Zovda
Almost every social dataset is collected through a ranked interface: a feed, a search result page, a recommendation. The ranking was built to maximize attention, and attention is not a random sample.
Then the dataset is analyzed as if it were one, and the conclusions are reported with the confidence of a survey.
Ranked retrieval is not a sample
A ranked list returns the items the system believes you will engage with, which correlates with popularity, recency, and what you already looked at.
That means the dataset over-represents the loud, the recent, and the familiar. Every claim about "what people think" inherits those three biases at once.
What to randomize
Where the interface allows it, collect across time rather than across popularity: sample the stream at fixed intervals instead of taking the top of it.
Where it does not, collect the full set for a narrow slice rather than a partial set for a wide one. A complete small window is worth more than a top-heavy large one.
Report the frame, not just the size
The size of a sample says nothing about what it represents. Report the frame: which accounts were eligible, which were reached, and which were returned.
Then report what the frame probably excludes. That paragraph is the most useful thing in the memo.
Test with a boring query
Run a deliberately uninteresting query and look at what comes back. If the results are all high-engagement, the interface is ranking, and your collection is a highlight reel.
Zovda assembles these collections through Datuva, where the collection frame stays with the export. The social media data collection shows how the frame is recorded.
Tell us the population you need to describe. We reply within one business day.