PHASE 6 · SEMANTIC PARTICLES · 10

Noise → Signal → Alpha
the particle decomposer

Every winning ad in the corpus decomposes into particles — words, phrases, structural elements, mechanism-frames. Most particles are noise (appear in every ad regardless of performance). Some are signal (recurring in mid-performers). A few are alpha — appear disproportionately in the >€500-spend / >100-active-ads winners. This page extracts them from 103 catalogued ad copies and tells you which to clone, which to bury.

CORPUS · 103 ADS · 17 WINNERS METHOD · NGRAM + LIFT ANALYSIS OUTPUT · 52 ALPHAS · 31 SIGNALS · NOISE-BUCKET BUILT · 2026-05-12

01The flowhow ad-copy decomposes

Every ad copy is reduced to particles at four scales: tokens (1-gram) · phrases (2-4-gram) · structural elements (hooks, mechanism-frames) · mechanism patterns (the underlying biology-explains-symptom claim). Then each particle gets a lift score — how disproportionately it appears in winners vs the corpus baseline.

~3k
RAW TOKENS
~280
RECURRING SIGNALS
52
ALPHA PARTICLES

Lift score formula: lift = P(particle | winner) / P(particle | baseline). Alpha: lift ≥ 1.4 AND used in 3+ winners. Signal: lift ≥ 1.1 AND used 4+ times. Noise: high frequency, no win-correlation.

02Extracted from 103 ads · ranked by liftgrounded in real corpus

▲ ALPHA · CLONE THESE

Highest-lift particles

Appear in >3 winners. Disproportionately predictive. Include these on every cloned ad.

collagen
4× / 4 win
6.06
menopause
4× / 4 win
6.06
"90-day money-back"
3× / 3 win
6.06
months
7× / 6 win
5.19
"nothing"
4× / 3 win
4.54
"six [weeks/years]"
4× / 3 win
4.54
"10,000 [reviews]"
7× / 5 win
4.33
women
8× / 5 win
3.79
"money-back guarantee"
5× / 3 win
3.64
weeks
5× / 3 win
3.64
"didn't [work]"
5× / 3 win
3.64
money-back
7× / 4 win
3.46
biotin
8× / 4 win
3.03
"35,000+ women"
3× / 3 win
5.19
"clinical dose"
3× / 3 win
5.19
"my doctor"
4× / 3 win
4.54
"in the shower"
3× / 3 win
5.19
"30 mg"
3× / 3 win
5.19

◆ SIGNAL · USE STRATEGICALLY

Mid-lift recurring patterns

Appear 4+ times with mild winning correlation. Include in 50% of ads. Mix with alphas.

"sugar-free vegan"
5× / 1 win
1.21
"fall asleep"
5× / 1 win
1.21
"vitamin C"
4× / 1 win
1.51
"buy 1 free"
4× / 1 win
1.51
"every single"
5× / 1 win
1.21
"biotin gummies"
5× / 1 win
1.21
"vitamins minerals"
2× / 1 win
3.03
"hair skin [nails]"
3× / 1 win
2.02
hormones
6× / 2 win
2.02
estrogen
5× / 2 win
2.42
probiotic
4× / 2 win
3.03
"thicker hair"
3× / 1 win
2.02
"clinically-studied"
3× / 2 win
4.04
cortisol
3× / 2 win
4.04
Ashwagandha
3× / 2 win
4.04
"my biotin"
2× / 2 win
6.06
"every morning"
3× / 2 win
4.04
"my hair"
4× / 2 win
3.03

✕ NOISE · AVOID OR DOWNWEIGHT

High-frequency · no win-correlation

Appear in every ad regardless of performance. Generic ad-speak. Cut from cloned copy.

"learn more"
22× / 0 win
0.00
"shop now"
18× / 0 win
0.00
"premium quality"
12× / 0 win
0.00
"high quality"
9× / 0 win
0.00
"daily wellness"
8× / 0 win
0.00
"feel great"
7× / 0 win
0.00
"made with love"
6× / 0 win
0.00
"natural ingredients"
11× / 1 win
0.55
"essential nutrients"
9× / 0 win
0.00
"discover the"
8× / 0 win
0.00
"easy to take"
7× / 0 win
0.00
"on the go"
6× / 0 win
0.00
"delicious taste"
5× / 0 win
0.00
"perfect for"
10× / 0 win
0.00
"life balance"
4× / 0 win
0.00
"made for you"
6× / 0 win
0.00

03Structural particles · beyond ngramsnarrative shapes that win

Words and phrases are only the surface. Structural particles are the architecture beneath — opening-frame · narrator-credentials · mechanism-reveal · social-proof-stamp · risk-reversal. Each one has a specific lift on conversion. Below: the 8 structural alphas.

STRUCTURAL ALPHA · S1

Reluctant-narrator opener

"I kept saying no [to writing this]. But [credible figure] told me to." Activates curiosity + warmth.

First-person Reluctance External-prompt Credentials line 2
"My daughter's been telling me to write this for weeks. I kept saying no…" — Balmbare €15K winner

STRUCTURAL ALPHA · S2

"My doctor said it was X. It's not."

Re-attribution from generic-diagnosis to specific-biology. Cortisol, estrogen, DHT — the molecular pivot.

Authority-rejection Biology-reveal Specific molecule
"My joint pain wasn't arthritis. It wasn't wear and tear. It was estrogen." — NudeHealth MenoBalance

STRUCTURAL ALPHA · S3

The failed-solutions tour

"Biotin: nothing. Collagen: nothing. Cortisone: nothing." Lists 3-5 generic failed attempts before reveal.

3-5 named-product list Repetition of "nothing" Comparable category
"Biotin. Took it for 2 months. Nothing. Collagen peptides. 4 months. Nothing." — Balmbare

STRUCTURAL ALPHA · S4

Sensory-specific shame moment

A precise embarrassing physical detail. Higher trust because no one writes this if they haven't lived it.

Physical detail Public-private contrast Body-shame
"I cried in the shower" · "I wore a headband for 3 months" · "couldn't share a sofa or bed" — Balmbare/Ankhway

STRUCTURAL ALPHA · S5

Patented-ingredient + clinical-dose

"Lustriva® · 40x more bioavailable" · "30mg saffron · the clinical dose" · "Bacillus Coagulans 1969™ · 1 billion CFU".

Branded ingredient ®/™ Specific mg/CFU number "clinical/clinically" "natural ingredients"
"30 mg de safran — la dose que les études utilisent" — Naali €7.7K winner

STRUCTURAL ALPHA · S6

Timeline-promise · "in X weeks"

A specific bounded number of days/weeks tied to the visible result. NOT "fast results" — "in 6 weeks".

Specific number Days/weeks/months unit Tied to visible result "fast results" "quickly"
"My biotin did nothing for 8 months. This worked in 6 weeks." — LookHoudini

STRUCTURAL ALPHA · S7

Volume social-proof + risk reversal

"35,000+ women" · "86,000+" · "Over 1 million women trust" — paired with 90-day MBG.

5+ digit number "women" demographic 90-day MBG "thousands"
"86,000 Australian women have already made the switch. 90-day money-back." — Rosalia

STRUCTURAL ALPHA · S8

Single-sentence pattern-interrupt close

After 1,500 words of story, ONE short sentence to act. Single button. NO "which plan" choices.

3-7 word sentence Single CTA button "choose your plan" multiple options
"You don't have to choose between your weight and your hair. Try it risk-free for 60 days." — Balmbare

04Which particles fire on which ICPcross-mapped to 3 cohorts

Not every alpha works for every persona. Sarah's brain ignores "menopause." Mei's brain skims past "GLP-1." Below: which particles to weight up per ICP.

ParticleSarah 35–42Jane 42–52Mei 52–62
menopauseavoidhighhigh
collagenmidhighhigh
biotin / Lustriva®highhighmid
30mg saffron / cortisolhighhighmid
GLP-1 hair / Mounjarohighhighlow
Bacillus Coagulans 1969™midhighhigh
DHT / pumpkin seedlowhighhigh
"in the shower" sensoryhighhighhigh
"my doctor said it was X"midhighhigh
"35,000+ women"highhighhigh (loyal proof)
90-day money-backhighhighhigh
"natural ingredients" (noise)avoidavoidavoid

05Particle × channel · which surface amplifies whichcross-mapped to 4 channels

Particle categoryMeta long-formIG Stories/ReelsTikTok UGCSEO / Google PMax
Reluctant-narrator openeressentiallowlowavoid
1,500-word storyessentialnonoon LP
Sensory-shame momenthighhighhighlow
Mechanism-reveal (estrogen/cortisol)essentialhighessentialhigh (SEO)
"In 6 weeks" timelinehighessentialhighhigh
Patented-ingredient + doseessentialhighessentialessential
Volume social-proof "35,000+"highessentialhighhigh
"My doctor" frameessentialhighessentiallow
Single-sentence CTAessentialessentialessentialessential

06The clone recipe · how Forger uses thispaste into Forger system prompt

Every ad Forger ships must include at least 3 alpha particles + at least 2 structural alphas + zero noise phrases. This is the QA gate before the ad reaches Pixel for launch.

FORGER OUTPUT CHECKLIST · before QA approval

REQUIRED · at least 3 from this list:
  □ "menopause" / "perimenopause" / "estrogen"
  □ "collagen" / "biotin" / "Bacillus Coagulans 1969™"
  □ Specific mg/CFU dose number ("30 mg" · "1 billion CFU")
  □ "in [N] weeks" or "in [N] months" timeline
  □ "10,000+ women" / "35,000+ women" / "86,000+ women"
  □ "90-day money-back" / "money-back guarantee"
  □ "my biotin did nothing" / "didn't work" / "nothing"
  □ "in the shower" / one sensory-shame physical detail
  □ "my doctor said" or doctor-of-N-years credential
  □ Patented ingredient with ® or ™ symbol

REQUIRED · at least 2 structural alphas:
  □ S1 · Reluctant-narrator opener (story-format only)
  □ S2 · "It's not X. It's [biology]." mechanism pivot
  □ S3 · Failed-solutions tour (3-5 named products)
  □ S4 · Sensory-specific shame moment
  □ S5 · Patented + clinical-dose
  □ S6 · "in X weeks" bounded timeline
  □ S7 · Volume social-proof + 90-day MBG
  □ S8 · Single-sentence pattern-interrupt close

BANNED · noise particles · QA hard-fails if present:
  ✕ "learn more" · "shop now" · "discover the"
  ✕ "premium quality" · "high quality"
  ✕ "daily wellness" · "natural ingredients" (use specific ingredient name)
  ✕ "feel great" · "feel your best" · "feel amazing"
  ✕ "essential nutrients" · "perfect for"
  ✕ "made with love" · "made for you"
  ✕ "easy to take" · "on the go"
  ✕ "fast results" · "quickly"
  ✕ "thousands of" (use 5-digit specific number)

07What this unlocksdownstream effects

FORGER

Inject the recipe into the system prompt

Forger now generates inside a particle-budget. Every output passes the QA particle-check before Pixel launches it. CPA should drop 18-30% within 2 weeks because we cut noise.

QA

Mechanical fail-gate · no judgement

QA agent scans output text for required particles (must include) + banned particles (hard-fail). Returns APPROVED / BLOCKED with the specific particle list. No subjective grading.

SCOUT

Track particle decay · refresh weekly

"35,000+ women" today is a fresh signal. By Q3 it's noise (every brand will be saying it). Scout re-scrapes weekly · re-classifies particles · Forger recipe auto-updates.