PAX2

5852–64

050 = reference average100

Engage

6660–72

Interest

5245–59

Act

5850–66

Sample figures

	Engage · 25%	Interest · 40%	Act · 35%
Counts	33,600	0:17	425
Rate	30.0%	18.0%	1.4%
Ref. avg	24.0%	17.5%	1.1%
Weighted	16.5	20.8	20.3
Distance	+1.0 sd	+0.1 sd	+0.5 sd
Score	66	52	58
Weight	25%	40%	35%

Sample figures

Spend$900

Impressions112,000

Reach74,300

3s plays33,600

ThruPlays13,400

25% play17,900

50% play10,800

75% play7,000

100% play4,600

Avg. Time0:17

Shares270

Saves95

Comments60

Reactions980

Clicks1,570

CTA1.4%

Spend

$900

Impressions (1k)

$8.04

3s play

$0.027

Per Person

$0.012

Sample figures

Opening
Direct and urgent.
Tight framing on the speaker, outdoors, in daylight.
A map graphic shows districts being redrawn.
Bold captions repeat the line about losing the country.
The speaker delivers the ask; a donate link appears on screen.

setting: Outdoors, in daylight.
on screen people: One speaker.
captions: Yes, burned in throughout.
text overlays: Yes, on the key lines.
music: Low, rising bed under the voice.
pace: Fast, with quick cuts.
call to action: Get involved and donate.
visual variety: Talking head with one map graphic.

Sample figures

+1.1 ptsVote choice, average tested ad

78%Ads that moved voters their way

Politician as messenger

+0.06 (−0.35 to +0.46)

Advocates an opinion

−0.08 (−0.43 to +0.27)

Negative tone

−0.12 (−0.38 to +0.15)

Fear

−0.01 (−0.59 to +0.56)

Text on screen

−0.69 (−1.40 to +0.01)

Sounds scripted

−0.08 (−0.41 to +0.26)

574 ads in 132 randomized studies, 2018–2020. Points of vote choice, ads with the trait against ads without it. Together the traits predict almost none of the gap between one ad and the next; only a test does.

Sample figures

Raise moneyGoal

Goal	Ads	Spend	Median impressions
Persuade	77%	81.8%	6K–7K
Raise money	12.6%	6.6%	0–1K
Both	2.1%	1.1%	1K–2K
Neither	8.3%	10.5%	2K–3K

100,831 pro-Democratic video ads on Meta in 2022, $101 million in spend.

Sample figures

1.9×Clicks, best headline over worst

12,256Headline tests

Starts with “This” or “These”6.8% of winners · 5.8% of losers

Carries a number19.1% of winners · 17.2% of losers

Asks a question14.2% of winners · 18.2% of losers

Says “you” or “your”25.3% of winners · 25.3% of losers

Quotes someone1.2% of winners · 1.2% of losers

Shouts a word in capitals5.1% of winners · 6% of losers

Starts with “What”, “Why” or “How”6.3% of winners · 8.3% of losers

Says “watch”, “see” or “listen”8.2% of winners · 6.7% of losers

Trails off or breaks midline4.6% of winners · 4.9% of losers

First person: “I” or “my”14% of winners · 12% of losers

The same story under different headlines, shown at random, 2013–2015.

Pushed Baseline Killed

Label

r_{v} = \frac{views _{v}}{median _{u \in c (v), age_{u} \geq 7 d} views _{u}} y_{v} = 1 [r_{v} \geq 3]

Within a channel

Δ_{c} = \frac{1}{a _{c}} v \in A_{c} \sum y_{v} - \frac{1}{b _{c}} v \in B_{c} \sum y_{v} a_{c}, b_{c} \geq 3, a_{c} + b_{c} \leq 300

Weight

w_{c} = \frac{2 a _{c} b _{c}}{a _{c} + b _{c}}

Pooled, in points

Δ^= 100 \cdot \frac{\sum _{c} w _{c} Δ _{c}}{\sum _{c} w _{c}}

90% range

Δ^^{* (1)}, \dots, Δ^^{* (1000)} channels drawn with replacement [Δ^_{(50)}^{*}, Δ^_{(950)}^{*}]

Worked example

Δ_{Holler} = \frac{13}{70} - \frac{19}{208} = 0.094 w = \frac{2 \cdot 70 \cdot 208}{278} = 104.7 Δ^_{reform, CTA} = \frac{2073.7}{252.5} = + 8.2

A guess, written down before the numbers come in: will this video do better or worse than usual? When the numbers arrive, PAX2 checks the guess against them, and every check makes the next guess sharper.

Every round on file is already loaded; forecasting reopens when a new round is expected.

Video