An ad-transparency archive is a repository, published by an advertising platform, of the ads it carried: the creative, who paid, when it ran, and sometimes how it was targeted and how many people saw it. Meta's Ad Library, Google's Ads Transparency Center, TikTok's Commercial Content Library and the ten other repositories the Digital Services Act obliges are all of this kind.

For a measurement researcher, an archive is an unusually good instrument and an unusually treacherous population. It is good because it is bulk, retrospective, queryable and free: no crawler, no sock puppets, no rate limit worth the name, and a window that reaches years further back than you could ever crawl. It is treacherous because the platform decides what goes in it, using a classifier of its own, and both directions of that classifier's error land in your denominator. A prevalence computed on an archive is a prevalence among ads the platform put in the archive, which is not the same as ads that ran, and is very much not the same as ads that were submitted.

This page is about that gap: which archives exist and what each one covers, what an archive structurally cannot contain, how the field has measured the archives' own error rates, and what you have to hold alongside an archive to say anything about the ads it is missing. It is a child of Platforms, which covers the other ways into a platform; the archive is the one route that page identified as a distinct instrument deserving its own treatment.

Five papers. The first four are chosen because each attacks the archive's accuracy from a different side; the fifth is what you do when you decide the archive cannot answer your question.

• [1Edelson, Laura; Lauinger, Tobias; McCoy, Damon (2020): "A Security Analysis of the Facebook Ad Library", in: Proceedings of the IEEE Symposium on Security and Privacy. (DOI)] (IEEE S&P 2020) — the archive as the object of study, and still the foundational paper. Analyses Facebook's US political Ad Library (3,685,558 ads) and finds that “9.7 % of all ads in the Ad Library do not include a disclosure string”, that “15.8 % of total ad spend in the Ad Library cannot be attributed due to missing disclosure strings, or would be misattributed due to disclosure string fragmentation”, and that “we could not retrieve 16,160 ads on 7,515 pages using the API, despite repeated attempts”. Read it for the taxonomy of ways an archive can be complete and useless at the same time.
• [2Le Pochat, Victor; Edelson, Laura; Van Goethem, Tom; Joosen, Wouter; McCoy, Damon; Lauinger, Tobias (2022): "An Audit of Facebook's Political Ad Policy Enforcement", in: Proceedings of the USENIX Security Symposium. (Link)] (USENIX Security 2022) — both error directions, measured simultaneously, on 33.8 million ads retrieved from the Ad Library. A 55% false-positive rate among the ads Facebook itself flagged as political in the US, a 4.5% false-negative rate worldwide, and — the finding that should change your study design — false negatives of 0.85% in the United States against “considerably worse in other countries”. A single global recall figure hides a factor of fifty.
• [3Bouchaud, Paul; Liénard, Jean F. (2024): "Beyond the Guidelines: Assessing Meta's Political Ad Moderation in the EU", in: Proceedings of the ACM Internet Measurement Conference. (DOI)] (IMC 2024) — the same two-sided analysis on the DSA-expanded archive, which since August 2023 holds “all advertisements targeting individuals within the EU”. Across 16 EU countries, “only 7.7% of undeclared political ads … were moderated as political by Meta” while “60.4% of ads moderated by Meta did not align with Meta's criteria for political advertising”. Read it for how to calibrate a classifier against the platform's own published guidelines rather than against your intuition.
• [4Benzaamia, Abir; El Fraihi, Asmaa; Abdelaziz, Ines; Goga, Oana (2026): "A Year Under the DSA: Ad Transparency's Uneven Landscape", Proceedings on Privacy Enhancing Technologies 2026(2):517-532. (DOI)] (PoPETs 2026) — the archives measured against a donated ground truth, across four platforms at once. Matching 48,511 real ads collected through the Who Targets Me extension back to the platforms' DSA repositories yields “181 out of 351 matches on Facebook (48% missing from the repository) and 399 out of 438 on Instagram (9% missing)”. Two repositories under the same legal obligation, five times apart on completeness.
• [5Gkiouzepi, Eleni; Andreou, Athanasios; Goga, Oana; Loiseau, Patrick (2023): "Collaborative Ad Transparency: Promises and Limitations", in: Proceedings of the IEEE Symposium on Security and Privacy. (DOI)] (IEEE S&P 2023) — the alternative, and its limits. Argues that ad libraries cannot deliver targeting transparency and proposes inferring the targeting formula from a panel of monitored users instead. Honest about the cost: correct inference in 17 of 45 controlled campaigns, rising to “65%” only “where our ad has been received by ten or more monitored users”.

This is the first thing to get straight, because the four have different populations, different failure modes, and live on different pages of this wiki. Papers use them interchangeably in prose and not at all interchangeably in method.

The confusion is not academic. [6Bryson, Kevin; Borem, Arthur; Moh, Phoebe; Akgul, Omer; Edelson, Laura; Lauinger, Tobias; Mazurek, Michelle L.; McCoy, Damon; Ur, Blase (2025): "Characterizing the Usability and Usefulness of U.S. Ad Transparency Systems", in: Proceedings of the IEEE Symposium on Security and Privacy. (DOI)] (IEEE S&P 2025) is a usability study of 22 US ad-transparency systems — the second and third rows, not the first — and finds that “more than two-thirds of all questions and more than half of desired actions went unanswered or unaddressed”. [4Benzaamia, Abir; El Fraihi, Asmaa; Abdelaziz, Ines; Goga, Oana (2026): "A Year Under the DSA: Ad Transparency's Uneven Landscape", Proceedings on Privacy Enhancing Technologies 2026(2):517-532. (DOI)] measures the second row (48,511 explanations) and uses it to audit the first. [7Breuer, David; Becker, Lucas; Hollick, Matthias (2026): "Ad Personalization and Transparency in Mobile Ecosystems: A Comparative Analysis of Google's and Apple's EU App Stores", Proceedings on Privacy Enhancing Technologies 2026(1):604-630. (DOI)] (PoPETs 2026) wanted the first row as ground truth for an experiment on the second and concluded it could not be used. If you write “ad transparency” in a methods section without saying which of the four you queried, a reviewer cannot tell what your denominator is.

For the third row see consent and Platforms; for the fourth, Biases on what a donation panel is a sample of.

DSA Article 39 requires an ad repository of every platform the European Commission designates as a very large online platform or search engine, and the inventory below is the thirteen repositories Mozilla and Check First listed when they tried them all in 2024. Read that as the set somebody has actually exercised, not as the current designation list: the Commission's list has moved since and was last updated 24 July 2026.¹⁾

These are not comparable instruments, and the differences are exactly the ones that decide whether your study is possible.

The Papers column counts papers in this corpus that used or measured the archive, from the hand-audit in Use in Publications. Statuses and scopes were checked against each provider's own documentation on 2026-09-08; every URL and quote is on ad_archives.

Three readings of that table matter more than the rows.

• Nobody in these seven venues has published from any archive other than Meta's and Google's. TikTok's Commercial Content Library, X's Ads Repository, LinkedIn's, Bing's, Snap's and the retail platforms' appear in zero of 5,855 papers. That is not evidence they are unusable — X's repository hands you the whole EU commercial-communications dataset as a CSV with no application at all, which is the least gated bulk access of anything on this page. It means there is no methods section to copy, and if you use one you are the first, in these venues.
• The scopes do not line up, so a cross-platform rate is a cross-scope rate. Meta's archive is worldwide for political ads and EU-only for everything else; TikTok's and X's are EU-only for everything; Google's political section is worldwide but restricted to verified advertisers. A “share of ads that are X” computed across platforms is comparing different populations unless you cut all of them down to the intersection, which in practice means the EU and one year.
• Bulk access is the discriminator, not the existence of a repository. Of the thirteen, four offer something you can compute on at scale: Meta's API, Google's BigQuery dataset, TikTok's API and X's CSV. The rest are search boxes. A repository you can only query one advertiser at a time is a transparency gesture, not an instrument.

The one independent cross-platform comparison is not from these venues: Mozilla commissioned Check First to stress-test twelve of them against DSA-derived criteria, published 16 April 2024, and concluded that “even the best approaches don't meet our baseline” and that “our accuracy testing found many cases where ads in the user interface were not located in the ad repository”.¹⁴⁾ It is now two years old, which is why the statuses above were re-checked rather than taken from it — but it is the only source that has tried them side by side, and it should be your starting inventory rather than a vendor listicle.

Concrete, because the archive-by-archive table above does not say how long you will be waiting. Verified against each provider's own documentation on 2026-09-08.

• Meta Ad Library API. You need a Facebook account, then identity and location verification at facebook.com/ID — the same process an advertiser must complete to run political ads, and Meta's own page says “It can take a few days to confirm the information you submit” — then a Meta for Developers account and an app. Queries then go through the Graph API, and Meta publishes a script repository with a command-line interface. So: days, and a verified identity tied to your real name and location. Plan it before the deadline, not after.
• Google's BigQuery public datasets. No application, no account review: you need a Google Cloud project and you pay for bytes scanned, not for the data. bigquery-public-data.google_political_ads (table creative_stats) is the political subset; bigquery-public-data.google-ads-transparency-center is the whole centre. Minutes, and a credit card. This is the cheapest bulk archive access described anywhere on this wiki.
• TikTok Commercial Content API. A TikTok for Developers account and an application TikTok reviews. Platforms records a stated turnaround of “within 4 weeks” for the sibling Research API. Weeks, and an institutional eligibility test.
• X Ads Repository. No application at all: pick a country and a date on the page and download the commercial-communications CSV. Minutes. Nobody in these seven venues has done it, so nobody has reported what the file actually contains — see Open Questions.

The pattern is worth naming: the archives with the most published research behind them are the ones with the highest access friction, and the two with no friction at all have none. That is not a statement about data quality. It is a statement about who has looked.

This is the part to put in your paper as a limitation, and the part reviewers of archive studies now ask about. None of these are bugs; they follow from what an archive is.

• Only ads that ran. An archive is populated at delivery. Ads that were submitted and rejected, ads that were never submitted because the advertiser expected rejection, and ads that ran on a platform outside the archive's scope are all invisible. If your question is about the platform's gatekeeping, the archive is the wrong side of the gate.
• Not the content of removed ads. DSA Article 39(3) is explicit: where the platform “has removed or disabled access to a specific advertisement based on alleged illegality or incompatibility with its terms and conditions, the repository shall not include” the ad's content, the person it was presented on behalf of, or the payer.¹⁵⁾ So the ads a compliance study most wants to read are precisely the ones the law says must be redacted. Google's own FAQ says the same in product terms: an ad may be missing because it “was later taken down for a policy violation”, and where an ad is withheld you get “Sorry, we're not able to show you this ad” with the advertiser name and spend retained.
• Only advertisers the platform verified. Google's political archive contains “Election ads … that were paid for by verified election advertisers”, and the FAQ's answer to why an advertiser is missing is “Only advertisers that have been verified through the verification process for election ads appear”. A prevalence over that archive is a prevalence among verified political advertisers, and the interesting cases are often the unverified ones.
• Whatever the platform's classifier decided was in scope. For political archives this is the dominant error term, and it runs in both directions at once; see The Two-Sided Error Problem. [8Silva, Márcio; Oliveira, Lucas Santos de; Andreou, Athanasios; Melo, Pedro Olmo Stancioli Vaz de; Goga, Oana; Benevenuto, Fabrício (2020): "Facebook Ads Monitor: An Independent Auditing System for Political Ads on Facebook", in: Proceedings of the ACM Web Conference. (DOI)] put it plainly for Brazil: “not all political ads we detected were present in the Facebook Ad Library for political ads”.
• Aggregates, not counts, and rounded ones. Spend and impressions come as ranges. Google's impressions are explicitly non-unique — “if it was shown to the same user 10 times, that counts as 10 and not 1” — per-platform impression breakdowns exist only from 4 September 2023, and are “published after a delay of 90 days”. Any spend total you compute is the sum of range midpoints, and you should report which end you took.
• A window, and a moving one. One year for the DSA all-ads repositories, seven for political archives. That is short enough that a longitudinal design has to collect prospectively rather than reach back.
• A history that changes under you. Google states that “Google is continuously reviewing its data and improving its classification of election ads. As we constantly increase the accuracy of our data, historical data may slightly fluctuate.” An archive is not an immutable snapshot: two queries a year apart can disagree about 2022. Date and archive your pull, and see Longitudinal.
• One archive can have two surfaces that disagree. [7Breuer, David; Becker, Lucas; Hollick, Matthias (2026): "Ad Personalization and Transparency in Mobile Ecosystems: A Comparative Analysis of Google's and Apple's EU App Stores", Proceedings on Privacy Enhancing Technologies 2026(1):604-630. (DOI)] queried Google's Ads Transparency Center both ways and reports that “The data shown on the web interface seems incomplete compared to the records in the BigQuery dataset”, with “multiple instances where the web interface does not show any text ads for an advertiser in the Play Store. The dataset, however, lists multiple matching entries” — and the dataset's own creative_page_url for those entries “resolves to a page-not-found error”. Say which surface you queried; a reviewer replicating you on the other one will get a different number.
• No targeting, for most of them. Meta and TikTok publish targeting parameters; several others publish country and language only. [5Gkiouzepi, Eleni; Andreou, Athanasios; Goga, Oana; Loiseau, Patrick (2023): "Collaborative Ad Transparency: Promises and Limitations", in: Proceedings of the IEEE Symposium on Security and Privacy. (DOI)] exists because “Similar limitations exist in Political Ad Libraries (public repositories aggregating political ads run on the platform)” — the archive tells you an ad ran, not who it was aimed at.

An archive built around a category — “political”, “social issue”, “election” — is the output of a classifier the platform runs and does not publish. That classifier has a false-positive rate and a false-negative rate, and an archive study inherits both. Papers that report only one of them are reporting half the error.

Four papers in this corpus have measured it. They used different platforms, years, regions and ground truths, and they agree that both rates are large.

Three consequences for a study design, in the order they will bite you.

1. A single global rate is not a rate. [2Le Pochat, Victor; Edelson, Laura; Van Goethem, Tom; Joosen, Wouter; McCoy, Damon; Lauinger, Tobias (2022): "An Audit of Facebook's Political Ad Policy Enforcement", in: Proceedings of the USENIX Security Symposium. (Link)]'s spread from 0.85% to 45% across countries is the most important number on this page. If your study is about one country and you cite a platform's global transparency performance, you have imported an error term that may be fifty times off in either direction. Report per-region.
2. Which side you can measure depends on what you hold beside the archive. False positives need a definition of the category that is independent of the platform's — [3Bouchaud, Paul; Liénard, Jean F. (2024): "Beyond the Guidelines: Assessing Meta's Political Ad Moderation in the EU", in: Proceedings of the ACM Internet Measurement Conference. (DOI)] used Meta's own published guidelines as the definition and a calibrated text classifier as the instrument, which is the cleanest trick on this page. False negatives need ads the platform did not archive, which the archive cannot give you by construction: you need a second collector.
3. Do not silently treat the archive as ground truth. [7Breuer, David; Becker, Lucas; Hollick, Matthias (2026): "Ad Personalization and Transparency in Mobile Ecosystems: A Comparative Analysis of Google's and Apple's EU App Stores", Proceedings on Privacy Enhancing Technologies 2026(1):604-630. (DOI)] states that the authors “initially intended to refer to these datasets as ground truth to evaluate our experiments but found them affected by several issues”, and for the ads they were observing directly the failure was total: “the ads shown in Google's Play Store are entirely missing from the Ad Transparency dataset”. If your design has the archive on both sides of a comparison, it cannot detect its own gap.

Since the archive cannot tell you what it is missing, an archive study needs a second, independent view of the same ads. The field has used four, and they trade off in the obvious way: the closer you get to what a real user saw, the smaller and more biased your sample.

Two practical notes. Matching is where these studies actually fail, not collection: [4Benzaamia, Abir; El Fraihi, Asmaa; Abdelaziz, Ines; Goga, Oana (2026): "A Year Under the DSA: Ad Transparency's Uneven Landscape", Proceedings on Privacy Enhancing Technologies 2026(2):517-532. (DOI)] links each donated ad to a repository record “when a unique match can be determined using shared identifiers”, and its headline 48%-missing figure is really “48% we could not uniquely match”, which is an upper bound on missingness. Say which it is. And the donation route is the one that is growing — 25 papers in this corpus mention data donation, 11 of them in 2025 — but only 4 mention Who Targets Me and 5 Ad Observer or an Ad Observatory, so the panels themselves are still nearly unused in these venues.

Both Who Targets Me and Ad Observer were still operating when checked on 2026-09-08; the Australian Ad Observatory's first phase closed in March 2024 and a second phase runs to 2027.¹⁶⁾

Ad archives are one of the few measurement instruments on this wiki that exist because a law requires them, which means their scope is set by legal text you can read rather than by product decisions you cannot.

For services the European Commission has designated as very large online platforms or search engines, Article 39 of the Digital Services Act requires an ad repository “through a searchable and reliable tool that allows multicriteria queries and through application programming interfaces”, covering “the entire period during which they present an advertisement and until one year after the advertisement was presented for the last time”. The mandated fields are the ad content, the person it is presented on behalf of, the payer if different, the run period, whether it was targeted “including where applicable the main parameters used to exclude” groups, branded content identified under Art. 26(2), and “the total number of recipients of the service reached and, where applicable, aggregate numbers broken down by Member State”.

Three things follow that a fresh reader gets wrong.

• “Through application programming interfaces” is in the text. A platform that offers only a search box is not obviously compliant. That is a lever, and Mozilla's stress test found several platforms with no API at all.
• One year is the floor, not the practice. Political archives at Meta and Google run to seven years. The DSA all-ads repositories run to one.
• Article 39 is not Article 40. Art. 39 is the public ad repository, open to anyone; Art. 40 is the vetted-researcher regime for non-public data, which Platforms covers and which nobody in this corpus has published through. Conflating them is the most common error in this area, and this corpus shows how thin the DSA literature really is: “Digital Services Act” appears in 29 of 5,855 papers, and Article 39 appears near a DSA mention in 4.

The EU's political-advertising regulation — Regulation (EU) 2024/900, the “TTPA” — applies from 10 October 2025 and mandates a European repository for online political advertisements run by the Commission, with transparency notices retained “for a period of seven years after the last publication”. Commission Implementing Regulation (EU) 2026/818 of 9 April 2026 sets out its common data structure and API.

Both of the platforms whose archives this entire literature is built on responded by leaving the market.

• Meta, 25 July 2025: “From early October 2025, we will no longer allow political, electoral and social issue ads on our platforms in the EU”, citing “significant operational challenges and legal uncertainties”. Its developer blog confirmed the switch: “Beginning on October 6, 2025 at 6:00 PM Central European Time (CET) … social issue, electoral and political ads will no longer be able to be delivered in the EU and associated territories.”¹⁷⁾
• Google, 14 November 2024: “Google will stop serving political advertising in the EU before the TTPA enters into force in October 2025”, because the TTPA “defines political advertising so broadly that it could cover ads related to an extremely wide range of issues that would be difficult to reliably identify at scale”. The corresponding policy change is in the Advertising Policies change log for September 2025.¹⁸⁾

For a measurement plan in 2026 the consequences are concrete, and they cut in opposite directions.

• EU political-ad archives are now historical corpora. The seven-year retention means [3Bouchaud, Paul; Liénard, Jean F. (2024): "Beyond the Guidelines: Assessing Meta's Political Ad Moderation in the EU", in: Proceedings of the ACM Internet Measurement Conference. (DOI)]'s and [14Capozzi, Arthur; Morales, Gianmarco De Francisci; Mejova, Yelena; Monti, Corrado; Panisson, André (2023): "The Thin Ideology of Populist Advertising on Facebook during the 2019 EU Elections", in: Proceedings of the ACM Web Conference. (DOI)]'s windows remain queryable, but nothing new arrives. A “current state of EU political advertising” study is not available through this instrument, and a longitudinal series must be labelled as ending in October 2025.
• The EU commercial-ad repositories are unaffected and under-used. DSA Art. 39 covers ads of every kind, and that obligation did not move. Meta's all-EU-ads scope, X's CSV and TikTok's API are all still there, and zero papers in these venues have used the latter two.
• Political-ad measurement outside the EU is unaffected, and Meta's worldwide political archive still has its 7-year window. [15Coelho, Bruno; Lauinger, Tobias; Edelson, Laura; Goldstein, Ian; McCoy, Damon (2023): "Propaganda Política Pagada: Exploring U.S. Political Facebook Ads en Español", in: Proceedings of the ACM Web Conference. (DOI)] and [2Le Pochat, Victor; Edelson, Laura; Van Goethem, Tom; Joosen, Wouter; McCoy, Damon; Lauinger, Tobias (2022): "An Audit of Facebook's Political Ad Policy Enforcement", in: Proceedings of the USENIX Security Symposium. (Link)] are US studies and their route is intact.
• The corpus has not noticed any of this. Regulation (EU) 2024/900 appears, under any phrasing, in zero of 5,855 papers — including the two 2026 papers whose subject is EU ad transparency. Whatever you write about the EU here in 2026, there is no prior methods section that accounts for the cliff.

All figures below come from the publication corpus of seven venues (CCS, IMC, NDSS, PETS, USENIX Security, TheWebConf, IEEE S&P), 2010–2026, 5,859 extracted papers of which 5,855 have extractable full text. The script is scripts/adarchives_report.mjs and its unedited output, every query and the full hand-audit are on ad_archives.

A paper counts if an ad-transparency archive is a data source for it, or if the archive itself is what it measures. A bibliography entry does not count. A mention of an Android advertising SDK does not count, and that is the whole problem: “ad library” is a homonym, and the SDK sense is far more common in these seven venues than the archive sense.

The 44 lowercase-only papers sit outside the candidate set by construction, so they were screened rather than audited: the 9 whose lowercase mention falls within 80 characters of a transparency-archive word were read by hand, and they are 7 Android advertising SDKs, 1 lowercase bibliography entry (USENIX's small-caps reference style lowercases titles, so [9Ali, Muhammad; Goetzen, Angelica; Mislove, Alan; Redmiles, Elissa M.; Sapiezynski, Piotr (2023): "Problematic Advertising and its Disparate Exposure on Facebook", in: Proceedings of the USENIX Security Symposium. (Link)]'s citation of [1Edelson, Laura; Lauinger, Tobias; McCoy, Damon (2020): "A Security Analysis of the Facebook Ad Library", in: Proceedings of the IEEE Symposium on Security and Privacy. (DOI)] lands here), and 1 pure substring collision with no advertising content at all — “popular file uplo[ad librar]ies (e.g., express-fileupload)” in a paper about Node.js file uploads. The other 35 have no archive word near the mention. No archive user is hidden among the 44, but that is a screen and this page says so rather than calling it an audit.

Auditing all 29 papers matched by a wider, case-sensitive candidate probe against their own full text:

Precision of the wide candidate probe: 11 of 29 = 37.9%. Precision of the probe Platforms used, for this question: 10 of 69 = 14.5% — and it also missed one genuine paper, [7Breuer, David; Becker, Lucas; Hollick, Matthias (2026): "Ad Personalization and Transparency in Mobile Ecosystems: A Comparative Analysis of Google's and Apple's EU App Stores", Proceedings on Privacy Enhancing Technologies 2026(1):604-630. (DOI)], which uses Google's archive and never says “Ad Library”. Every rejected candidate is printed with its reason on the provenance page, because a residue that lives only in a local script output is a residue nobody looks at.

Per venue: TheWebConf 3, IEEE S&P 2, IMC 2, PoPETs 2, USENIX Security 1, NDSS 1. Per year: 2020:2, 2021:1, 2022:1, 2023:3, 2024:1, 2025:1, 2026:2. 2025 and 2026 are provisional venue-years — CCS 2026 and IMC 2026 have not been held, and IEEE S&P 2026 and TheWebConf 2026 abstracts are not in OpenAlex, so those years are under-represented by construction rather than by relevance (corpus).

Eight of the eleven used Meta's archive; three used Google's; none used any of the other eleven DSA repositories. Set against a floor of 180 papers in this corpus whose measured phenomenon is advertising-shaped, the audited set is 10 of them — about 5.6% of advertising measurement in these venues touches an archive.¹⁹⁾

Each figure with the paper's own denominator, and each quote verified against the paper's full text (see ad_archives for the check, which needed a pypdf fallback for 11 of 51 strings).

Every query, the report script and its unedited output, the full hand-audit with its rejected residue, the quote checks with their fallback modes, and every external source with its fetch date and the ones we rejected are on ad_archives. Corpus-level caveats — the seven-venue scope, the provisional 2025–2026 slice, field stability — are on corpus.

Four limits worth restating here. The corpus is seven security, privacy and measurement venues: political-advertising measurement also lives in ICWSM, CHI, FAccT, Political Communication and the internet-policy journals, none of which are in it, so eleven papers is a count for these seven venues and certainly not for the field. The audited set is small enough that every figure about it is a count, not a rate — do not compute confidence intervals on eleven papers. The archive-by-archive table rests on provider documentation, which is the primary source but also the interested party. And the negative rows — TikTok 0, X 0, LinkedIn 0 — are probe results over full text: a paper that used X's CSV without naming the repository is invisible to them, though at these counts that is unlikely to hide more than one or two papers.

An archive study should let a reader reconstruct the population the platform handed you. Report:

1. Which archive, which access route, and which version — web portal, API and API version, BigQuery dataset, or bulk CSV. The three routes into Meta's archive return different things: [1Edelson, Laura; Lauinger, Tobias; McCoy, Damon (2020): "A Security Analysis of the Facebook Ad Library", in: Proceedings of the IEEE Symposium on Security and Privacy. (DOI)] used both the API and the aggregate Report and found them inconsistent.
2. The archive's own scope clause, quoted: region, ad category, the platform's definition of that category, and the retention window. This is your denominator sentence.
3. The pull date, and re-pull once. Historical data mutates. Two pulls a month apart, with the diff reported, costs nothing and pre-empts a reviewer.
4. The matching procedure, if you compared against anything, and how many ads failed to match. “Missing from the repository” and “we could not match it” are different claims and only one of them is about the platform.
5. Both error directions, or an explicit statement of which one you cannot measure and why.
6. Per-region rates, not a global one, wherever the archive spans regions.
7. Which end of the range you used for spend and impressions, and whether impressions are unique.
8. What you can deposit. Archive terms constrain redistribution; deposit your ad IDs, your classifier and your matching code even where you cannot deposit creatives. See Artifacts.

• Platforms — the other routes into a platform, and why an archive is treated separately.
• Existing datasets — an archive is somebody else's dataset with somebody else's inclusion rule; the same discipline applies.
• Website classification — the classifier you will need to define “political” independently of the platform.
• Interrater agreement — validating that classifier against human labels, which every two-sided error study on this page had to do.
• Biases — donation panels, verified-advertiser populations, and what a convenience sample of consenting users is a sample of.
• Longitudinal — an archive whose history mutates, and a series that has to stop in October 2025.
• Requests — measuring the ad as a network request on a page, which is the other half of advertising measurement and a completely different instrument.
• Legal enforcement — what a regulator does with an archive, and what DSA non-compliance proceedings have turned on.
• Artifacts — depositing from a source whose terms restrict redistribution.
• ad_archives — every query, the hand-audit with its full residue, the quote checks and the external sources with fetch dates.

[1]

Edelson, Laura; Lauinger, Tobias; McCoy, Damon (2020): "A Security Analysis of the Facebook Ad Library", in: Proceedings of the IEEE Symposium on Security and Privacy. (DOI)

[2]

Le Pochat, Victor; Edelson, Laura; Van Goethem, Tom; Joosen, Wouter; McCoy, Damon; Lauinger, Tobias (2022): "An Audit of Facebook's Political Ad Policy Enforcement", in: Proceedings of the USENIX Security Symposium. (Link)

[3]

Bouchaud, Paul; Liénard, Jean F. (2024): "Beyond the Guidelines: Assessing Meta's Political Ad Moderation in the EU", in: Proceedings of the ACM Internet Measurement Conference. (DOI)

[4]

Benzaamia, Abir; El Fraihi, Asmaa; Abdelaziz, Ines; Goga, Oana (2026): "A Year Under the DSA: Ad Transparency's Uneven Landscape", Proceedings on Privacy Enhancing Technologies 2026(2):517-532. (DOI)

[5]

Gkiouzepi, Eleni; Andreou, Athanasios; Goga, Oana; Loiseau, Patrick (2023): "Collaborative Ad Transparency: Promises and Limitations", in: Proceedings of the IEEE Symposium on Security and Privacy. (DOI)

[6]

Bryson, Kevin; Borem, Arthur; Moh, Phoebe; Akgul, Omer; Edelson, Laura; Lauinger, Tobias; Mazurek, Michelle L.; McCoy, Damon; Ur, Blase (2025): "Characterizing the Usability and Usefulness of U.S. Ad Transparency Systems", in: Proceedings of the IEEE Symposium on Security and Privacy. (DOI)

[7]

Breuer, David; Becker, Lucas; Hollick, Matthias (2026): "Ad Personalization and Transparency in Mobile Ecosystems: A Comparative Analysis of Google's and Apple's EU App Stores", Proceedings on Privacy Enhancing Technologies 2026(1):604-630. (DOI)

[8]

Silva, Márcio; Oliveira, Lucas Santos de; Andreou, Athanasios; Melo, Pedro Olmo Stancioli Vaz de; Goga, Oana; Benevenuto, Fabrício (2020): "Facebook Ads Monitor: An Independent Auditing System for Political Ads on Facebook", in: Proceedings of the ACM Web Conference. (DOI)

[9]

Ali, Muhammad; Goetzen, Angelica; Mislove, Alan; Redmiles, Elissa M.; Sapiezynski, Piotr (2023): "Problematic Advertising and its Disparate Exposure on Facebook", in: Proceedings of the USENIX Security Symposium. (Link)

[10]

Chen, Baiyu; Tag, Benjamin; Xue, Hao; Angus, Daniel; Salim, Flora (2026): "When Ads Become Profiles: Uncovering the Invisible Risk of Web Advertising at Scale with LLMs", in: Proceedings of the ACM Web Conference. (DOI)

[11]

Venkatadri, Giridhari; Sapiezynski, Piotr; Redmiles, Elissa M.; Mislove, Alan; Goga, Oana; Mazurek, Michelle L.; Gummadi, Krishna P. (2019): "Auditing Offline Data Brokers via Facebook's Advertising Platform", in: Proceedings of the ACM Web Conference. (DOI)

[12]

Zeng, Eric; Wei, Miranda; Gregersen, Theo; Kohno, Tadayoshi; Roesner, Franziska (2021): "Polls, Clickbait, and Commemorative \$2 Bills: Problematic Political Advertising on News and Media Websites Around the 2020 U.S. Elections", in: Proceedings of the ACM Internet Measurement Conference. (DOI)

[13]

Bitaab, Marzieh; Karimi, Alireza; Lyu, Zhuoer; Oest, Adam; Kuchhal, Dhruv; Saad, Muhammad; Ahn, Gail-Joon; Wang, Ruoyu; Bao, Tiffany; Shoshitaishvili, Yan; Doupé, Adam (2025): "ScamMagnifier: Piercing the Veil of Fraudulent Shopping Website Campaigns", in: Proceedings of the Network and Distributed System Security Symposium. (Link)

[14]

Capozzi, Arthur; Morales, Gianmarco De Francisci; Mejova, Yelena; Monti, Corrado; Panisson, André (2023): "The Thin Ideology of Populist Advertising on Facebook during the 2019 EU Elections", in: Proceedings of the ACM Web Conference. (DOI)

[15]

Coelho, Bruno; Lauinger, Tobias; Edelson, Laura; Goldstein, Ian; McCoy, Damon (2023): "Propaganda Política Pagada: Exploring U.S. Political Facebook Ads en Español", in: Proceedings of the ACM Web Conference. (DOI)