Abstract: We contribute a comprehensive dataset to study user attention and purchasing behavior on Search Engine Result Pages (SERPs). Previous work has relied on mouse movements as a low-cost large-scale behavioral proxy but also has relied on self-reported ground-truth labels, collected at post-task, which can be inaccurate and prone to biases. To address this limitation, we use an eye tracker to construct an objective ground-truth of continuous visual attention. Our dataset comprises 2,776 transactional queries on Google SERPs, collected from 47 participants, and includes: (1) HTML source files, with CSS and images; (2) rendered SERP screenshots; (3) eye movement data; (4) mouse movement data; (5) bounding boxes of direct display and organic advertisements; and (6) scripts for further preprocessing the data. In this paper we provide an overview of the dataset and baseline experiments (classification tasks) that can inspire researchers about the different possibilities for future work. [ACM DL] [arXiv]
Please download the dataset from here. Each type of data (e.g. eye movement, mouse movement etc) has this naming convention:
[PARTICPANT_ID]-b[BLOCK_NUMBER]-t[TRIAL_NUMBER]
Example: p004-b1-t3
If you are looking for the corresponding data of a participant, for example p004 in block number 1 and in trial number 3, then the corresponding HTML file is in data/serp/p004-b1-t3.html, the fixation file is in data/fixation-data/p004-b1-t3.csv, the AOIs of the ads are in data/ad-boundary-data/p004-b1-t3.json and so on.
The (comma-delimited) data/participants.csv file has 8 columns:
participant_id: (string) Participant's ID.gender: (string) Participant's gender; ['Female', 'Male', 'Prefer not to say'].age: (integer) Participant's age.english: (string) Participant's English level; ['Intermediate', 'Advanced', 'Native speaker'].education: (string) Participant's education level; ['High school diploma', "Bachelor's degree", 'Graduate'].pref_device: (string) Participant's preferred device for online shopping; ['Smartphones', 'Computers or laptops'].pref_website: (string) Preferred website for online shopping; ['Amazon', 'eBay', 'Alibaba', 'Other'].online_shopping_freq: (string) Participant's online shopping frequency; ['Rarely', 'Weekly', 'Monthly'].
Example:
participant_id,gender,age,english,education,pref_device,pref_website,online_shopping_freq
p005,Female,36,Intermediate,Graduate,Computers or laptops,Rarely,Amazon
...The data/serps/ folder contains page snapshots (in HTML format) of the search engine results.
Rendered full-page screenshots of the corresponding SERPs are in data/screen-recordings/. All screenshots have the same viewport width (1280 px) but vary in height, depending on the document size.
Example:
These files in data/ad-boundary-data contain the slot boundaries of organic and direct-display ads on the screenshots in JSON format.
Each file contains 3 entries (ad types):
native_ad: Refers to organic (or native) adsdd_top: Refers to direct display ads that appear on the top-left of the SERPdd_right: Refers to direct display ads that appear on the right side of the SERP
For each ad type, the boundaries are defined with a list of dictionary-like items that contain:
location:xandydenote the top-corner coordinates of the AOI (relative to the top-left corner of each screenshot in pixels)size:widthandheightdenote the width and height of the ad, respectively, in pixels.
Note: For an ad type if there is no instance on the corresponding SERP, the value is an empty list [].
{
"native_ad": [
{
"location": {
"x": 162,
"y": 1900
},
"size": {
"height": 129,
"width": 540
}
},
{
"location": {
"x": 162,
"y": 2056
},
"size": {
"height": 108,
"width": 540
}
},
{
"location": {
"x": 162,
"y": 2191
},
"size": {
"height": 108,
"width": 540
}
}
],
"dd_top": [
{
"location": {
"x": 162,
"y": 158
},
"size": {
"height": 351,
"width": 586
}
}
],
"dd_right": []
}
An example of how the AOI information can be loaded as a Python dictionary:
import json
import pprint
def load_dict_from_file(file_path):
with open(file_path, "r") as f:
dict_obj = json.load(f)
return dict_obj
# example usage:
trial_ads_boundaries_dict = load_dict_from_file('path_to_the_ad_boundary_json_file')
pprint.pprint(trial_ads_boundaries_dict)The (comma-delimited) pupil dilation files in data/pupil-data has 7 columns:
timestamp: (int) Unix timestamps in millisecondsBPOGX: (int) Rhe best position of gaze x-coordinate, relative to the top-left corner of the screenshot in pixelsBPOGX: (int) Rhe best position of gaze y-coordinate, relative to the top-left corner of the screenshot in pixelsLPD: (int) Left eye pupil dilation in pixelsLPV: (int) The valid flag of the left pupil with the value of 1 if the data is valid, and 0 if it is notRPD: (int) Right eye pupil dilation in pixelsRPV: (int) The valid flag of the right pupil with the value of 1 if the data is valid, and 0 if it is not
Example:
timestamp,BPOGX,BPOGX,LPD,LPV,RPD,RPV
1671196599208,676,204,15.85818,1,14.75008,1,1065,463
...The (comma-delimited) eye fixation files in data/fixation-data has 4 columns:
timestamp: (int) Unix timestamps in millisecondsFPOGX: (int) Fixation positions of X, relative to the top-left corner of the screenshot in pixelsFPOGY: (int) Fixation positions of Y, relative to the top-left corner of the screenshot in pixelsFPOGD: (int) Fixation duration in milliseconds
Example:
timestamp,FPOGX,FPOGY,FPOGD
1671196599208,676,198,476
...The data/mouse-movement-data folder has all the log files, as recorded by the evtrack software.
Here you can find space-delimited files (CSV) with information about each event type:
Each CSV file has 8 columns:
timestamp: (int) Timestamp of the event, with millisecond precision.xpos: (int) X position of the mouse cursor.ypos: (int) Y position of the mouse cursor.event: (string) Browser's event name; e.g.load,mousemove,click, etc.xpath(string) Target element that relates to the event, in XPath notation.
timestamp,xpos,ypos,event,xpath
1671196599208,0,0,load,/
...The XML files in data/trial-metadata contain information about the user’s browser metadata (e.g. viewport size, user agent string).
Example of the XML files:
<data>
<ip>::1</ip>
<date>12/16/2022 14:16:39</date>
<timestamp>1671196599</timestamp>
<url>http://localhost/page.php?q=buy-momentum-brands-9-led-flashlight</url>
<ua>Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/108.0.0.0 Safari/537.36</ua>
<screen>1280x1024</screen>
<window>1422x1137</window>
<document>1403x2642</document>
<task>batch_20_trial_1 | buy-momentum-brands-9-led-flashlight</task>
</data>How to: (1) load the metadata from XML file as a dictionary, (2) load mouse movement data and make the coordinate relative to the top-corner of the screenshots:
import pandas as pd
import xml.etree.ElementTree as ET
def extract_trial_info(file_path):
tree = ET.parse(file_path)
root = tree.getroot()
task = root.findall("task")
batch = task[0].text.split(" | ")[0].strip().split("_")[1]
trial_number = task[0].text.split(" | ")[0].strip().split("_")[3]
slug = task[0].text.split(" | ")[1].strip()
info_dict = {'date': root.find("date").text,
'user-agent': root.find("ua").text,
'batch': int(batch),
'trial-number': int(trial_number),
'slug': slug,
'screen': root.find("screen").text,
'window': root.find("window").text,
'document': root.find("document").text,
}
return info_dict
mt_data_df = pd.read_csv('path_to_mouse_movement_csv_file')
trial_info_dict = extract_trial_info('path_to_corresponding_xml_file')
# screen size of the monitor
DISPLAY_WIDTH = 1280
DISPLAY_HEIGHT = 1024
# window size of the web browser
win_width = trial_info_dict['window-width']
win_height = trial_info_dict['window-height']
# extract those rows of data that a mousemove happend
mt_data_df = mt_data_df[mt_data_df['event'].isin(['mousemove', 'mouseover'])]
# converting the xs, ys to absolute positions on the corresponding screenshot
for index, row in mt_data_df.iterrows():
ratio_x = row['xpos']/win_width
ratio_y = row['ypos']/win_height
mt_data_df.loc[index, 'xpos'] = int(DISPLAY_WIDTH*ratio_x)
mt_data_df.loc[index, 'ypos'] = int(DISPLAY_HEIGHT*ratio_y)
# let's get each column as a list of values
timestamps = mt_data_df['timestamp'].values
events = mt_data_df['event'].values
cursor_xs = mt_data_df['xpos'].values
cursor_ys = mt_data_df['ypos'].values
xpaths = mt_data_df['xpath'].valuesIn data/screen-recordings you can find screen recordings for each trial.
Example (sped up and converted a GIF):
If you use the dataset, please cite us using this bibliographic reference:
- Kayhan Latifzadeh, Jacek Gwizdka, Luis A. Leiva. 2025. A Versatile Dataset of Mouse and Eye Movements on Search Engine Results Pages. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. https://doi.org/10.1145/3726302.3730325
@inproceedings{latifzadeh2025adserp,
title={A Versatile Dataset of Mouse and Eye Movements on Search Engine Results Pages},
author={Latifzadeh, Kayhan and Gwizdka, Jacek and Leiva, Luis A.},
booktitle={Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval},
year={2025},
doi={10.1145/3726302.3730325}
}The AdSERP dataset has been used in the paper(s) listed below. If you use the dataset in your publications, please email us, and we will include your paper here.
-
Mario Villaizán-Vallelado, Matteo Salvatori, Kayhan Latifzadeh, Antonio Penta, Luis A. Leiva, Ioannis Arapakis. 2025. AdSight: Scalable and Accurate Quantification of User Attention in Multi-Slot Sponsored Search. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. [DOI] [arXiv]
-
Stefan Lengauer, Sarah Annabelle Von Götz, Marie-Therese Hoesch, Florian Dieter Steinwidder, Mariia Tytarenko, Michael Bedek, Tobias Schreck. 2025. GUIDAETA—A Versatile Interactions Dataset with Extensive Context Information and Metadata. arXiv preprint arXiv:2511.20328. [arXiv]