Zovda
Every data product shows a chart with two years on the axis. Almost none of them explain that the left half came from a different collection process than the right half.
History is the assumption that breaks comparisons, because the further back a dataset goes, the more it was assembled by other means.
Backfill is a separate pipeline
Collecting going forward and reconstructing backwards are different jobs. Forward collection can be scheduled and audited. Reconstruction depends on what still exists, what the source will still return, and how far back it is willing to go.
A set that does both without labeling which rows came from which is a set where the trend line is partly an artifact.
Deleted content is gone
The older the window, the more of the original material has been removed: posts deleted, accounts closed, items made private. A count from three years ago is a count of what survived, not of what happened.
That failure is not uniform. It hits controversial topics hardest, which is exactly where the history is most valuable.
Coverage declines going back
Plot the density of items per day across the whole archive. The curve almost never stays flat, and where it bends is where your comparisons stop being valid.
That bend belongs in the data card, in the same font as the total.
Write the coverage curve down
For each month: how many items, from how many accounts, and which fields were present. Three numbers per row, and the archive becomes readable.
Zovda maintains these collections through Datuva, so the points where the collection changed stay attached to the series. The data research collection shows how a defined historical window is assembled.
Tell us how far back your question goes. We reply within one business day.