Lutiq Research
First-party measurements and worked arithmetic — not listicles. Numbers we measured, with the caveats attached.
Start here: Holdout Protocol, what A/B tests cannot see, DTC landing-page motion audit, sample size, generation diversity collapse, ad features need variation, and cold-start priors.
- Ad creative embeddings cannot help when every click shares one creative setA paired ablation on 8,609 paid impressions: when ad creative does not vary, rich creative embeddings add no allocation signal beyond noise — a structural limit, not a model-size problem.
- A new landing page's best prior is brand mean, not design similarityLeave-one-page-out measurement across 83 trafficked pages and 343 conversions: informed Beta priors are worth 80–160 pseudo-observations, but text embeddings do not beat a simple brand-mean prior.
- Lutiq Holdout Protocol for paid landing pages (v0.1)A citable standard for measuring multi-variant and predictive landing-page systems against a live holdout — required fields, selection vs contextual gain, and claims we will not make.
- 7,768 impressions, 52 conversions: why landing-page allocator backtests go quietA public-safe account of evaluating contextual multi-page allocation under extreme conversion rarity — corrected walk-forward methods, and why impressive lift percentages often fail uncertainty checks.
- Our page generator produced 15 identical pages and scored them 82-95 on distinctiveness.Fifteen consecutive production runs returned the same 18 feature assignments. Every page scored highly on a per-page distinctiveness metric. A diagnosis of why per-page quality scores cannot detect cross-page convergence.
- We rebuilt 230 CRO benchmarks against the research. The median real effect is +1–2%.Our first set of landing page priors came from vendor case studies: video hero +30%, PAS headline +20%, sticky CTA +20%. Recalibrating against experimental medians cut nearly every number by 3-10x. Full table and method.
- We instrumented the 25 best DTC landing pages. Only 2 were playing video.A runtime audit of 25 leading DTC landing pages. We measured what actually ran in the browser at five scroll depths — not what was in the code. 15 sites shipped autoplay video; 2 were playing it.
- How much traffic does a landing page test actually need?The arithmetic of landing page test sample size, worked end to end: why 300 impressions per variant, why you need 5–10 conversions before deciding, and how many variants a given ad spend can support.
- Two ways people buy from an ad: 7 minutes or 3 daysAn analysis of 51 converter journeys found two distinct purchase systems, not one funnel. One converts in about 7 minutes with no quiz. The other takes multiple sessions and reaches 2+ minutes on page at 4.5x the rate.
- Two landing pages tied at 17.5%. The right routing rule got 27.5%.A worked example of contextual uplift — the conversion gain that is mathematically invisible to an A/B test because it cancels out in the aggregate. Includes a real holdout result across 19,708 users.