Where the Paid-Search Evidence Lives, and For How Long

Overview

Paid-search evidence sits in three different places, on three different clocks. Where a record lives decides whether you can still get it, what it costs to produce, and how much argument it will take to defend once you have it.

Six pages here divide that way. Platform records are held in the advertising account and export from it, intact, for as long as access and the retention window both hold. Measurement records are composites — the account, the tags on the advertiser's own site, and the advertiser's own order data — and they have to be rebuilt before they mean anything. Quantum — the money — is part measured and mostly modeled, and the part everyone actually fights about is the modeled part. It should be argued as a model rather than presented as a measurement.

The most useful thing a page like this can tell you is how long you have.

Platform records: export them, do not reconstruct them

The advertising account itself is the only tier where the work is retrieval rather than reconstruction. Change history, performance reporting, search terms, auction data, billing and invoices are all held inside the account and come out of it in a form that can be checked against the interface it came from.

Two properties of that tier are easy to miss. The first is that access is a status somebody else may control — the manager account sitting above an advertiser's account can revoke a login in one action. The second is that the same record has different windows depending on how it is pulled. A change history read in the interface goes back two years; the field-level change detail available through the platform's API reaches back thirty days. An attorney who plans a production around the longer figure and gets the shorter one has lost most of the log.

The clocks, with the source and the date read

Every figure below was read from the platform's own published documentation on August 14, 2026, or carries the effective date the platform published. Two of these schedules tightened within the last year, so a retention figure learned even two years ago is likely wrong now.

Three things follow. Windows run from today, not from the date of the conduct — a two-year log covers the two years ending now, and it gets shorter every month a matter sits. A change record on one platform survives four times as long as the same record on another, so a matter touching both has two preservation deadlines. And one row above is not a platform policy at all: the analytics retention period is a dropdown a party's own administrators set, which means a record can be destroyed on schedule, automatically, with nobody taking a further act.

Measurement records are composites, and composites get rebuilt

Conversion and attribution data, and paid social data, behave differently from platform records because the number an attorney is looking at was assembled rather than observed. A reported conversion figure blends conversions the platform observed with conversions it estimated using a model — the platform's own documentation says the column contains both — and no standard report separates them. Anyone treating that figure as a count of sales is treating a statistical estimate as a business record.

Rebuilding it means going to three systems with different owners: the conversion action settings and their change record inside the account, the tag container on the advertiser's site with its own version history, and the advertiser's own order or customer records. A composite is only as old as its shortest component, and the tag container is very often that component — and the one nobody thinks to preserve.

Two figures from two systems will not reconcile to zero, and that is not a finding of wrongdoing. Advertising reporting counts a conversion against the date of the click; analytics counts it against the date of the conversion itself. A report claiming to have reconciled them exactly has either used one shared definition or is over-claiming.

Quantum is modeled, and should be argued as a model

The money divides in two, and merging the halves is the most common error here. Spend that should not have gone out is measured: money left an account, the platform recorded it, the invoices record it, and the two can be set side by side. Lost profit asks what the business would have earned had the program run correctly, and no system holds a record of what would have happened. Every method for producing that number — comparing the plaintiff against its own earlier performance, against a comparable unaffected entity, or against a modeled market — is an argument with assumptions in it, and each assumption is a place the figure can be tested.

Preservation sits in this group rather than with the platform records, because preservation is what makes any of these numbers possible at all. The order matters more than the technique: establish who actually holds access, export while it holds, and only then demand what is missing from the other side.

What will not work

A preservation letter does not stop any of this. It does not change a retention setting, it does not extend a platform's published window, and it does not restore access that has already been revoked. The deletion runs on schedule regardless of what has been served on whom.

Two demands come back empty routinely. A request for the complete list of queries that triggered an advertisement returns an incomplete list, because low-volume terms are withheld. A request for the platform's reasoning behind a filtered click returns a count and a credit, because there is no ledger to produce.

What obligations attach to these records, and what follows from one that no longer exists, is a question for counsel — I am not an attorney. What I can tell you is which tier a given question sits in, what it takes to get the record out, and how many months are left on it.

Published 2026-08-14