Imagine a creative meeting that begins with thirty screenshots from three ad libraries. The team can say which competitors are using reviews, discounts or new talent. It cannot say when each ad appeared, who paid for it, where it sent people, which market it addressed, or what decision the next production should test.
The problem is not a shortage of examples. It is the absence of a data model. Enforcement of the EU Digital Services Act is pushing large-platform ad repositories toward fuller creative records, better search and API access. For global brands and production partners, that shift creates a chance to build durable creative intelligence—provided public delivery records are not mistaken for private performance data.
This is more than another place to browse ads
Article 39 of the Digital Services Act requires very large online platforms and search engines to maintain public ad repositories through a searchable, reliable tool that supports multi-criteria queries and APIs. The required information includes the ad content, the advertiser, a different payer where applicable, the delivery period, main targeting parameters and reach. Records must remain available for one year after an ad was last presented.
The law specifies fields; it does not make every repository equally useful on day one. On 5 December 2025, the European Commission accepted TikTok commitments on advertising transparency, including full ad content and targeting criteria selected by advertisers. On 15 July 2026, the Commission accepted an X corrective-action plan covering search filters for ad content and targeting criteria, full ad content and destination URLs, and repository API access. X was given six months to implement the plan, so these are supervised commitments—not proof that every feature is already live everywhere.
The confirmed structural signal is narrower and more useful: public ad activity can increasingly be observed as dated, sourced records rather than occasional screenshots. The sources do not establish that cross-platform normalization, historical completeness or performance attribution has been solved.
What a repository record proves—and what it does not
A record can establish that a specific creative was presented during a period on a platform. Depending on the repository, it may also identify the advertiser, payer, destination, disclosed audience parameters and reach. A sequence of records can reveal when a brand changes its promise, adds a format or localizes the same product differently across markets.
It does not reveal profit. Reach is not conversion, and longevity is not ROAS. An ad may remain active because it performs, because it has a small always-on budget, because approval or replacement is slow, or because the team has not refreshed it. Public repositories do not normally provide the full spend, margin, attribution window, incrementality design or internal test decision.
That is why an intelligence note should separate two sentences:
- Observation: Three product-demonstration openings appeared for the same range in Germany and led to one product-family landing page.
- Hypothesis: The advertiser may be testing “result first, explanation second,” which is worth comparing with our own product and audience.
Remove “may,” and intelligence turns into a story the data never told.
The ONCE TRACE method
TRACE is a five-step workflow for moving from public records to an original production question. It is not a competitor score and not a copying queue.
T — Trace the delivery facts
Save the source platform, repository URL or record ID, collection date, advertiser, payer, start and end dates, market, disclosed targeting and reach fields, destination URL and creative version. A screenshot can preserve visual state, but it should remain attached to its source and timestamp. Detached screenshots lose decision value quickly.
R — Read how the creative works
Replace “premium,” “authentic” or “UGC-like” with comparable mechanics. What does the opening establish? When does the product appear? What is promised? Is the proof a demonstration, person, review, data point or scenario? Where does the offer appear, and where does the CTA lead?
For video, record how voice, on-screen text and imagery divide the explanation; whether the message survives without sound; and whether the production is live action, creator-led, animation or a hybrid. Two unboxing videos may do different jobs: one proves setup simplicity, while another creates launch ritual. A generic “unboxing” tag cannot guide production.
A — Assign stable, defined tags
A tag needs an inclusion rule. Use “product proof” only when the creative shows a checkable action, structure or result. Use “social proof” only when a source, identity or third-party judgment is actually present. Automated classification can draft tags, but it should not invent a new vocabulary every week.
Durable fields usually include market, category, channel, opening mechanism, core promise, proof type, role of the person, offer structure, CTA, production mode, claim risk and landing-page type. Human review is still needed to merge synonyms, interpret context and mark uncertainty.
C — Challenge every inference
Add a “missing condition” beside each observation. No spend means no investment comparison. No conversion means no winner claim. Platform scope does not equal the full media plan. Failure to find an ad does not prove it never ran.
If the creative makes environmental, health, safety or performance claims, flag the underlying substantiation for separate review. A competitor’s live ad is not permission to repeat its claim.
E — Experiment with your own question
The output should not be “make one like this.” It should be a testable production hypothesis: for a product that requires setup explanation, does showing the first visible result in the opening outperform a pain-point-first version for our audience?
The brand then defines its own product truth, audience, channel and single variable before buying the required shots. A competitor repository supplies questions; only first-party testing can supply the answer.
Separate the library into three layers
The public-record layer preserves original records, collection time and platform fields with minimal interpretation. The interpretation layer stores tags, shot analysis, analyst, confidence and unresolved checks. The decision layer records which observations entered a brief, what was tested, the asset version and whether the result supported the hypothesis.
Separation matters when the law, platform interface or taxonomy changes. Teams can reprocess source records without rewriting every historical decision. It also makes permissions clearer: public competitor observations may support a client discussion, while one brand’s private performance data must not leak into another client’s library.
For a production company, the compounding asset is not the largest collection of ads. It is a stable creative grammar: what counts as demonstration proof, what is promise-first, which shots carry atmosphere, and which shots carry a purchase decision. Public repositories supply observations; authorized client work supplies outcomes. They should be stored separately and connected only within agreed permissions.
A useful cost model is:
Cost per adopted creative hypothesis = (collection + normalization + human analysis + review + maintenance) ÷ hypotheses used in approved briefs.
The denominator is not the screenshot count. If research changes no brief, shot or test, more automated collection is inventory, not intelligence.
When a continuous system is worth it
Ongoing monitoring is more valuable when a category changes quickly, a brand sells across several markets, competitors refresh creative often, or the team launches new tests every month. In that case, define the watchlist, collection cadence, tag dictionary and review owner, and periodically remove fields that no longer answer a decision.
A one-time project scan may be enough for brands with a few major launches each year, stable competitors, or decisions driven mainly by channel research and sales feedback. Do not build a “data asset” nobody will maintain. A minimum viable pilot can start with one decision question, a small comparable brand set, stable fields and one complete handoff from observation to test brief. Any sample count is a working choice, not an industry benchmark.
Before the next creative review, remove labels such as “best ad,” “winner” and “viral” unless the evidence actually supports them. Replace them with observed mechanism, missing condition and test hypothesis. The discussion will sound less certain—and become much more useful.
If your team needs to turn public creative research into scripts, shot matrices and controlled variants, see ONCE’s cross-border e-commerce product video solutions. A strong ad creative intelligence library does not tell you whom to copy. It makes every production brief answer one question: what exactly are we trying to learn this time?
Sources were checked on 22 October 2026. The core evidence is DSA Article 39, the European Commission’s 5 December 2025 TikTok announcement and its 15 July 2026 X action-plan announcement. This article addresses public-record research and production decisions, not legal advice, and it does not claim identical feature availability across every market or account.
