LinkedIn Ads Experiment: 42 Split Tests That Changed Our CPQL Model
A structured experiment log of 42 LinkedIn Ads split tests across creative, audience, and offer variables — with lifts, losses, and no cherry-picking.
Executive summary
Problem — Public LinkedIn Ads guidance is dominated by winner-only anecdotes. Operators have no honest ledger of what actually beats control.
Why it matters — LinkedIn is now the most expensive B2B channel — decisions must be data-backed, not vibes-based.
- Paid media leads
- Demand gen managers
- Growth consultants
- Only 38% of tests beat control at p < 0.05. Half of what teams ship is noise.
- Creative accounts for 62% of the total observed CPQL improvement — audience tweaks are ~2× less impactful than teams assume.
- Video ads outperformed static in 71% of paired tests — but only when ≤15 seconds.
38% of tests beat control at p<0.05.
62% of aggregate CPQL improvement came from creative variables.
Video ≤15s beat static in 71% of paired tests.
Native document ads had a median 34% lower CPQL than single-image.
Skills targeting beat job-title targeting in 8/12 head-to-heads.
CPQL degraded 22% once frequency exceeded 8 impressions per user in 14 days.
Landing page changes contributed 21% of aggregate CVR improvement — not the ad.
Research objectives
- Publish an honest, pre-registered LinkedIn Ads experiment log.
- Attribute lift by variable category (creative / audience / offer / landing).
- What variable class contributes the most CPQL improvement?
- How often do LinkedIn Ads changes actually beat control?
- Where are the frequency and creative-fatigue ceilings?
Methodology
- LinkedIn Campaign Manager conversion data
- Downstream CRM opportunity data (HubSpot)
- Minimum $8k spend per variant
- Minimum 100 conversions per variant before significance testing
- Two-tailed z-test for proportions
- Bonferroni correction for multi-variant tests
- External review by 2 independent paid media consultants
- Learning-curve bias — later tests benefit from earlier ones.
Data & visualizations
Tables
| # | Variant class | Change | Median CPQL lift | p-value |
|---|---|---|---|---|
| 1 | Creative | Video ≤15s vs static | -38% | 0.002 |
| 2 | Format | Document ad vs single image | -34% | 0.004 |
| 3 | Creative | Face in first frame | -28% | 0.01 |
| 4 | Landing | Above-fold value prop rewrite | -24% | 0.02 |
| 5 | Creative | Bold data point in headline | -22% | 0.03 |
| 6 | Audience | Skills > Job title | -18% | 0.04 |
| 7 | Offer | 'Free audit' vs 'Book demo' | -16% | 0.03 |
| 8 | Creative | Carousel vs single image | -14% | 0.04 |
| 9 | Landing | Removed nav on LP | -11% | 0.04 |
| 10 | Audience | Company size 200-1000 | -9% | 0.04 |
Analysis & insights
- 01Most LinkedIn teams over-invest in audience tinkering and under-invest in creative velocity.
- 02The 'won' bar is lower than everyone thinks — expect ~4 in 10 tests to move the needle.
- 03Landing pages are the second-largest lever and the least tested.
Recommendations
The 62% creative contribution is the clearest signal in the dataset.
21% of aggregate improvement is currently untouched at most companies.
Frequency-driven decay is real — the 22% CPQL degradation is not recoverable.
- Sample is B2B SaaS only. Consumer verticals will differ.
- Learning-curve bias may inflate later-test win rates.
Download the full research package
FAQ
Are the raw experiment logs shared?
Yes — the CSV includes all 42 experiments, including losers and inconclusives.
Why publish losers?
Because a 38% win rate is the honest number. Winners-only libraries lie to operators.
Get the monthly research digest
One email per month with the latest SaaS marketing research, benchmarks, and experiments. No spam.
Join the newsletter