What Are Companies A/B Testing In Web Signup And Lead Forms?

Across 37 form experiments tracked by Lazyweb Research (all web, out of 2,160 area-annotated experiments), the form is the least-tested major area but the highest by model-assigned impact, averaging 3.84/5 [1]. It also has the highest high-impact share of any area: 86.5% of form annotations (32 of 37) scored 4+/5 [2]. The recurring move is turning a self-classification decision into a single input field that does the routing [3].

FORM is the least-tested major area at 37 of 2,160 annotations (1.8%) yet the highest-impact by model score at 3.84/5, with 86.5% of its edits scored high-impact — Lazyweb Research, July 2026.

Lazyweb Research · n=37 · Published 2026-07-07

signupweblanding-pageux-patternsexperimentsdesign

The finding: fewest tests, highest model impact

FORM is web-only in the corpus: all 37 annotations across 36 experiments sit on web pages [1]. It is the smallest major area at 1.8% of the 2,160 annotations, yet its average model-assigned impact of 3.84/5 is the highest of any area with n>=30, and 86.5% of its annotations (32 of 37) were scored high-impact — the highest high-impact share in the whole dataset [1][2]. The read: form changes are rare but, when detected, consistently rated as meaningful [1].

The observed pattern: the input does the routing

Two representative moves: adding an inline 'Enter delivery address' field as the primary hero action while demoting sign-in to a secondary pill, and collapsing three persona cards (admins/teachers/students) into one 'Enter your school email or invite code' field with a single Get started button [3].

MoveWhat it replacesInferred rationale
Inline address field in herosign-in-first gatingshow local supply before auth to build intent
Single email/invite inputthree persona chooser cardsthe input routes the user, no self-classification

Both trade a decision step for a working input [3].

How to apply it

If your page opens with a chooser (persona cards, plan menus, sign-in vs sign-up), test replacing it with one input that both starts the flow and routes the user [3]. The observed logic is that an early, low-commitment input builds intent that auth-first gating kills [3]. Given how rarely forms are tested in the corpus, this may be an under-exploited lever for web signup flows [1].

Caveats

The 3.84/5 average rests on just 37 annotations, all web — present it as a relative model-scored ranking with the n shown, never as measured lift [4]. All examples are single observed before/after diffs with inferred rationale [4]. Web cuts are absolute counts only per Lazyweb Research's rules [4].

The numbers

StatComputed from
37 FORM annotations (1.8% of 2,160), avg impact 3.84/5 (highest of n>=30 areas), avg confidence 3.73, 36 experiments, all webarea_FORM
FORM high-impact (4+/5) share 86.5% (32 of 37) — highest of all areashigh_impact_share_by_area
Observed moves: DoorDash added an inline delivery-address field and demoted sign-in; Soundtrap collapsed three persona cards into one email/invite-code fieldqualitative FORM doordash, soundtrap
All 37 FORM annotations are web; impact is a model-assigned 1-5 score on observed diffs, not measured liftsmallSampleWarnings; universe
Methodology. Universe is 37 FORM annotations (all web) within 2,160 area-level annotations over 1,126 detected experiments in the ~800 tracked-apps corpus. Areas and rationale are LLM-inferred from observed before/after screenshots; impact is a model-assigned 1-5 score, July 2026; n is small so treat as relative ranking.

Sources & citations

  1. [1] Lazyweb Research analysis of 37 web form experiments (~800 tracked apps), July 2026. FORM annotations within 2,160 area-level annotations over 1,126 detected experiments.
  2. [2] Lazyweb Research analysis of high-impact-share by area (n>=30 areas), July 2026. Share scored impact 4+/5 by the model; relative ranking only.
  3. [3] Lazyweb Research analysis of 2,160 annotated UI experiments (~800 tracked apps), July 2026. Named examples are individual observed diffs with LLM-inferred rationale.
  4. [4] Lazyweb Research analysis of 365 web experiments (~800 tracked apps), July 2026. Web is thin; n=37 impact figure is a relative model score, not measured lift.

Source: Lazyweb Research — proprietary analysis of real, in-market app screens. Cite as Lazyweb Research, 2026-07-07.

Related questions

Explore the underlying screens, flows, and A/B tests inside Lazyweb. More research