Data in the Upworthy Research Archive

The Upworthy Research Archive is a dataset of headline A/B tests conducted by Upworthy from early 2013 into April 2015. This page documents the archive and answers common questions. We have also published an academic paper that reports the details of the archive and our work to validate the data. Please cite this paper when using the archive.

You can download the archive on the Open Science Framework at osf.io/jd64p/.


For background on the Upworthy Archive, please consult and cite the following sources:

Units of Observation

The Upworthy Research Archive contains packages within tests. On Upworthy, packages are bundles of headlines and images that were randomly assigned to people on the website as part of a test. Tests can include many packages.

The archive only includes aggregate results on the number of viewers a package received and how many of those viewers clicked on that package. It does not include any individual-level information to differentiate between viewers.

Inclusion/Exclusion Criteria

This research archive includes valid tests conducted by Upworthy in the study period. We have omitted tests that were never shown to viewers (zero impressions) and packages that had missing test IDs.

Exploratory and Confirmatory Datasets

To support reliable scholarly research and education, we are releasing the Upworthy Research Archive as a partial exploratory dataset. We will share a confirmatory dataset with researchers whose analysis plans have been peer reviewed (read more about the process).

The exploratory dataset includes 22,666 packages from 4,873 tests. The confirmatory dataset includes 105,551 additional packages from 22,743 tests.

To support time-series research, both datasets are a random sample stratified by week number.

illustration showing that the exploratory and confirmatory datasets are a random sample stratified by week number

Selecting Headlines for Comparison

We expect that many researchers will want to data-mine the archive for specific headline types and compare them to other headlines within the same tests. We created this task as a workshop and homework assignment for students in a Cornell class on the design and governance of experiments. Students were asked to meta-analyze the effect of including a notable person’s name in a headline, and the effect of including a number in a headline. We offer the materials below as food for thought when developing your own data-mining approach:

Columns in the Upworthy Research Archive

The dataset of packages contains the following columns:

Time-related columns:

Experiment-related columns:

Stimuli shown to viewers:


Miscellaneous columns that may be of interest. To our knowledge, none of these columns represent information shown to viewers as part of A/B tests:

Columns we learned about through conversations with former staff:

Talk To Us Before You Scrape & Merge Data from the Web

We have also been scraping Upworthy and the Internet Archive in search of supplementary information, including images. Since only some tests and packages can be supplemented in this way, we are doubtful that this data will be useful for confirmatory research. Please contact us if you think that these columns might be important to your research.