Zovda
A researcher finds the thread that proves the point. The link is dead. The account is gone. The post that started the discussion was removed before the study began.
This is normal, not unlucky. Public conversation is written, edited, and removed continuously, and any dataset built once from today's state will be a dataset of whatever survived.
Everything decays, and it decays unevenly
Deletion rates are not uniform. They cluster by account type, by topic, and by time. A panel that samples evenly today and then ages will quietly drift toward the kinds of accounts that never delete anything.
That drift is the single most common flaw in longitudinal work built on conversation data, and it is invisible from inside a single snapshot.
Coverage is a grid, not a number
"Ten million records" says nothing about whether the records cover the period and the segments you need. Coverage is a grid: time window on one axis, account segment on the other, content type on a third.
A supplier who cannot draw that grid cannot tell you what the dataset will support. They can only tell you how big it is.
Stable ids are the product
If the same post gets a new internal id on every export, you cannot track it across time, and every follow-up study starts from zero.
Ask for ids that survive re-export, plus a version field per record. Then a longitudinal panel becomes possible: the same objects, measured repeatedly.
Snapshots answer different questions than panels
A snapshot tells you what was visible on one date. A panel tells you how something changed. Most research questions that sound like the first are actually the second, and the difference is decided at collection time, not at analysis time.
Decide which one you are buying before the collection window opens. Six months later it is too late to add the earlier observations.
The checklist we run
- Can the provider show coverage as a time × segment grid?
- Do record ids stay stable across exports?
- Is there a version field for records that changed?
- Does the manifest state how deletion was handled?
- Can the same panel be extended later without restarting?
If a supplier cannot answer these, the row count is real. The study it supports may not be.
Where to start
Start from the question you need to answer over time, and work backwards to the coverage window it requires. The data research catalogue describes how the research collections are organised, and the social media data API covers the case where you want to build the panel yourself.
Zovda maintains these longitudinal collections through Datuva, so the dates above survive a refresh and an item deleted upstream does not become a hole downstream.
Tell us the period and the segments you need. We reply within one business day, and the field documentation is shared once the scope is confirmed.