Fundraising emails mentioning “billionaire(s)”, weekly
% of each party's political fundraising emails.
Bold line = 4-week trailing average; faint line = raw weekly value.
Show the weekly numbers
Words most distinctive of emails mentioning “billionaire(s)”
Fundraising emails substantively about AI, weekly
% of each party's political fundraising emails. Counts only emails a
model classified as substantively about AI, not raw keyword matches.
Bold line = 4-week trailing average; faint line = raw weekly value.
Show the weekly numbers
Words most distinctive of emails mentioning AI
Fundraising emails mentioning data centers, weekly
% of each party's political fundraising emails. Counts mentions in
either direction, and both directions occur: of 19 Republican data-center emails in
2026, 12 attack them (electric bills, land use — Hawley, and Virginia candidates)
and 6 promote them as economic development (Tuberville, the Indiana GOP).
Bold line = 4-week trailing average; faint line = raw weekly value.
Show the weekly numbers
Words most distinctive of emails mentioning data centers
Recent emails
Notes on the data
- Denominator. Every rate is a share of political fundraising emails in
this archive — not of campaigns, and not of voters. The Republican side carries a
lot of list-rental operations, so read these as "the fundraising inbox," not "the
parties."
- Incomplete weeks are dropped, not drawn. The archive updates daily, so the
current week is always partial. It is excluded until complete rather than plotted as
a collapse.
- Distinctive words are keyness, not frequency. Weighted log-odds with an
informative Dirichlet prior (Monroe, Colaresi & Quinn 2008), against all other
political fundraising email in the same window. The sample is capped at 6 emails per
sender: counted per email, the "distinctive" words are just the branding of whichever
committee mails most; counted once per sender, prolific senders drag in the whole
generic campaign vocabulary. The category's own defining word is excluded.
- Mentions are counted over the full email body. Worth stating because it
changes a number: an earlier version of the data-center series was computed over a
2,000-character preview of each email and undercounted by roughly 1.5–3.6× (June 2026
Democrats: 0.70% full-body vs 0.40% truncated). The figures here are full-body.
- A mention is not a stance. These lines count who is talking about
something, not whether they are for or against it.
- AI comes from a labeled census. Every AI-mentioning email is classified by
gpt-5-mini; cross-model agreement on the villain task was κ=0.91. Months
whose AI emails have arrived but not yet been labeled are excluded entirely rather
than shown low.
- Everything is recomputed. No number here is typed in by hand; the page is
generated from piece_metrics.json,
content_profiles.json and
recent_examples.json, rebuilt each weekday.
The code that produces them is
public.
How to cite
This page is a living instrument: the figures move every weekday as new
mail arrives. A citation should therefore record both the version you used (the
data-through date) and the date you accessed it — the numbers you quote will not be the
numbers a reader sees later. Both are filled in below.
Please also credit the underlying archive. Every email counted
here comes from the Archive of Political Emails, collected and published by
Derek Willis. This dashboard
is a measurement layer on top of that collection, not a source of it.
Code and data are public. The full pipeline, the derived series,
and the 546,769 LLM labels behind these figures are at
github.com/andybhall/fundraising-inbox
— enough to reproduce every number on this page without re-running any paid labeling.