The seven activity types Adobe Target ships with — what they actually do, when each is the right pick, and what they cost you in traffic or setup.
Winner-declaring test of two or more variants on the same page or flow.
When to reach for it: You have a clear hypothesis and want a single causal answer — does the change move the metric, yes or no.
What it costs: Standard two-proportion sample-size math; works on most traffic. The default starting point for any optimization program.
An A/B test where Adobe Sensei shifts traffic toward leaders during the run.
When to reach for it: Same setup as an A/B test, but you'd rather pay forward to a leader than hold a strict 50/50 split — useful when you have confidence in the candidates and want compounding lift while the test is live.
What it costs: Significance interpretation gets less clean because traffic is unequal by design. Be explicit with stakeholders about how to read the result before you start.
Tests every combination of multiple element changes at once.
When to reach for it: You suspect specific elements interact — a CTA copy change might only work with the new hero image — and you want to find the best combination rather than test each change in isolation.
What it costs: Sample size scales with the number of combinations. Two elements with three variants each means nine cells, so traffic requirements multiply quickly. Only worth it on high-volume pages.
Delivers different content to different audiences without a statistical test.
When to reach for it: Your goal is personalisation, not picking a winner — e.g., logged-in vs anonymous, returning vs new, geo-specific promos.
What it costs: Minimal. No significance math, no winner declared. Trade-off: you never learn whether the targeted variant actually outperformed the default for that audience.
Sensei picks the best variant per visitor based on profile and contextual signals.
When to reach for it: You have pre-built variants and want the platform to serve each visitor whichever one is best for them, rather than treating everyone the same.
What it costs: Training period before the model is reliable; harder to explain post-hoc since there's no single winning variant. You learn "personalised lift," not "the new hero beat the old one."
Always-on ML that matches offer variations to each visitor at scale.
When to reach for it: You have at least 3 viable variants, your audience is heterogeneous enough that no single experience will win for everyone, and you have rich profile data to feed the model.
What it costs: Same as Auto-Target with more emphasis on always-on, cross-activity scale. Needs enough traffic per profile slice for the model to stabilise.
Surfaces products or content via collaborative-filtering or content-based algorithms.
When to reach for it: Catalog-driven sites where the right item to surface varies by visitor and context — retail PDPs, content discovery, related articles.
What it costs: Needs a product catalog feed plus behavioural data to train the recommendation model. Different evaluation model than A/B — you're comparing recommendation strategies, not single variants.