GovFeeds Data: Collection and Quality FAQ

Figures below that refer to "the research snapshot" come from the data snapshot of 8 July 2026 used in our academic paper on local government Facebook communication (posts published 1 October 2025 to 8 July 2026). GovFeeds coverage has grown since then.

1. How collection works

Where does GovFeeds data come from?

GovFeeds records publicly available posts from the official social media accounts of U.S. city and county governments: content any member of the public can see. For each post we record its text, its publication time and its public engagement counts. We do not collect from private accounts, private groups or messages.

How often is an account collected?

Collection is continuous and rolling rather than a single fixed schedule, so posts are observed at varying times after publication, and a post can be observed more than once. Engagement counts are snapshots taken at the time of observation. In the research snapshot, the median time from publication to engagement capture was 4.4 days, and the paper's supplementary analyses show its estimates are robust to capture timing.

Because engagement keeps growing as a post ages, comparisons should use posts of similar age where possible.

Are edited or deleted posts detected?

Each post is stored once, under its platform post ID. When a post is observed again, its record is updated with the latest values, so engagement counts can rise over time and a later text edit replaces the earlier text. GovFeeds does not keep an edit history.

On Facebook, when a post is no longer publicly available (deleted by the government or made private), it is flagged as removed once that has been confirmed on two separate checks at least a day apart, and it stops appearing in the GovFeeds app. Removed posts stay in the research archive, marked as removed, with their last observed counts. If a removed post becomes available again, the flag is lifted.

How are engagement counts measured?

Engagement is recorded as it is publicly displayed at the time of capture: total reactions (all reaction types), comments (including replies) and shares. Because capture is rolling, a count reflects engagement accumulated up to the moment of capture, not a fixed interval after publication.

We checked accuracy with an independent re-observation of 1,199 sampled posts. In the strata representative of the panel (780 posts), recorded and re-observed counts correlate at Spearman ρ = 0.847 for reactions, 0.853 for comments and 0.892 for shares; 66% of comment counts and 65% of share counts agree exactly; and disagreements skew upward (the re-observed count is higher), the pattern expected when engagement keeps accruing after capture. For the most engaged posts, recorded counts can exceed other public displays because GovFeeds counts all reaction types and all replies.

2. Which governments are included

How is the list of government accounts built?

GovFeeds keeps a registry of the official accounts of U.S. city and county governments. Each government is identified by its stable 2022 Census of Governments ID rather than by its name, since many place names occur in more than one state. An account is attributed to a government only on evidence from the account itself or the government's own website: the account's displayed name reads as that government or one of its departments, or the government's official website links to it.

How were accounts verified for the research snapshot?

The snapshot registry held 5,252 Facebook pages, and all of them were verified. 89.1% matched the registered government directly. The remaining 571 were checked by the name each page displays: 512 confirmed the registered government, 18 tracked a wrong or department-level page or carried a corrupted registry name (0.3% of the registry), and 41 could not be confirmed. The 59 governments that could not be confirmed were excluded from all analyses. Separately, 6,498 duplicate rows from Virginia independent cities, tracked under both a city and a county-equivalent identity, were deduplicated, keeping the city record.

3. The research corpus

Which posts were excluded, and why?

From a raw corpus of 1,065,286 rows:

1. 6,498 duplicate Virginia independent-city rows were removed, keeping the city record.
2. 182 rows with unreadable publication timestamps, and 58,345 rows published outside the 1 October 2025 to 8 July 2026 window, were dropped.
3. 16 rows whose engagement capture time preceded publication were dropped. Rows were retained when engagement capture completed, including rows where no engagement values could be read (see "Metrics-Missing/Zero" below), giving engagement coverage of 78.5% of the remaining rows.
4. 50 accounts (9,642 rows, 1.2%) that turned out not to be the general-purpose page of their registered government were excluded: pages belonging to another entity, another government or a private individual.
5. 13,500 rows with a false-zero signature (zero reactions recorded alongside nonzero comments or shares) were removed; re-observation confirmed this signature as a capture error.

The final research panel is 761,919 posts from 4,746 governments.

What does "Metrics-Missing/Zero" mean?

It is the status recorded when a post was captured but no engagement values could be read from it. Such a post enters the panel with reactions, comments and shares set to zero. It differs from the false-zero rows removed above, which recorded zero reactions alongside nonzero comments or shares and so contradicted themselves.

A stratified sample of 180 posts in this class was re-observed: 118 displayed positive reaction counts (median 6.5, maximum 240) and 62 displayed none. So roughly two-thirds of the class may record a zero for a post that did draw engagement. For that reason every estimate in the paper was re-run on the 566,796 posts whose capture fully completed, and the results hold when the class is excluded. Separately, 63 of 90 sampled posts that were never captured displayed positive engagement: a missing capture does not mean zero engagement.

How many false-zero rows remained after correction?

None. The false-zero rule is deterministic and was applied to every row, not to a sample, and all 13,500 matching rows were removed. What the rule cannot catch is a post that recorded zero on all three counts despite drawing engagement; that case belongs to the Metrics-Missing/Zero class above and is covered by the robustness re-estimate.

What content categories are used, and how were they validated?

Each post carries one of thirteen mutually exclusive content categories: Events & Community Activities; Public Safety & Police; Sustainability & Environment; Recreation, Parks & Tourism; Infrastructure & Planning; Government Information & Boards; Business & Economic Development; Community Engagement & Volunteer; Health & Wellness; Technology & Digital; Library & Education; Animal Services; and Other/Miscellaneous.

To validate them we drew a stratified sample of ten posts per category (130 in total, each at least 40 characters long). A language model (Claude Opus 5) assigned each post to a category from the written taxonomy alone, given the post text, the government's name and its state, without seeing the GovFeeds label. It reproduced the GovFeeds label for 93 of 130 posts (71.5%; Cohen's κ = 0.69), or 74.2% excluding the residual Other/Miscellaneous class. Almost every disagreement was a boundary case between two defensible labels. This shows the labels are substantially reproducible from the text and the taxonomy; it does not show they are the labels an independent human coding team would assign.

Questions

Researchers with questions about GovFeeds data, or interested in using it, can email bruce@govfeeds.com. See also How Governments Are Included.

Last updated November 2026.