personalization does not impress the customer with how much the brand knows. It helps the customer see, faster and with less friction, why the product fits the task in front of them.
Sources and scope
- “Enabling and Understanding Personalization in AI-Generated Advertising Imagery”, preprint released September 11, 2026.
- “I Am AdMan: A Pipeline for Automatic Generation of Personalized Advertising Imagery,” companion technical preprint, September 11, 2026.
- Google Ads, “Restricted targeting in Personalized advertising.”
- US Federal Trade Commission, “Bringing Dark Patterns to Light.”
The research findings above apply only to the study design and sample. The five-level ladder, production contract, and pilot method are ONCE’s analytical recommendations, not industry standards or guarantees of performance.
