Constraints are all you need: Exploring AI-powered landing page design

Landing pages are central to digital marketing, yet producing high-quality pages traditionally demands scarce, cross-disciplinary expertise in design, copywriting, and front-end development. Large language models (LLMs) can now generate web pages from natural-language descriptions, but unconstrained generation is often inconsistent and not deployment-ready. This study investigates whether imposing structural constraints on generative AI improves landing-page quality, comparing three generation methodologies: unconstrained generation using LLMs, prompt-engineered generation, and structured generation with component-based constraints. We evaluate these methods through an expert panel (65 participants) assessment of 12 generated landing pages across multiple dimensions. Statistical analysis reveals that constrained generation significantly outperforms other approaches across all evaluation dimensions, with improvements of 18.9% in visual design, 15.7% in content quality, 13.3% in user experience, and 19.6% in business effectiveness compared to unconstrained generation. The constrained approach yielded 78.60% high satisfaction ratings versus 56.98% for unconstrained generation. To account for the repeated-measures structure of the data, we additionally report rater-level aggregated tests, linear mixed-effects models, and leave-one-company-out and bootstrap sensitivity analyses, all of which confirm the constrained approach’s advantage. We also conducted a parallel assessment using four advanced LLMs, Claude 3.5 Sonnet, GPT-4o, Gemini, and DeepSeek, which demonstrated a statistically significant positive correlation with human evaluations (Pearson r = 0.67, p < 0.01) while revealing model-specific evaluation patterns. Despite exhibiting rating inflation compared to human assessors, LLMs consistently identified the superiority of constrained generation approaches, providing further validation of our central thesis. Our findings demonstrate that imposing appropriate structural constraints on generative AI systems leads to superior outcomes in automated web design, providing valuable insights for developing more effective AI-powered design tools.

Authors

Institutions

Publication Details

Journal
PLoS ONE
Published
2026-11-19
DOI
https://doi.org/10.1371/journal.pone.0362573
Primary Topic
Digital Marketing and Social Media
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Constraints are all you need: Exploring AI-powered landing page design

Miodrag Vujkovic, Maja Drakic-Grgur, Ivan Jovović, Tomo Popović
PLoS ONE
Digital Marketing and Social Media
article

Constraints are all you need: Exploring AI-powered landing page design

Miodrag Vujkovic, Maja Drakic-Grgur, Ivan Jovović, Tomo Popović
article en

Abstract

Landing pages are central to digital marketing, yet producing high-quality pages traditionally demands scarce, cross-disciplinary expertise in design, copywriting, and front-end development. Large language models (LLMs) can now generate web pages from natural-language descriptions, but unconstrained generation is often inconsistent and not deployment-ready. This study investigates whether imposing structural constraints on generative AI improves landing-page quality, comparing three generation methodologies: unconstrained generation using LLMs, prompt-engineered generation, and structured generation with component-based constraints. We evaluate these methods through an expert panel (65 participants) assessment of 12 generated landing pages across multiple dimensions. Statistical analysis reveals that constrained generation significantly outperforms other approaches across all evaluation dimensions, with improvements of 18.9% in visual design, 15.7% in content quality, 13.3% in user experience, and 19.6% in business effectiveness compared to unconstrained generation. The constrained approach yielded 78.60% high satisfaction ratings versus 56.98% for unconstrained generation. To account for the repeated-measures structure of the data, we additionally report rater-level aggregated tests, linear mixed-effects models, and leave-one-company-out and bootstrap sensitivity analyses, all of which confirm the constrained approach’s advantage. We also conducted a parallel assessment using four advanced LLMs, Claude 3.5 Sonnet, GPT-4o, Gemini, and DeepSeek, which demonstrated a statistically significant positive correlation with human evaluations (Pearson r = 0.67, p < 0.01) while revealing model-specific evaluation patterns. Despite exhibiting rating inflation compared to human assessors, LLMs consistently identified the superiority of constrained generation approaches, providing further validation of our central thesis. Our findings demonstrate that imposing appropriate structural constraints on generative AI systems leads to superior outcomes in automated web design, providing valuable insights for developing more effective AI-powered design tools.

PLoS ONEVol. 21(11)
University of Donja Gorica (ME)
Digital Marketing and Social Media
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon