2026 ad tracker

What this is

Political ads that began running on Meta or Google on or after January 1, 2026, and that the classifier marked anti-trans. It is the live counterpart to the project's finished 2018–2025 political-ad corpus; the 2025 weeks in the chart come from that corpus.

How sponsors are typed

Every advertiser is given a type (super PAC, candidate committee, party committee, 501(c)(4), 501(c)(3), media company and so on) and the party it is tied to. Where the sponsor is a federal committee the type comes from its Federal Election Commission record; otherwise it comes from the sponsor's own name and “paid for by” line, or from public records, and a person has checked the largest sponsors. A page run by a PAC under another name is typed by the PAC that paid. Republican-aligned and Democratic-aligned mean a party committee, a candidate's own committee, or a group that exists to elect one party's candidates (for a super PAC or 501(c)(4) that is judged from public record or from whom its ads back and attack). A conservative issue group with no such tie is shown as no party tie, and where nothing was found the type or party says not determined. A sponsor shown as a super PAC may be a hybrid PAC; a 501(c)(3) or (c)(4) label is public record, not an IRS check made here.

How it is collected

• Nightly from midnight Eastern, on the project's desktop: a delta pull from the Meta Ad Library (the archive's 165 tracked pages and 72 lexicon terms) and from Google's Transparency Center (about 250 listed advertisers, within a daily call budget). Confirming a flagged ad, by a person or by the keyword rule below, adds its advertiser to that pull from the next night on: a Meta ad's page, or a Google ad's advertiser account. A confirmed Meta advertiser also moves to the front of the search for its Google account, and joins the Google pull once that account is found.
• Ad text comes from the platforms, from image OCR and from video transcripts. Each ad says which.
• Qwen 3.8 27B reads every ad once with the archive's prompt. An ad it does not flag, but that uses a term from the tracker's lexicon, or that it judged with low or medium confidence, gets a second, slower read with the model's step-by-step reasoning switched on, and that answer stands. Then the editorial overrides apply. Until September 11, 2026 the tracker used the archive's own chain (Qwen 2.5 7B, then 32B re-reads); those ads are being re-run with the 27B over the following nights, and each verdict records the model that made it.
• The run then publishes here. The status line at the top says when that last happened.

Every flag here comes from that automated classifier; an ad marked ✓ reviewed has been checked and confirmed by a person, or runs the same wording as an ad that was. An advertiser usually runs one script as many separate ads — the largest group here is 72 copies of one text — so a person's decision carries to the copies: at least four fifths of the copy's wording must come from the ad that was read. The keyword rule: a flagged ad that says “transgender” or “sex change” is about trans people beyond doubt, so the nightly run confirms it without a person and marks it ✓ verified.

Reading the numbers

• Meta and Google publish spend as brackets, not amounts. Totals are sums of those brackets; the true figure lies inside the range.
• "Shown in" lists the US states the platform reports delivering the ad in. Delivery outside the US is counted but not listed.
• Google coverage is thinner than Meta's: its Transparency Center has no date filter and is walked within a daily call budget.
• There is no broadcast TV here. The Wesleyan/Kantar data arrives after the cycle and cannot be redistributed ad by ad.