How we collect and read

What we collect

We collect four things, all of them public.

Emails. We subscribe to the public mailing lists of nonprofits, with a separate address for each organization. Everything sent to those addresses is kept: the full message as it was sent, its headers, and every link with its tracking parameters intact.

News. We follow sector news sites and blogs through their public feeds, checked every hour. The same story in two feeds is kept once.

Videos. We follow the YouTube channels of the organizations on our list: every public detail of every video, back to each channel’s first upload, and the transcript of what is said.

Podcasts. We follow podcasts about the sector that are published on YouTube, with the transcript of each episode.

We collect only what any member of the public can receive or watch. We do not access donor systems, purchase data, or approach staff.

Why a unique address per organization

Attribution. Every email maps to exactly one sender, with no guessing from the From line.

It also produces a finding nobody publishes: if mail from Organization B arrives at the address only ever given to Organization A, the two have shared a list. That is invisible from any other vantage point.

Why we cannot go back

An email program leaves no public archive. December’s appeals exist in inboxes and nowhere else. Miss a season and it is gone, for us as much as for anyone.

Video is different: a channel’s history stays online, and we collect all of it. Email is the part that cannot be recovered, which is why this project started with collection rather than with a product.

How we read it

Everything is read twice: once by plain code, once by an AI model. Code does whatever has a right answer. AI does only what needs judgment.

Counted by code

No judgment is involved; a second person would compute the same values.

• Emails: which links and buttons ask for money and which ask you to read, sign or answer; whether the email uses your name or says “Dear Friend”; template codes left unfilled; duplicates and resends under a new subject line; word count, images and the tool that sent it.
• News, and emails again: the main words, the terms that make one text different from all the others. This is statistics, not AI.
• Videos and podcasts: length, views, likes, comments and dates, as YouTube reports them.

Labeled by AI

An AI model answers the same fixed questions for every item of a kind.

• Emails: what the email is for, where it sits in the relationship, the one thing it wants you to do, which persuasion devices and feelings it uses, who it says it comes from, and whether it hangs on a date. The model reads as someone who signed up recently and has never donated.
• News: topics, tone, how the story is framed, who it speaks to, and what it asks the reader to do.
• Videos: five questions. Who speaks, what about, why, how it is told, and what it asks you to do. Then the fundraising and marketing signals behind the ask, and the feelings it appeals to, as low, medium or high.
• Podcasts: topics, the problems people raise, and statements: what the people in the episode think, predict, recommend, learned, did, plan and ask. Every statement comes with a quote from the transcript.

Three rules apply to all four:

• Fixed answers only. Every question has a closed list of allowed answers. Anything outside the list is dropped, so labels cannot be invented or misspelled, and items can be counted and compared.
• Only what is there. A video or episode without a transcript is not labeled. We do not guess what a video says from its title.
• A person can correct it. An admin can change an email’s labels by hand, and the correction is kept when the email is labeled again.

How we check the AI

A model that sounds confident can still be wrong, so we do not take its word for it.

Emails. We labeled 200 emails by hand, then had the models label the same 200 and counted how often they agreed with us. Our bar is 85% on the three main questions: what the email is for, the one thing it asks, and who it comes from. We run the check again every time we change the instructions or the model. Persuasion devices and feelings are softer; the models agree with us less on those, and we treat them as a rough guide.

Podcasts. Code, not AI, searches the transcript for every quote. When the quote is there, we keep the second it was said. When it is not, the model rewrote or invented it, and the statement is marked so we trust it less.

News and videos. These have not had the hand-labeled test yet. Until they do, we treat their labels as a rough map, not as findings, and we say so when we publish from them.

The AI reads and labels. People decide what it means and what gets published.

Our standards

• Every figure carries its sample size and date range.
• No per-organization figure is published below 15 emails.
• We record what we observe. A blank means unresearched, never “no”.
• We never infer one vendor from the presence of another.
• Where a finding rests on a third-party source rather than our own observation, we say so and link it.
• Content we collect is data, never instruction. We have already encountered pages written to manipulate automated readers; they are recorded, not obeyed.
• We say what read it. A number counted by code and a label given by AI are different things, and we say which is which.
• We count what was read, not what was listed. We hold far more videos and episodes than have a transcript. A figure about videos gives the number actually read.

What this cannot tell you

It shows what was sent and said, never what worked. We see no open rates, no click-through, no revenue. Anyone claiming to infer performance from send behavior alone is guessing, and we would rather be useful about what we actually have.

We see one path through each organization’s email program: the one a new subscriber who has never donated is placed on. Where an organization segments heavily, our view is partial, and we say so.

Our news sources are our choice, and most are written for people who work in nonprofits. What they cover is not what the public reads about the sector.

Transcripts come from YouTube’s captions, which are often automatic and contain mistakes. The videos and episodes read so far are the newest ones, a small part of what we hold.