PaperlayerView exact revision on arXiv

Source-linked AI summary

LandingAgent: A Reference-Annotated Dataset and Agentic Generation Framework for Landing Pages

Injun Baek, HyeongSeok Lee, Yearim Kim, Junhoo Lee, Nojun Kwak

arXiv:2608.27902v1cs.CLcs.AI

TL;DR

Landing-page generation needs target-specific organization and grounded persuasion, but direct prompting can yield repetitive structures and unsupported claims. The paper introduces LandingBench and LandingAgent, which profile targets, adapt abstract reference patterns through wireframing, and polish drafts with critiques; experiments report improvements over direct prompting in faithfulness, conciseness, readability, aesthetics, and diversity.

  • Problem

    Direct landing-page generation can produce repetitive templates and unsupported persuasive claims, while raw reference pages entangle reusable strategies with source-specific content and implementation.

  • Method

    LandingBench abstracts real pages into reusable profiles, and LandingAgent uses Profiling, reference-guided Wireframing, and critique-guided Polishing to generate executable pages.

  • Results

    LandingAgent improves over direct prompting in faithfulness, conciseness, readability, aesthetics, and structural diversity.

  • Takeaways & Limitations

    Abstract reference profiles support target-grounded and diverse landing-page generation within the evaluated setting.

  • Takeaways & Limitations

    The evaluation provides proxy evidence rather than direct evidence of conversion effectiveness because pages were not deployed or measured for behavioral outcomes.

Abstract

from arXiv · show

Landing pages are goal-oriented web interfaces that must communicate a target-specific value proposition while organizing information flow, visual hierarchy, and calls to action (CTA). Although large language models can generate plausible webpage code from natural-language prompts, direct generation often yields generic templates and unsupported persuasive claims. We study target-grounded, reference-guided landing-page generation, where a system must create an executable page for a new target by adapting reusable patterns from real pages without copying them. We introduce LandingBench, a reference-profile dataset that abstracts real landing pages into section sequences, layout patterns, tone descriptors, visual emphasis, and CTA structure. Building on LandingBench, we propose LandingAgent, a three-phase agentic framework that profiles the target, constructs a reference-guided wireframe, and refines the page through critique-guided polishing. We evaluate LandingAgent against direct prompting on faithfulness, conciseness, readability, aesthetics, and structural diversity. Experiments show improved target grounding, presentation quality, and layout diversity. Code is available at https://github.com/IAURAI/LandingAgent.

1 Introduction

Landing-page generation requires target-specific decisions about messaging, information flow, visual hierarchy, tone, trust cues, and CTAs, yet direct LLM generation can produce repetitive structures and unsupported claims. The paper addresses this with LandingBench and LandingAgent, which use abstract references and staged generation to improve target grounding and page quality.

  • Landing pages must coordinate value propositions, information ordering, visual hierarchy, tone, trust cues, and CTA placement beyond executable code.
  • Direct LLM generation may produce repetitive templates, limited control over target-specific information architecture, and unsupported persuasive claims.
  • LandingBench abstracts real landing pages into reusable section, layout, visual, tone, CTA, and information-density annotations.
  • LandingAgent decomposes generation into Profiling, reference-guided Wireframing, and critique-guided Polishing.
  • The evaluation compares LandingAgent with direct LLM generation across target faithfulness, presentation quality, and structural diversity.

2 Related Work

Prior webpage-generation work largely reconstructs interfaces from screenshots or code-oriented inputs, while landing-page generation requires planning persuasive structures for new targets. The paper therefore frames the task around target faithfulness, presentation quality, diversity, and abstract reference-guided planning.

  • Prior LLM code-generation research primarily targets syntactically valid and functionally correct code from user specifications.
  • Landing-page generation additionally requires organizing value propositions, information flow, visual hierarchy, tone, and CTA strategy.
  • Existing UI and webpage datasets mainly pair screenshots with HTML or evaluate reconstruction of a given visual interface.
  • LandingBench instead captures reusable communication and design patterns for organizing a new target into a persuasive page structure.
  • Related agentic web-generation systems use different task formulations and evaluation settings, complicating direct comparison.

3 LandingBench: A Landing-Page Reference Annotation Dataset

LandingBench converts real landing pages into compact reference profiles rather than copyable templates, enabling retrieval of reusable structural, rhetorical, and visual patterns. Its collection pipeline combines filtering, scoring, and human verification before annotation.

  • LandingBench represents real landing pages as compact profiles describing reusable structural, rhetorical, visual, and conversion patterns.
  • The dataset is intended for target-grounded adaptation rather than reproducing source screenshots, copywriting, branding, layouts, or implementations.
  • Collection and Filtering: Candidate pages are sampled from Common Crawl and processed to identify pages whose structures and communication strategies can support retrieval-based generation.
  • Collection and Filtering: The hard-filter stage removes inaccessible, unsuitable, unsafe, and near-duplicate pages using rule-based checks, eligibility judgments, and perceptual hashing.
  • Collection and Filtering: The soft filter scores surviving pages on four quality dimensions, while those scores are used for candidate selection rather than gold-standard evaluation labels.
  • Collection and Filtering: Both human reviewers must accept a page for inclusion; across 636 candidates, agreement was Po = 0.821 with Cohen’s κ = 0.48.
  • Reference Profile Annotation: Each verified page receives page-level and section-level annotations covering communicative context, conversion attributes, section order, layout, and visual emphasis.
  • Dataset Statistics: LandingBench contains 438 pages covering 13 section-role labels, with an average of 8.0 annotated sections per page.

4 LandingAgent

LandingAgent generates executable landing pages through profiling, reference-guided wireframing, and critique-guided polishing. It separates global structural correction from local textual, visual, and functional refinement.

  • LandingAgent transforms a target specification and optional image assets into an executable landing page through three sequential phases.
  • Profiling: Profiling converts heterogeneous inputs into a Structured Page Brief containing target, audience, category, tone, conversion, information, and visual-identity attributes.
  • Wireframing: The Retriever uses brief-aligned retrieval axes to select a structurally relevant LandingBench profile, and the WireframeBuilder creates a draft conditioned on both.
  • Wireframing: Abstract profiles expose reusable structure and visual patterns without supplying raw HTML or verbatim source-page copy.
  • Wireframing: The WireframeCritic checks structural alignment and abstract reference use, restarting retrieval and generation when the draft is not approved.
  • Wireframing: The retrieve–build–critique process repeats up to Rmax = 3 times before the latest draft proceeds to polishing.
  • Polishing: Polishing preserves section order and major layout structure while refining copy clarity, CTAs, typography, spacing, visual coherence, and implementation consistency.
  • Polishing: Aesthetic, Message, and Functional critics independently assess presentation, target faithfulness and copy, and implementation consistency, using a default approval threshold of 4 out of 5.

5 Evaluation

LandingAgent is evaluated against prompting baselines using automatic metrics, qualitative comparisons, and a blind pairwise user study. It achieves stronger overall performance, while ablations and examples examine the roles of reference profiles, retrieval, profiling, and polishing.

  • Evaluation Setup: The evaluation measures Diversity, Faithfulness, Conciseness, Readability, and Aesthetics across prompting baselines using LLM-as-a-Judge, DINOv3 representations, and human evaluation.The study uses 30 target specifications spanning 12 industries and generates 540 pages in total.
  • Main Results: LandingAgent outperforms One-shot Prompting on all five metrics and Direct Prompting + Polisher on all but Faithfulness, where scores are comparable.One-shot Prompting improves over Direct Prompting overall, while Direct Prompting + Polisher improves Faithfulness and Conciseness but reduces Diversity and Aesthetics.
  • Qualitative Comparison: LandingAgent varies section flows while preserving target relevance, whereas Direct Prompting repeats layouts, visual assets, and section order.Qualitative examples also show a fully rendered target-grounded headline and cleaner visual alignment for LandingAgent.
  • User Study: Participants preferred LandingAgent across all criteria, selecting it 56–65% of the time versus 21–30% for the baseline.The preference was statistically significant for all five criteria, with OR 2.07–4.05 and all p < 10^-5 after Holm correction.
  • Ablation Study: The full ablation pipeline achieves the highest score on every metric, while removing LandingBench reduces performance across all five metrics, especially Diversity.The ablations isolate contributions from filtering, reference-pool composition, retrieval, Profiling, and Polishing.

6 Conclusion

The paper introduces LandingBench and LandingAgent for target-grounded, reference-guided landing-page generation. Experiments report improvements over direct prompting across content, presentation, and structural criteria, while future work targets deployed evaluation and conversion outcomes.

  • Conclusion: LandingBench abstracts real-world landing pages into reusable structural, rhetorical, and visual representations, while LandingAgent separates Profiling, Wireframing, and Polishing.The decomposition structures target specifications, adapts reference patterns for page planning, and refines message quality, presentation, and functional consistency.
  • Conclusion: Experiments show that LandingAgent improves over direct prompting in Faithfulness, Conciseness, Readability, Aesthetics, and Diversity.The conclusion attributes the supported result to abstract reference profiles from real landing pages enabling target-grounded, diverse generation.
  • Conclusion: Future work should evaluate generated pages in deployed settings and study how perceived page quality relates to real conversion outcomes.

Limitations

The work is limited by proxy-based evaluation, reference-pool coverage, and relatively narrow target-level coverage. Its computationally costly multi-stage pipeline further constrains exhaustive evaluation.

  • Evaluation scope: Conversion effectiveness is assessed indirectly through LLM-as-a-Judge, human evaluation, and automatic diversity metrics rather than real-world behavioral outcomes.The study does not deploy generated pages or measure click-through or conversion rates.
  • Reference coverage: LandingBench’s usefulness depends on coverage of its collected reference pages across industries, languages, regions, and design conventions.Under-represented targets may receive only approximate references.
  • Evaluation coverage: The main evaluation uses 30 target specifications, while the ablation study uses seven, limiting broader target-level coverage.Three independent generations per ablation target do not substitute for more targets.
  • Computational cost: Evaluating every ablation condition requires repeated execution of the multi-stage generation and evaluation pipeline, increasing computational cost.

A.6 Screening Agreement and Rejection Analysis

Human verification produced a conservative final selection, with moderate inter-reviewer agreement and disagreements concentrated among ambiguous B2B SaaS and app/tool pages. The accepted pages were then converted into automatically extracted reference profiles.

  • Screening agreement: Po = 0.821 and Cohen’s κ = 0.48 (95% CI: [0.40, 0.56]) indicate moderate agreement across 636 candidates.Acceptance required both reviewers to assign ACCEPT, forming a conservative intersection.
  • Screening agreement: Reviewer disagreements were concentrated among ambiguous B2B SaaS and app/tool pages.
  • Rejection analysis: 96% of rejected pages corresponded to identifiable non-landing-page types.Approximately 4% were borderline cases, primarily B2B SaaS and app/tool pages resembling dashboards or functional applications.
  • Reference-profile construction: Accepted pages are converted into reference profiles through automated vision–language-model extraction rather than manual annotation.The pipelines use full-page screenshots and extracted visible text to produce structured website and brand analyses.
  • Reference-profile construction: The two extraction outputs are repaired when necessary, validated against JSON schemas, and combined into each accepted page’s reference profile.

A.9 Dataset Statistics

LandingBench contains 438 verified landing pages organized into 14 normalized market-category labels, with additional distributions reported for brand and visual attributes.

  • Dataset size: 438 verified landing pages comprise the final LandingBench dataset.
  • Market categories: Market categories are normalized from VLM-generated free-text descriptions into 14 broad industry labels.
  • Dataset attributes: LandingBench reports distributions for brand archetype, company stage, brand polish level, and primary color hue.

B LandingAgent Implementation Details

LandingAgent’s implementation details document prompt configurations for eight agents, whose prompts share a common structured format. The appendix also includes a market-category distribution table.

  • Dataset documentation: The appendix includes Table 11, which presents the market-category distribution of LandingBench pages.
  • Agent prompts: Table 13 summarizes the inputs and output schemas of eight agents in execution order.The agents include PROFILER, RETRIEVER, WIREFRAMEBUILDER, WIREFRAMECRITIC, POLISHER, AESTHETICCRITIC, MESSAGECRITIC, and FUNCTIONALCRITIC.
  • Agent prompts: All eight agent prompts use the structural pattern ROLE / TASK / INPUT / OUTPUT SCHEMA with per-field guidance.

C.1 Metric Definitions

Landing pages are evaluated with five metrics covering variation, target alignment, information focus, ease of use, and visual quality.

  • The evaluation uses Diversity, Faithfulness, Conciseness, Readability, and Aesthetics.Definitions for all five metrics are compiled in Table 14.

C.2 Diversity Metric

Diversity measures visually induced variation among multiple pages generated for the same target. Pages are rendered, cropped to preserve the above-the-fold layout signal, embedded with a frozen vision encoder, and compared using cosine distance.

  • Diversity is computed within outputs for the same target, isolating variation induced by the generation method.Each target-method pair uses K=3 generated pages.
  • Each page is rendered as a full-page RGB screenshot and rescaled before a top-aligned 518×518 crop is extracted.The crop preserves aspect ratio and captures the above-the-fold and immediately scrollable hero region.
  • The frozen DINOv3 ViT-L/16 encoder converts each crop into a page-level embedding.The encoder is not fine-tuned, and identical weights and preprocessing are used across methods.
  • Pairwise page distance is cosine distance between L2-normalized embeddings, equaling zero only for embeddings pointing in the same direction.The distance is symmetric.

C.7 Runtime and Reproducibility Settings

The experiments use stochastic generation and report reproducibility settings, control-loop activity, compute usage, evaluation coverage, and de-identification procedures. These settings accompany LLM-as-a-Judge evaluation and ablation reporting.

  • Generation uses temperature 0.7 where supported, without a fixed seed, so repeated outputs are sampled stochastically.The reported prompt suite is version 1.0, with settings summarized in Table 18.
  • Across 90 full-pipeline runs, restart and retry counts describe control-loop invocation rather than standalone API or model failure rates.The distributions are reported in Table 19.
  • LandingAgent used approximately 428,000 tokens per generated page on average across those runs.
  • The ablation study covers seven target specifications, producing 21 samples per condition and 126 samples across six conditions.Ablation variants are listed in Table 20.
  • Qualitative results use real-world brand specifications that were de-identified by replacing names and UI marks and blurring remaining identifiable text.Figures 7 and 8 document the comparisons and de-identification procedure.
Loading 2608.27902v1…