How political are shutdown alerts?

If you visited a .gov website during the 2025 government shutdown, you may have noticed alert banners on the top of the page that indicate a suspension or interruption of normal operations for the agency.

Some agencies vary from the standard, however, and instead use this opportunity to deliver a political message.

From a cursory analysis, of the 248 shutdown banners I could find from about 1,150 agency websites, 92 were political to some extent (37%).¹

I wanted to explore this in the context of the following questions:

• what do you think the appropriate way to communicate an interruption of services to the American public?
• Do you think that government agencies should be allowed to include political messages of one or another political party like this? What are the risks of such messaging, if any?
• Should the Executive-branch appointed agency heads have more freedom in their communications strategies, or should agencies be expected to remain apolitical - like federal employees?

• Findings
• Follow-Ups
• Limitations

Findings

• About 37% of the close to 250 website shutdown alerts posted were politicized.

• Agencies with the most politicized shutdown alerts: (listed by very similar wording, messages not shown here for these sites are listed later on in the post)

  • sba.gov and business.gov
  • usda.gov, farmers.gov, and forestsandrangelands.gov
  • cfius.gov, fsoc.gov, sigpr.gov, treasury.gov

  • The radical left has chosen to shut down the United States government in the name of reckless spending and obstructionism. As a result, the U.S. Department of the Treasury’s websites will only be sporadically updated until this shutdown concludes. Please refer to Treasury’s contingency plans for more information.

  • minoritywhip.gov and democraticwhip.gov were also politically charged, albeit in the opposite direction than most agency websites.

• Few government websites highlighted any other impact of the shutdown; if ever, only as means to politicize the debate further:

  • See for instance, sba.gov:

    Special announcement Senate Democrats voted to block a clean federal funding bill (H.R. 5371), leading to a government shutdown that is preventing the U.S. Small Business Administration (SBA) from serving America’s 36 million small businesses.

    Every day that Senate Democrats continue to oppose a clean funding bill, they are stopping an estimated 320 small businesses from accessing $170 million in SBA-guaranteed funding. As a result of the shutdown, we wanted to notify you that many of our services supporting small businesses are currently unavailable.

    The agency is executing its Lapse Plan and as soon as the shutdown is over, we are prepared to immediately return to the record-breaking services we were providing under the leadership of the Trump Administration. If you need disaster assistance, please visit sba.gov/disaster."

• Minor variations were interesting to see; truthfully, at the outset of this project I expected to see a uniform banner with the same wording on every website. Data collection was made much more complex because this was not true.

  • I am curious about how the instructions flowed downward from the Executive branch (I am assuming?) to the front-end developers or CMS managers who had to add these. Who wrote the text for each one? Assistants to agency heads?

  • A “Democrat Shutdown is Imminent” banner with a ticking clock seemed to have caught on in the offices of the White House, as it appeared on a number of websites including whitehouse.gov, omb.gov, budget.gov, maha.gov, eop.gov, crypto.gov, and others.

  • Yet even non-partisan wording varied, with “federal government shutdown”, “lapse of appropriations” and “lapse of government funding”, “lapse of federal funding”. Agencies were often “unable to maintain [this] website.” Banners occasionally described continuing work as “mission-critical activities”, using “available resources”, while websites may only be “sporadically updated” or “reviewed to comply”.

    • In the case of CBP, at cbp.gov, they used this as a opportunity to advertise they were hiring; apparently applications “will not be delayed by the lapse in federal funding in accordance with the One Big Beautiful Bill Act (H.R.1).”
    • This is even as the website would “not be actively managed”.

• The most commonly politicized terms used were “radical-left”, “Democrat-led” or “Democrat shutdown”.

• The most common non-political message was that a lapse of appropriations meant the website was not actively being updated.

• The “largest” message (in terms of screen space) was from the USDA. If anything, any future analysis should use the shape and presentation of the alert to as a factor in the partisan score!

Follow-up questions

• Better partisan ranking
  • for example, minoritywhip.gov had the text:

    WE’RE IN A REPUBLICAN GOVERNMENT SHUTDOWN. They are refusing to protect your health care. They are refusing to negotiate on solutions. Democrats will keep fighting for everyday families because you deserve better.

  • I decided this wasn’t worth looking into further due to how few Democrat-supporting alerts appeared in the data.
• I am curious about how a LLM-driven web data collection platform would have performed collecting these. I would like to do a deeper analysis, but this would require some data clean-up that is time better spent on other projects.
• I don’t find LLM-ranked partisanship estimates, e.g., “give a rating of 1-10”, particularly effective relative to a embeddings-based approach. However, the switch to using embeddings compared to Python libraries which relied on traditional NLP methods was valuable.
• A indicative failure of the data collection and analysis methodology is from dems.gov, which the data collection pulled “https://democrats.house.gov/ house-democrats https://democrats.house.gov/" from as a banner, and the keywork-based, transformers, and LLM approach all ranked as highly partisan.
  • I think a higher quality model would have caught this, though, and ranked it as non-partisan.
• How does this compare to previous government shutdowns?
  • Is it possible to get the archived snapshots from shutdowns in former President Obama’s administration & Trump’s first term, then check those banners to see the change in politicization?

Limitations

This is just a quick exploration to satisfy my curiosity which a) reflects a point-in-time analysis, b) has not been thoroughly checked for correctness, and c) includes LLM-generated code from Claude Sonnet 4.5 and GPT-5.

Snapshots of websites were pulled from visiting each page in this list of domains. For some domains that I had trouble accessing, snapshots from the Internet Archive through the waybackpy Python library were used instead, relying on the most recent snapshot available if one was available since the shutdown began.

Note that there is wide variety in how each website implements these alert banners, if at all. CSS classes and element IDs were not consistently applied; additionally, some websites used a JavaScript implementation, while others used HTML. This complicated parsing; for example, omb.gov has a time ticker that measures the length of the government shutdown.

Other websites merge existing fields into the alert banner, or have other elements on the website that use the same CSS classes or element IDs. As such, parsed banner text may contain additional text, links, or information irrelevant to the government shutdown. Given I also use keyword-search, filtering, and LLM-based partisanship estimates, I don’t think this is likely to have a large effect on results => irrelevant banners will receive neutral designations or not be included in the data.

The banner alerts have been changed during the government shutdown, so this data may not reflect what you see live on the website. Data was pulled between 11/5/2025 and 11/6/2025.

For example, websites that I could not access may not be random; e.g., websites with stricter protections may reflect agencies with more authority that add political shutdown banners at a higher rate than others.

Domains were de-duplicated, but agencies may manage multiple consistently formatted websites. E.g., the USDA (usda.gov) also managed nutrition.gov. Domains do not reflect agencies. The domain source list also includes many domains which may no longer be hosted.

Alerts were also filtered through the presence of set keywords generated by an LLM and manual edits after reviewing the text.² Because this is not a formal analysis, and the shutdown alerts were largely consistent in their messaging, a regular expression approach was sufficient.

On partisanship, outside of keyword presence indicators, I also tested distilbert from HuggingFace and Llama 2 (7b) running locally.

Messages were not normalized for their word length; the longest messages often meant parsing errors that pulled additional text.