The effects of the aesthetics and composition of hotels' digital photo images on online booking decisions

The study investigates how the aesthetics, composition, and content of hotels’ digital photos affect customers’ online booking decisions. Motivated by the dominance of survey- and experiment-based approaches and the emergence of AI-powered computer vision, the authors aim to identify specific photographic elements that drive actual booking outcomes and to develop a predictive model grounded in visual data mining. Research question: What aesthetics, composition and image elements relevant to hotel photography determine the positive effects on hotel online bookings? The work positions hotel imagery as a powerful communication tool that influences perception, emotions, and behavior throughout the booking journey, with implications for marketing theory and practice.

The review covers computational aesthetics and photographic composition rules (directional lines, framing, rule of thirds, symmetry, depth/angle, planning of objects, light/time of day) and their relevance to hotel photography. In hotel e-commerce, images reduce cognitive load, enhance webpage appeal and trust, and can significantly raise bookings; however, too many images may induce choice overload. Prior findings indicate that larger photos increase booking intentions (especially without people), while human presence can help smaller images; natural light, multiple angles, and images of public spaces/exteriors appeal to customers. Existing studies largely use surveys/experiments, leaving gaps regarding precise aesthetic properties and causal links to actual bookings; AI methods are emerging but underused in hotel marketing.

Design: Exploratory AI-driven visual analytics with predictive modeling and fuzzy cognitive mapping. Sampling and data collection: Random multi-stage sampling using a random digit generator. Destination selected: Barcelona, Spain (+3493). Hotel class randomly set to 4-star to avoid multi-class bias. Number of images per hotel set to 6. Data source: Booking.com. Identified 225 available 4-star hotels; collected 1350 digital photos (per Ren et al., 2021 distribution: 2 room, 1 bathroom, 2 exterior, 1 lobby). Target variable: binary booking event (mode of whether booked in the last hour), observed every 6 hours (four times in one day) per hotel, coded Yes=1, No=0. Image embedding: Google’s Inception v3 (48-layer CNN trained on ImageNet) used to embed each image into 2047-dimensional descriptor vectors (penultimate layer activations). Three images were unprocessed by Inception v3 and removed, leaving 1347 embedded images. Modeling: Supervised learning with cross-validation to predict booking (binary). Algorithms and key settings: Logistic regression (balanced distribution, L2 regularization, C=1); SVM (RBF kernel, epsilon=0.10, C=1.0, tol=0.001, max iter=100); MLP neural network (hidden units n=400, ReLU activation, Adam optimizer, alpha=0.0001, max iter=200). Performance assessed via AUC, classification accuracy (CA), precision, recall, and F1; confusion matrix analyzed for misclassifications. Fuzzy cognitive mapping (FCM): Based on high-scoring images (logistic regression predicted probability ≥0.6; 258 images), coded concepts included objects (e.g., pools, beach), shooting angle, time of day/light, human presence, and color scheme. Researchers mapped causal relationships among aesthetic properties and photo classes (exterior, lobby, room types) to derive practical composition rules.

Model performance: Logistic regression AUC=0.614, CA=0.578, F1=0.577; SVM AUC=0.881, CA=0.806, F1=0.806; MLP neural network AUC=0.903, CA=0.830, Precision=0.830, Recall=0.830, F1=0.830. Confusion matrix (MLP) on 1347 images showed high true positive and true negative rates (overall CA ≈83%). Selling image properties (four key dimensions identified by AI):

• Light/time of shooting: Natural light generally beneficial; exteriors can perform at any time with sunny conditions; rooms (bedrooms) favored daytime natural light; bathrooms often with artificial light.
• Color scheme: Subtle/monochromatic palettes perform well. Exteriors: monochromatic with light brown/dark gray harmonizing with surroundings. Rooms: white, brown, gray dominate; accents (pillows/flowers) for suite living areas. Lobbies: monochrome brown or black-and-white with red elements present.
• Human presence: Avoid humans in exterior and room images; presence of uniformed reception staff is beneficial in lobby images.
• Shooting angle: Exteriors: straight angle for entrances; elevated/top-down for pools/rooftops. Rooms/bathrooms: angled from corner or doorway to maximize perceived spaciousness; balcony/window views shot from the window area. Lobbies: angled from the opposite corner to enhance spaciousness. Category-specific taxonomy (from 258 high-probability images):
• Exterior (n=23): Any time; sunny blue-sky for pools; straight angle for entrances; no people; monochromatic subtle tones; pools/beachside are strong objects; day/night sequences also helpful.
• Lobby (n=12): Any time; natural and artificial light; shot from opposite side/corner; reception staff present; monochrome with red accents; include distinctive elements.
• Room bathroom (n=10): Artificial light; corner/doorway angle; no people; white/brown/gray palette; tidy white towels/robes recommended.
• Room bedroom (n=7): Daytime natural light; corner/doorway angle; no people; white/brown/gray palette.
• Room balcony/window view or suite living area (n=4): Any time; from window; no people; white/brown/gray with strong accents (pillows/flowers). Overall, appropriately composed and lit images with restrained color schemes and context-appropriate human presence are associated with higher booking likelihoods.

The findings directly answer the research question by identifying concrete aesthetic dimensions—time/light, color scheme, human presence, and angle—that increase the likelihood of bookings, and by providing category-specific composition rules. The study advances hotel marketing theory by: (1) applying AI-based computer vision and predictive modeling to actual booking-linked outcomes, moving beyond self-reported intentions; (2) demonstrating that image aesthetics contribute meaningfully to booking behavior alongside other known determinants (price, reviews, location), thus complementing multi-factor hotel choice models; (3) offering a working model that quantifies the influence of image properties, validating computer vision as a suitable tool for marketing inquiries; and (4) aligning with and extending prior work on the effects of image size, human presence, and natural light by revealing nuanced, context-dependent rules (e.g., staff presence helpful in lobby but not in rooms/exteriors). The results translate into actionable guidance for marketers and photographers to craft images that enhance persuasion and reduce information asymmetry.

The study proposes an AI-driven approach to evaluate and predict the selling power of hotel images, identifying four core dimensions—time/light, angle, human presence, and color scheme—that underpin effective imagery. Using Inception v3 embeddings and machine learning (best-performing MLP; AUC≈0.90, CA≈0.83), the authors show that specific, context-sensitive photographic choices can increase booking likelihood. They synthesize these into practical recommendations for different hotel photo classes (exterior, lobby, room types), enabling marketers to brief photographers effectively and create high-quality visuals that improve web experience and stimulate bookings. Future research should expand datasets, integrate control variables (e.g., stars, ratings, price), develop hotel-specific vision models, broaden destinations, and employ advanced CV methods (object recognition, GANs) for greater accuracy and generalizability.

• Model accuracy: While MLP achieved ≈83% precision/accuracy, higher (>90%) accuracy is desirable; larger, multi-destination datasets could improve performance.
• Domain-specific models: Lack of specialized neural networks trained solely on hotel images may limit precision.
• Missing controls: The predictive analysis used image embeddings as exogenous variables; adding hotel-level controls (stars, guest rating, room rate) could better isolate image effects.
• Geographic scope: Data restricted to 4-star hotels in Barcelona; limits generalizability.
• Fuzzy cognitive mapping: Relies on researcher judgment; more automated object detection and GAN-based methods could enhance validity in future work.