A new landing page's best prior is brand mean, not design similarity

Summary: When a new landing page launches into a multi-variant race, what should its conversion prior be? Across 83 trafficked pages (20,341 impressions, 343 purchase-or-lead conversions within 24 hours), any reasonable informed prior is worth roughly 80–160 pseudo-observations against an uninformative Beta(1,1). Brand-mean pooling wins. Text-embedding nearest neighbours match brand mean only because same-brand pages share vocabulary — once same-brand neighbours are removed, embeddings add nothing over the global mean.

New pages are the hard case in multi-variant landing page systems. They enter with almost no traffic, compete with pages that already have hundreds of impressions, and often sit behind an exploration capacity limit. A common default is an uninformative prior such as Beta(1,1) — a 50% conversion rate mean that is almost never true for paid commerce traffic.

We measured how much information different prior sources actually carry.

Method

For each of 83 trafficked pages, we predicted that page's conversion rate from the other pages only (leave-one-page-out), using several prior sources, then fit — per source — the Beta prior strength τ (pseudo-observations) that maximises held-out Beta-Binomial marginal likelihood across pages.

Sources compared:

• Global pooled mean
• Brand pooled mean
• Style-family pooled mean
• kNN on page-copy embeddings (MiniLM, BGE, MPNet)
• Ablations: cross-brand kNN only; kNN restricted to earlier-launched pages

All pooled predictors used a uniform 10 pseudo-observation stabilizer at the global mean; kNN used cosine top-10 with traffic-weighted pooled counts. Comparisons refit τ inside each of 2,000 page-level bootstrap resamples.

This is a retrospective information-value measurement, not a full deployment simulation. Open data: cold-start-prior-loo-2026-08.json. Protocol language: Lutiq Holdout Protocol.

Results (24h endpoint)

Prior source	τ* (pseudo-obs)	ΔLL vs Beta(1,1)
Global pooled mean	80	+185.8
Brand pooled mean	160	+199.6
Style-family pooled mean	80	+180.1
kNN MiniLM (all pages)	160	+199.4
kNN BGE (all pages)	160	+199.9
kNN MPNet (all pages)	160	+198.4
kNN MiniLM, cross-brand only	80	+179.6
kNN MiniLM, earlier-pages only	80	+185.8

Bootstrap contrasts:

• kNN MiniLM vs global: +13.9 log-likelihood, p(no improvement) = 0.003
• kNN MiniLM vs brand mean: ~0, p = 0.517 — embeddings do not beat brand mean
• Brand mean vs global: +13.9, p = 0.002
• Cross-brand kNN vs global: −5.5, p = 0.823 — mildly worse than global

72-hour endpoint: same ordering, same conclusions.

Interpretation

1. Informed priors are strongly validated. 80–160 pseudo-observations is more traffic than many new pages see in their first day of exploration. Shipping Beta(1,1) for paid landing pages is a modelling own-goal.

1. The recoverable signal in this sample is brand identity, not design similarity via text. Embeddings look clever until you remove same-brand neighbours; then they collapse to (or below) the global mean.

1. Style-family pooling underperformed the global mean in this observational set — family labels here are confounded with brand mix. Do not ship attribute priors from observational pooling without a cleaner design.

1. This does not prove design never transfers across brands. It proves that text embeddings over a small brand set and a few hundred conversions cannot see that transfer. Structured attributes or visual embeddings remain open research.

What to do with this

• Prefer a brand-mean cold-start prior (or hierarchical brand/global partial pooling) over uninformative Beta priors.
• Do not market “semantic similar pages” as a cold-start miracle without a cross-brand ablation.
• For system claims, still measure against a holdout, not only relative winners inside the race.

Caveats

• Multi-brand commerce sample; not a public consumer dataset.
• Retrospective LOO, not online A/B of prior policies.
• Conversion endpoint is purchase-or-lead within 24h (72h sensitivity agrees).