Paid Social Ad Data

Where the paid social record actually lives

Paid social is the hardest of the major channels to work with as evidence, and the reason is a split between what is public and what is useful.

Meta's Ad Library is genuinely public and searchable by advertiser and by keyword, with an API for programmatic access. It is also narrow. The advertiser-side detail an attorney needs — spend, targeting configuration, attribution-window settings, delivery, conversion counts, audiences, and the account's own change log — is visible only inside Ads Manager and the Ads Insights API, which means inside the account or through production from whoever holds it.

So the record on this page is marked reconstructable rather than in the account, and the distinction is practical rather than pedantic. What an examination assembles is a composite: account exports for as far back as the windows still reach, billing and payment records for the spend those exports no longer cover, the advertiser's own tag and pixel configuration, and the advertiser's order or CRM records as an independent count of the outcomes the platform claims to have caused. Any one of those alone is a partial record. Together they support a finding.

The first practical consequence: an early, complete export while access exists is worth more than any later analysis, because the windows below are shorter than most litigation timelines.

The three windows, and the date they changed

Meta's retention for its reporting API is not one number. It is three, and they were cut effective January 12, 2026, across all API versions simultaneously with no legacy exemptions. From Meta's own developer announcement, read August 14, 2026:

Meta's wording on the first row is that total values “are unaffected by the above changes and will continue to be available for up to 37 months”; the second and third are “limited to 13-month's of historical data” and “limited to 6-months of historical data” respectively.

Do not carry a Google figure across to Meta or the reverse. They differ, they moved on different dates, and a memo that applies one platform's clock to another is wrong in a way opposing counsel can demonstrate in a sentence. Google's 37-month figure for hourly, daily and weekly reporting took effect June 1, 2026 and happens to match Meta's top row; nothing else lines up.

The same Meta change deprecated the 7-day and 28-day view-through attribution windows, 7d_view and 28d_view. That matters beyond retention: any comparison of conversion counts across January 12, 2026 is comparing two measurement regimes, and a report that does not say so is exposed.

What the Ad Library shows, and what it has stopped showing

The Ad Library is frequently over-read, in both directions. Precisely what it holds:

• For ordinary commercial ads: creative and run dates. It does not show spend. It does not show targeting. It does not show performance, and it does not show which ads were delivered to whom.
• For ads about social issues, elections or politics: additional disclosure, including spend — but reported in ranges, not as a figure. Meta's 2018 announcement described the archive as showing “the campaign budget associated with an individual ad and how many people saw it – including their age, location and gender,” which remains roughly the outer limit of what is publicly knowable about such an ad without account access. The announcement is worth reading where a matter turns on what the archive was meant to contain.

The more important point is that it is not a permanent archive. Meta committed at launch to seven years' storage for ads with political content, and removals on that schedule began in 2025. Regional rollout dates have been reported by third parties rather than published by Meta, so treat a specific regional date as reported rather than established — and treat the deletion itself as real.

Two working rules follow. Preserve Ad Library pages by capture, not by reference: a URL in a memo is not evidence, and the page behind it has an expiry. And do not state a retention figure for ordinary commercial ads — the seven-year commitment is documented for political and social-issue content, and no primary source gives a window for the rest.

Why a reported conversion is not an audited number

The short answer to carry into a deposition: platform-reported conversions are not auditable in the sense a forensic accountant would recognize.

A conversion is recorded when the platform's tag or SDK fires on the advertiser's property and the platform attributes that event back to an ad interaction using its own model and its own lookback windows. The firing is observable. The attribution step is proprietary and unobservable.

What can be independently examined: the advertiser's tag implementation, server logs, CRM records and order data. Comparing platform-reported conversions against the advertiser's own systems of record is the standard approach — but discrepancies of meaningful size are routine and expected, because the platform counts attributed events while the order system counts transactions. A gap is the baseline condition, not a finding. A finding needs a magnitude, a direction and a mechanism.

What cannot be independently examined: the attribution logic, the modeled portion of the count, and any de-duplication across surfaces. Two platforms measuring the same purchase will each claim it, and neither is wrong by its own definitions.

On accreditation, say only what is true. The Media Rating Council's accreditation letter for Google Ads, dated March 31, 2026, covers clicks and invalid clicks as reported in that product's interface. I have located no equivalent letter for Meta's advertiser-facing conversion metrics. Accreditation is process assurance against a published standard — not a click-level ledger, and not an audit of attribution.

Rebuilding the record from the advertiser's own side

Because the platform's record thins with age and cannot be audited at the attribution layer, the reconstruction runs through material the advertiser controls. In the order I would ask for it:

1. A full account export, taken now. Campaign, ad set and ad level, at daily granularity, across the longest range the API still returns. This is the item that expires, and the 13-month and 6-month fields go first.
2. Billing and payment records. Financial records rather than reporting artifacts, and they outlive the reporting windows. Where the analytics record has aged out, billing often still establishes what was spent and when.
3. The tag and pixel configuration, including its change history and any tag-manager container versions. A conversion definition change between two compared periods invalidates the comparison, and this is where it shows up.
4. Order, CRM or booking records for the same period, as an independent count of outcomes.
5. Server logs, entirely within the advertiser's control and usually the longest-lived record in the stack.
6. Captures of the creative and the landing pages, which change without leaving any platform trace on the advertiser's side.

Note what this does not produce: a verified version of the platform's numbers. It produces a second, independent count from records with their own custody, against which the platform's count can be compared and the difference described.

Headline metrics and the litigation around them

A separate caution applies to planning-side numbers, as distinct from performance-side ones. Meta's advertiser-facing reach estimates have been the subject of sustained class litigation: in DZ Reserve v. Meta Platforms, Inc., No. 3:18-cv-04978, advertisers allege the “Potential Reach” metric was inflated by the inclusion of duplicate and fake accounts. Class certification was granted March 29, 2022 and affirmed by the Ninth Circuit on March 21, 2024, and the Supreme Court declined review on January 13, 2025.

Those are allegations and a procedural history, not findings, and the case page I have relied on for that history is a plaintiff-side source. Its value here is narrower and still real: a platform's own headline estimate is a contestable object rather than a fixed fact, and a claim built on a planning estimate rests on softer ground than one built on billed spend.

The general form of the point holds for any paid social matter. Billed spend is a financial fact with a paper trail. Delivery counts are the platform's measurement. Reach and frequency estimates are the platform's modeling. Conversions are the platform's attribution of outcomes it did not observe end to end. Those four carry very different evidentiary weight and should never appear in a single undifferentiated table.

What I will not be able to tell you

The limits here are unusually hard, and stating them early is the difference between a defensible report and one that unravels under examination:

• Whether a reported conversion actually happened. The tag firing can be examined; the attribution cannot. The platform's count is its attributed estimate under its own model.
• What share of a conversion figure was modeled rather than observed. No platform exposes that split in a form a third party can verify, and any percentage asserted without account-specific evidence is a guess.
• What a competitor spent, unless the ads carry political or social-issue disclosure — and then only in ranges.
• Who saw an ad. The Ad Library does not show delivery, and the account shows aggregate delivery rather than a recipient list.
• What the account looked like beyond the windows. Once a window closes there is no backup an advertiser can request, and no preservation letter extends it. No platform's public documentation says a hold changes a retention period; assume it does not.
• Whether an export is what it purports to be. A file a party downloaded and produced carries no certification by itself. That is a procedural question rather than a data question, and it belongs with counsel.

Preserving it before the windows close

The preservation problem in paid social is front-loaded, and it is worse than in search on two counts: the windows are shorter for the granular fields, and the access is more often held by a person rather than an entity — a business manager tied to an individual's login, an agency's asset, a departed employee's account.

A hold that names only “the Facebook ads account” will miss most of it. What has to be reached: the ad accounts themselves, the business manager layer above them and its user and permission history, billing and payment records, the pixel and any tag-manager containers with their version histories, the analytics property and its retention setting, third-party reporting or dashboard tools that hold their own copies, the CRM, the communications where directives to pause or change budget live, and captures of the creative and landing pages.

Do the export first and the theory second. Everything else in a paid social matter is recoverable later; the granular platform record is not, and the 6-month frequency row is already gone for any conduct older than two quarters.

When this data is not where the answer is

Three situations where an examination of paid social data will not earn its cost.

The dispute is about what was promised. If the disagreement is over what a contract required — a delivery commitment, a cost-per-result target, a definition of a qualified outcome — the platform data shows what happened without touching what was owed. It can quantify the gap once the standard is settled, and not before.

The claim rests entirely on the platform's own conversion number. That figure is the platform's attributed estimate and cannot be audited at the attribution layer. With no independent order or CRM record to compare it against, an opinion built on it is built on the opposing narrative's own instrument.

The conduct is older than the windows. If the granular record is gone and no export was taken, an examination can describe what billing shows and little more — sometimes enough for a wasted-spend theory, almost never enough for a measurement one. Better established in the first conversation than in the third invoice.

Frequently Asked Questions

Published 2026-08-14 · Last reviewed 2026-08-15