What Our Malaysian Ad Archive Can and Cannot Tell You

The archive holds 1,428,677 Facebook and Instagram ads across all markets, of which 1,043,499 carry an advertiser classified as Malaysian. 222,921 of those were still running the last time we checked that specific ad. The start-date window matters more than the totals do. 98% of the Malaysian ads sit between 2025-05-25 and 2026-08-21, with a median start of 2026-04-20, so this is a recent archive rather than a historical one. That is a real limit on what can be asked of it: a question about how Malaysian advertising changed over several years cannot be answered here, and any figure framed as a trend across that span would be reading collection growth as market change. The deeper limit is that collection follows advertisers we track rather than a sampling frame. There is no version of this dataset that is a census of Malaysian advertising, and no weighting that would turn it into one, so every count published anywhere on this site is a floor. Where a page says how many ads do something, the honest reading is always at least this many.

A field that exists on the schema is not a field you can count with. The sample below is 2,000 Malaysian ads drawn evenly across the whole start-date range rather than off the recent end, because sampling the newest ads would measure the latest collection pass instead of the archive. Three results change what may be asked. Ad body copy is effectively universal at 99.8%, so a keyword study over body text has an honest denominator. The English renderings do not: bodyEnglish sits at 34% and hookEnglish at 33.4%, because they exist only where the original was not already English. A numerator drawn from those fields against a whole-corpus denominator is not a floor, it is uninterpretable, and that rules out a whole class of question. Video transcription is present on 36.9% for the same structural reason, and it was removed from the search index besides, so no keyword count can reach it at all. The model-generated labels are the opposite case: creative angle and industry are attached to 100% of sampled ads. Universal coverage is not the same as accuracy, though. They are model output, they are multi-label, and the classifier has changed over the archive's life, so they carry cross-sections and never trends.

The search index returns at most 100 distinct values for any field, and it returns them in alphabetical order rather than by frequency. On a field with more values than that, one query gives you a list that looks complete and has an arbitrary tail missing, with nothing in the response saying so. Measured here, theme, language, brandName all exceed the cap. That has direct consequences: the advertiser field cannot be enumerated in one pass, so a question about how many distinct Malaysian brands advertise needs a partitioned crawl rather than a single query; creative angle has to be counted against the classifier's own controlled vocabulary rather than by asking what values exist; and any language study has to partition until every partition comes back under the cap and then prove the parts add up to the whole. Industry, ad format and publisher platform stay under the cap and can be listed directly. Every study we publish either respects that or does not ship. The tooling behind these pages now refuses a facet response that comes back at exactly the cap, because a list that is silently short is worse than an error.

Run length and every seasonal timing question depend on one field, so it gets checked rather than trusted. The code that writes it falls back to the current time when Meta supplies no start date, which would stamp an ad with the moment we ingested it. Fake starts of that kind would cluster on the days we happened to be collecting, and in a study that buckets ad starts by week they would read as a launch ramp that never happened. It cannot be detected by filtering, because the ingest timestamp is not a queryable field. It can be detected by arithmetic. A real start date is a calendar day rendered at a fixed offset, so its remainder against a full day is the same constant for every ad written that way, while an ingest stamp carries whatever wall-clock time it happened to be written at and is therefore very nearly unique. Across the sample, 2 offsets account for every ad, 7 hours on 1,526 and 8 hours on 474, and the share carrying an offset that almost nothing else shares is 0%. The fallback is effectively never exercised on Malaysian ads, so weekly bucketing of start dates is sound. The second check is less comfortable. Ad format is written by two different code paths that disagree about whether a dynamic creative is filed as dynamic or as an image, so the boundary between those two values is writer-dependent. Any split between video and static on this data is a range rather than a point, and every page that cuts by format says so.

The list below is the part of this page worth citing. Most published ad datasets describe what they contain; almost none publish what they structurally cannot measure, which is what a reader actually needs in order to know whether a number is safe to repeat. The short version: no spend, no cost per anything, no impressions, no reach, no return, ever. Those fields are not in Meta's public Ad Library, so any figure of that shape derived from this archive would be invented, and any Malaysian benchmark quoting them without naming its source deserves the same suspicion. No targeting, because it is not published and the audience fields here are model inferences from the creative. No claim that an ad performed, because run length is right-censored and says only that an advertiser kept paying. Every one of those refusals is a row in the CSV with an empty value and its reason attached, so a machine reading this dataset gets the limits along with the numbers.

The whole audit is one CSV under CC BY 4.0: counts, the coverage window, per-field population rates with the sample size attached, the enumeration limits, both contamination checks, and the refusal rows. Use it including commercially with attribution to AdPlay.ai. If you cite a figure from any of our other data pages, cite this one beside it, because the caveats live here rather than being repeated in full on every page.