A supplier shows your team an AI assessment of an overseas ad that attracted criticism. The explanation is detailed and persuasive. Before reducing local review or signing an annual prediction service, ask when the assessment was produced. Was it recorded before publication, or after the audience response became available?
Both exercises can be useful. They support different claims. For brands developing campaigns with China-based production partners, ONCE recommends reviewing the prediction timeline alongside the creative report.
A current signal, rather than an accuracy guarantee
Digital Culture Group's 3 November 2026 company announcement describes Resonance Studio analysing creative and briefs against live audience signals, then returning scores and proposed changes. ONCE checked it on 5 November. This establishes a supplier's offering, not independently audited predictive accuracy.
An AdExchanger interview report dated 3 November discusses a model comparison using an already-public Converse ad and notes possible influence from actual audience reactions. That is the report's qualification, not ONCE's investigation of the provider's training data or inputs. We use 3 November as the announcement date, not proof of the product's first use.
The buying implication is broader than one vendor: retrospective explanation and prospective prediction should be reported separately. The framework here is ONCE's proposal.
Define which job you are buying
“This copy does not explain who benefits” identifies a discussable creative issue. “This proportion of the intended audience will object” predicts an outcome. The second claim needs a defined audience, market, observation period, outcome and supporting validation.
A service that helps find omissions can be assessed for specificity, verifiability and useful revision advice. It need not promise sales lift. A proposal to replace local reviewers, prevent reputational incidents or predict campaign performance requires evidence suited to those promises.
A score is not automatically a probability. Without an appropriate definition and calibration, 80 points does not mean 80% approval or 20% opposition.
The ONCE prediction timeline
| Stage | Keep the record | Question it resolves |
|---|---|---|
| Submission | Creative version, brief, market, audience and available context | What was actually assessed? |
| Recorded prediction | Full report, run date, stated data scope and outcome being predicted | Was the judgement preserved before the result? |
| Revision and publication | Changes, final version, release location and date | Did the later response concern the assessed version? |
| Follow-up | Observable outcomes, period, human adjudication and unknowns | What occurred, what was not observed and what remains unresolved? |
Suppliers need not disclose another client's confidential work. They should identify which timeline a demonstration represents. An assessment produced after public reaction belongs in a retrospective category. If prior model exposure cannot be established, retain that limitation. Asking a model to pretend not to know later events does not establish that information was excluded.
Use historical examples for the right purpose
Old campaigns can teach issue categories, demonstrate whether a report identifies particular shots and help teams turn comments into production tasks. A plausible retrospective explanation alone does not establish that the same finding would have been made before release.
Nor does the absence of recorded controversy make an old campaign a definitive negative example. Relevant viewers may not have encountered it, or feedback may not have been captured. Allow an unresolved category rather than forcing every case into correct or incorrect.
Run prospective review without turning publication into a gamble
Use authorised, unreleased work already scheduled for ordinary review. Ask the AI and existing reviewers to record their judgements separately before discussion. Keep the normal project owner responsible for publication. Do not deliberately release a known problematic version to create an experiment.
Define the questions before results arrive: possible misunderstanding, a local association needing further investigation, or a specified measurable audience outcome. Do not choose the most flattering criterion afterwards. Confirm which project materials the supplier is authorised to process.
If an AI finding leads to a revision and the released ad attracts no controversy, record that the finding was adopted and the revision completed. It does not directly prove a crisis was prevented: the unchanged version was not released under the same conditions. Useful editorial assistance and validated prediction are different records.
Convert abstract scores into production decisions
Imagine a hypothetical home-goods brand manufacturing in China and commissioning an English-language campaign. A report says a household interaction lacks resonance. Ask which shot or line is involved, who might interpret it differently, what dated context supports that concern and what change preserves the brand's intention.
The eventual issue may be a missing reason for one character to help another. A clarifying action could then be considered before reshooting an entire film. If the concern involves an uncertain local association, seek an informed contextual review. A model's suggested interpretation is not the shared opinion of a whole population. This is an illustration, not an ONCE client result.
Specify report runs, preparation, human adjudication, production changes and later review separately in the scope. More findings do not automatically mean a cheaper review process. Track time spent resolving false alarms as well as useful changes; do not invent an industry pass rate or standard budget percentage.
Keep what was known at the time
As ONCE develops managed content services, a useful asset is a record of what was predicted, why a revision was chosen and what could subsequently be confirmed. A small set of projects supports a limited conclusion, not a promise across every culture or product category.
At the next procurement meeting, ask for one report's submission, run, revision, release and feedback dates. Decide whether you are buying explanation and revision support or prediction backed by relevant validation. When discussing ONCE's production solutions, preserve this distinction in both the brief and the acceptance criteria.
Sources and limits
Digital Culture Group, “Digital Culture Group Launches Resonance Studio, an AI Tool Built to Get Brands Out of the Apology Game,” GlobeNewswire, 3 November 2026; Joanna Gerber, “This Ad Tech Company Wants To Ensure Brands Don’t Go Viral For The Wrong Reasons,” AdExchanger, 3 November. Checked 5 November. The timeline, prospective workflow and home-goods example are ONCE's proposals. No platform test, accuracy measurement or finding of data leakage is claimed. The original cover is conceptual.
