What is knowable, and how each answer is labelled.
Public ad libraries publish creative, dates and, inside the EU and UK, real targeting and reach. They publish no spend and no performance. Everything this product sells is built on the first list, and the second list is printed on every report.
Read directly
observedFacts as published. No model sits between the source and the answer.
- What does the ad say and show?
- The ad library plus a vision pass that states objective facts about the creative.
- How long has it run, and how many variants sit behind it?
- Library start and end dates, the variant count, and our own daily snapshots.
- What did the advertiser set as targeting?
- Meta's official Ad Library API: ages, gender and locations. Ads delivered in the EU or UK only.
- Who did the ad actually reach?
- The same API's reach breakdown by age, country and gender. EU and UK only, and there is no equivalent anywhere else.
Classified, with confidence
classifiedA model's judgment, scored. Every one of these carries a probability, and a low one is shown as unclear rather than rounded into a clean answer.
- Hook, format, offer, awareness stage, angle
- 330 narrow questions across 9 requests, each answered against written criteria rather than a free-text opinion.
- Does the landing page keep the ad's promise?
- The fetched page compared with the ad, plus a deterministic cross-check on the offer string, the discount and the product name. The bar for showing a flag is 90% precision on hand-labelled pairs, and a check that has not been measured yet is marked unverified.
- Who does the creative speak to?
- Estimated from the copy and the creative. Where EU reach data exists for the same brand, that estimate is scored against it rather than trusted.
Offered as a suggestion
suggestionA starting point for a test, never a prediction.
- What targeting would I test?
- The implied audience turned into interest suggestions and sized through Meta's own API. Since January 2026 Meta treats interests as hints: location and minimum age are the hard constraints, so these are briefing signals, not delivery guarantees.
Not available, at any price
never claimedNo ad library on any platform publishes these for commercial ads. A tool that shows them is modelling them.
- A competitor's spend, CTR, CPA or ROAS
- Nothing publishes it. We do not estimate it, and no report contains a number shaped like it.
- Share of voice, or share of a competitor's budget
- Meta's inventory has no denominator, so the fraction cannot exist.
- Which placement an ad ran in
- The libraries publish platform (Facebook, Instagram), never placement.
How a label is made
Narrow questions, written criteria, and a number attached to every answer.
A vision pass first states objective facts about the creative: what is on screen, whether a person appears, when the brand first shows. Code reads those facts directly. They are never re-asked as an opinion.
The judgment calls go to a classifier as 330 narrow questions in 9 requests, each with written criteria for every option. Accuracy on this kind of work depends almost entirely on how atomic the question is, which is why there is no single prompt that asks a model to “analyse this ad”.
Ad copy, captions and landing pages are untrusted text. They travel as data, never as instructions, so an ad cannot talk the classifier into anything.
Answers come back with probabilities. A hand-labelled gold set fixes the threshold per question, so an unclear answer stays unclear instead of being rounded up. Below the threshold, a second model acts as referee.
The model version is pinned, never floating. A nightly canary re-runs a fixed slice and flags drift, and any question that scores badly gets rewritten or split rather than quietly kept.
Statistics sit on top of the labels, not inside them. Counting is by distinct brand and concept, intervals resample brands rather than ads, and comparative wording is gated on a multiple-comparison correction.
The reporting rules, which are product features.
These are enforced in the report templates and in the API responses, not left to whoever writes the summary.
Traction, never winner.
A long-running ad is one someone keeps funding. It is a lead worth investigating. Independent evidence that longevity predicts performance is mixed, and we say so rather than quietly relying on it.
A pattern in 90% of ads is a market norm, not an insight.
If everyone does it, doing it will not move you. Norms are reported as norms.
No share-of-voice or market-share claims.
There is no denominator to divide by.
No placement claims.
Feed, Reels and Stories are not in the data.
At most three ads per brand in any top list.
Otherwise the brand with the biggest budget writes your whole report.
Every insight cites a number or a named advertiser.
A sentence that cites neither is an opinion, and it gets cut before you see it.
Claims ladder
Three rungs, and the product may only stand on the one it has earned.
| Rung | It may say | Which needs | It never says |
|---|---|---|---|
| Descriptive | N brands and M distinct concepts use this combination. Two brands account for X% of it. | De-duplicated counts | Anything about performance |
| Associated | Ads combining X and Y stay live longer than the niche baseline, with the interval and the evidence badge shown. | Support floor, shrinkage, corrected q-value, interval | This works. This will lower your CPA. X causes Y. |
| Predictive, inside one account | In your account, concept type X is associated with lower CPA, validated on held-out ads. | Your own ad account plus a holdout or a lift test | Carrying one account's result over to the whole niche |
The accuracy scorecard is not published yet, because it does not exist yet.
Every question in the bank will get its own score, measured against a gold set labelled by hand and blind. Until that pass has run, there is no honest number to show you, so there is no number on this page.
Validation in progress
The gold set is being labelled now. When it is done this page gets a per-question scorecard with agreement, calibration and the threshold each question runs at, including the questions that scored badly and what happened to them.
Until then, the honest position is the one on this page: confidences are shown on every classified tag, an unclear answer says so, and no accuracy figure appears anywhere on this site or in any report. If you want to test it before that, read the sample report and check the tags against the ads yourself.