We read $103,524 of this account's own spend history, found three things that overturned our first conclusions, then built 30 statics, 10 videos and 10 motion pairs against what survived. This page is the working, including the parts where we were wrong.
One account, Sep 2025 to Aug 2026. Two offers — a free 75-minute workshop (“Build A $10k/Month Newsletter”) and the Personal IPO Summit, a free two-day event. Two audiences the account itself names: Newsletter ICP and Agency ICP. 28 distinct pieces of body copy across 134 creatives.
Every figure on this page comes from a direct Meta Graph API pull against the ad account on 2026-08-27, not from a dashboard. The pull scripts are in the vault next to this build so any number here can be re-derived.
Our first pass ranked every ad on Meta's lead action and produced a tidy list of
losers. Top of it: DCT_107 — $2,325 spent, zero leads. Next to it
DCT_112, $2,560 spent, zero leads.
Both numbers were artifacts. This account has changed its conversion event three times:
| Era | Campaign | Conversion event |
|---|---|---|
| Nov 2025 | Qualified leads | Custom: “Lead q” + “Lead nq” |
| Mar 2026 | Mar 2026 $10k Newsletter Webinar | complete_registration |
| Jun 2026 → | NLA Webinar Funnel / NLA WINNERS | lead |
Normalized, the ranking inverts. DCT_107 drove
249 registrations at $9.34. DCT_112 drove
281 at $9.11. Both are top-quartile. Had we optimized on the raw pull, we
would have killed two of the account's better performers and briefed away from what worked.
Two more ads (PC 1, PC 2, ~$3,945) had no conversion event
at all — they were video-view campaigns. They are excluded here by objective, never
scored as failures.
Both were caught by adversarial review — three independent models (Gemini 3.1 Pro, GPT-5.6, and a Claude reviewer with repo access) scoring the analysis, plus our own re-testing. They are on this page because the corrections are more instructive than the original claims.
We measured corr(length, cost) = +0.570 across n=27 and nearly wrote a hard 20-second cap into the brief. But n=27 was ad rows, and those rows re-run only 9 distinct video files — one of them 9 times. At the level of distinct assets the correlation falls to +0.371 (n=9), and it isn't monotonic: the two longest videos in the account (55.8s → $12.38, 57.7s → $10.34) beat the shortest (10.2s → $12.24). The account also has no videos at all between 17.8s and 36.6s, so a “20-second” threshold sits in a gap with no data on either side of it.
It survives stratification (within-NLA r=+0.500, within-PIPO r=+0.814) so it is a real signal — but it is a hypothesis, not a finding, and it cannot carry a production rule. So we did the opposite of capping: six videos ship as a deliberate duration ladder — the same script at 12s, 20s and 35s, twice — so the question can actually be settled with spend instead of asserted.
corr(CTR, cost) = −0.027, which we initially read as “CTR has no predictive power.” The 95% confidence interval is roughly −0.30 to +0.25. That is absence of a detected linear association in this sample, not evidence of no relationship. The honest version: CTR is not usable as a proxy here — don't brief for scroll-stop alone. One ad in this account hit 21.4% CTR at $17.18 per registration.
Every number on this page — every one — is a cost per registration. Registrations are not the business. Attendance is a step, and a paid enrollment is the outcome.
For one window in Nov 2025, this account measured lead quality, via two confirmation-page conversions named “Lead q” and “Lead nq”. It is the only quality data the account has ever produced:
| Ad | Spend | Qualified | Unqual. | Qual. rate | Cost / qualified | Cost / any lead |
|---|---|---|---|---|---|---|
| DCT_112_Founders | $2,560 | 113 | 168 | 40.2% | $22.66 | $9.11 |
| DCT_111 | $704 | 18 | 50 | 26.5% | $39.12 | $10.35 |
| DCT_114_Social is HARD | $311 | 7 | 24 | 22.6% | $44.43 | $10.03 |
| DCT_113_Life Hack | $248 | 0 | 18 | 0.0% | — | $13.78 |
| DCT_115_Confession | $29 | 0 | 3 | 0.0% | — | $9.67 |
| Total | $3,853 | 138 | 263 | 34.4% | $27.92 | $9.61 |
The gap between $9.61 and $27.92 is a 2.9× multiplier, and the
qualified rate swings from 0% to 40% depending on which creative brought the person in.
Then the account stopped measuring it. Everything since June 2026 optimizes bare
lead.
What this does not show, and we want to be exact about it: there is
no ranking inversion here. DCT_112 is the cheapest ad on raw cost and the
cheapest on cost-per-qualified. The effective sample is four creatives, one of
which supplies 70% of all leads and 82% of qualified ones, in a single era on a single offer.
Calling it “66% of your leads are noise” would overstate it — 65.6% were classified
unqualified under one historical rule that nobody has validated against revenue.
The ask. The join that would settle this is
registrations → webinar attendance → paid enrollments, split by angle and
era — it lives in GoHighLevel, beehiiv and SamCart, not in Meta. With even a
directional version of it, the exploit arm below can be re-ranked by
cost per buyer instead of cost per lead. Without it, we are all optimizing a
metric nobody has tied to revenue. Restoring the q/nq split costs one landing-page change.
The account's best static and its worst video are 3× apart on cost per registration. They are also visually opposite, which turned out to be more useful than any correlation.
The pattern we took forward: premium and editorial wins; generic-social-template loses. Every static below is built in the blueprint system decoded from that first image — cream ground, navy Anton condensed, ochre CTA, drafting furniture — rendered as HTML through headless Chrome so the typography is exact by construction.
The brief was explicit: don't just iterate on what already works. So the split is structural,
not aspirational. Each unit carries an immutable arm and rationale, frozen in
manifest.json and hashed before the first render
(3ad30a69…), so post-hoc reassignment would be visible in version history.
The split also holds within each format, which is what stops “explore”
quietly becoming statics-only.
| Rationale | Evidence grade | Arm | Static | Video | Flex | Total |
|---|---|---|---|---|---|---|
| R1 — proof density (named person + number in the hook) | Finding | exploit | 10 | 0 | 2 | 12 |
| R2 — duration ladder, identical script at 12/20/35s | Hypothesis | exploit | 0 | 6 | 0 | 6 |
| R4 — blueprint system extended, plus breaks as controls | Finding | exploit | 6 | 0 | 2 | 8 |
| R5 — Agency ICP rebuilt on Newsletter-ICP structure | Hypothesis | exploit | 2 | 0 | 2 | 4 |
| R3 — qualification-forward copy that repels bad fits | Hypothesis | explore | 4 | 1 | 1 | 6 |
| R6 — whitespace angles never run in this account | Untested | explore | 8 | 3 | 3 | 14 |
| Total | 30 | 10 | 10 | 50 |
The explore arm is sourced deliberately, not invented: proof assets that have only ever appeared buried in body copy and never as a hook (Nezha, Julia, Daria, the $2,000-a-call anchor), curriculum material that has never been in an ad at all (sponsor sales, digital products, media kits), and framings absent from all 18 primary bodies. Three units are reserved to contradict our own conclusions: two videos at 35s against the length hypothesis, and two statics that deliberately break the design system.
Meta is already rejecting ads on this account. Rejection notices landed on 7/2 (7 ads), 7/3 (3), 7/30 (4) and 8/3 (3), with multiple failed appeals through July and August 2026. The specific triggering clause was never diagnosed internally.
Two policies are in play. The obvious one is earnings claims. The one we think is actually biting is Personal Attributes — Meta prohibits asserting or implying a viewer's financial status, and the account's live copy does it constantly: “you're stuck under $10K a month”, “IF YOU'RE AN AGENCY OWNER UNDER $10K A MONTH”, “your newsletter isn't making you AT LEAST $10K/month”. The offer headline itself — “Build A $10k/Month Newsletter In 90 Days” — is a second-person income promise with a timeframe.
So we did not treat the existing copy as a safe ceiling. 48 of the 50 units carry no second-person income framing at all: claims are attributed to a named person and stated in the past tense about what that person did. That is a deliberate compliance posture, and it is also why the copy reads differently from what has run before.
Proof-density, the blueprint system, the Agency-ICP reframe, and the duration ladder. Each card carries its rationale and, where relevant, why it exists.
Qualification-forward copy designed to reduce registrations in exchange for better ones, plus angles absent from all 28 historical copy variants.
All rendered from HTML through HyperFrames — deterministic, no video model. Six of them are the duration ladder: identical scripts at 12s, 20s and 35s so the length question becomes measurable. No founder footage exists in any system we can reach, so these are typographic motion pieces rather than the cinematic treatment the account's best video used. That is a real constraint on this batch, not a stylistic preference.
The statics above are HTML rendered by headless Chrome, which makes typography exact. The fair question is whether an image model could do the same job. We re-issued six of the briefs verbatim to two models running locally on an RTX 5090 — SDXL 1.0 and Z-Image Turbo — same brief, same aspect, same seed, same QA rubric.
The honest read is more interesting than “local models are worse.” Both models understood the art direction immediately — cream stock, navy and ochre, blueprint furniture — and a 2026 local model (Z-Image) renders a headline correctly where a 2023 one (SDXL) cannot. The failure is specific and it is the same failure in both: everything below the headline. Secondary text, spec tables, CTA labels and exact typeface are all out of reach.
And on this account the failure is not just aesthetic. Asked for “$70,000 Per Month”, SDXL rendered “$767009”. On an ad account already being rejected for earnings claims, a model that invents a number inside a financial claim is a compliance liability, not a quality problem. That is the argument for HTML rendering on text-heavy creative — not that it looks better, but that it cannot say something you didn't write.
Where local models do earn their place on this build: art direction exploration, background plates, and any surface with little or no literal text. At ~4 seconds and no marginal cost, they are excellent for deciding what a thing should look like — just not for producing the final artwork that a compliance reviewer reads.
ANALYZE), so none of this has been published and none of it has outcome data.
This is a portfolio and a test design, not an experiment — and it is labelled that way
deliberately.