Website Performance Analysis involves evaluating various metrics related to a website’s functionality, user engagement, and overall success in achieving business goals. This form of analysis is critical because it directly impacts user experience, conversion rates, and the profitability and reputation of a business. So, if you want to learn how to analyze the performance of a website, this article is for you. In this article, I’ll take you through the task of Website Performance Analysis using Python.
Website Performance Analysis: Getting Started
The problem I am working on in this article is my take on a problem I found at statso. The dataset we are working on contains the following columns:
- Session primary channel group: The marketing channel (e.g., Direct, Organic Social)
- Date + hour (YYYYMMDDHH): The specific date and hour of the session
- Users: Number of users in a given period
- Sessions: Number of sessions in that period
- Engaged sessions: Number of sessions with significant user engagement
- Average engagement time per session: The average time a user is engaged per session
- Engaged sessions per user: Ratio of engaged sessions to total sessions per user
- Events per session: Average number of events (actions taken) per session
- Engagement rate: The proportion of sessions that were engaged
- Event count: Total number of events during the period
You can download the dataset and read the complete problem statement here.
Website Performance Analysis using Python
Now, let’s get started with the task of Website Performance Analysis by importing the necessary Python libraries and the dataset:
import pandas as pd
data = pd.read_csv("data-export.csv")
print(data.head())# ---------------------------------------- \
0 Session primary channel group (Default channel...
1 Direct
2 Organic Social
3 Direct
4 Organic Social
Unnamed: 1 Unnamed: 2 Unnamed: 3 Unnamed: 4 \
0 Date + hour (YYYYMMDDHH) Users Sessions Engaged sessions
1 2024041623 237 300 144
2 2024041719 208 267 132
3 2024041723 188 233 115
4 2024041718 187 256 125
Unnamed: 5 Unnamed: 6 \
0 Average engagement time per session Engaged sessions per user
1 47.526666666666700 0.6075949367088610
2 32.09737827715360 0.6346153846153850
3 39.93991416309010 0.6117021276595740
4 32.16015625 0.6684491978609630
Unnamed: 7 Unnamed: 8 Unnamed: 9
0 Events per session Engagement rate Event count
1 4.673333333333330 0.48 1402
2 4.295880149812730 0.4943820224719100 1147
3 4.587982832618030 0.49356223175965700 1069
4 4.078125 0.48828125 1044
There are some errors in the first row of the dataset, which usually occurs while collecting the data from websites. The data starts from the second row, let’s prepare it accordingly:
new_header = data.iloc[0] # grab the first row for the header data = data[1:] # take the data less the header row data.columns = new_header # set the header row as the df header data.reset_index(drop=True, inplace=True) print(data.head())
0 Session primary channel group (Default channel group) \
0 Direct
1 Organic Social
2 Direct
3 Organic Social
4 Organic Social
0 Date + hour (YYYYMMDDHH) Users Sessions Engaged sessions \
0 2024041623 237 300 144
1 2024041719 208 267 132
2 2024041723 188 233 115
3 2024041718 187 256 125
4 2024041720 175 221 112
0 Average engagement time per session Engaged sessions per user \
0 47.526666666666700 0.6075949367088610
1 32.09737827715360 0.6346153846153850
2 39.93991416309010 0.6117021276595740
3 32.16015625 0.6684491978609630
4 46.918552036199100 0.64
0 Events per session Engagement rate Event count
0 4.673333333333330 0.48 1402
1 4.295880149812730 0.4943820224719100 1147
2 4.587982832618030 0.49356223175965700 1069
3 4.078125 0.48828125 1044
4 4.529411764705880 0.5067873303167420 1001
Now, let’s have a look at the column info and the summary statistics of the data:
data.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 3182 entries, 0 to 3181
Data columns (total 10 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 Session primary channel group (Default channel group) 3182 non-null object
1 Date + hour (YYYYMMDDHH) 3182 non-null object
2 Users 3182 non-null object
3 Sessions 3182 non-null object
4 Engaged sessions 3182 non-null object
5 Average engagement time per session 3182 non-null object
6 Engaged sessions per user 3182 non-null object
7 Events per session 3182 non-null object
8 Engagement rate 3182 non-null object
9 Event count 3182 non-null object
dtypes: object(10)
memory usage: 248.7+ KB
print(data.describe())
0 Session primary channel group (Default channel group) \
count 3182
unique 7
top Direct
freq 672
0 Date + hour (YYYYMMDDHH) Users Sessions Engaged sessions \
count 3182 3182 3182 3182
unique 672 147 180 103
top 2024042417 1 1 0
freq 6 335 340 393
0 Average engagement time per session Engaged sessions per user \
count 3182 3182
unique 2823 808
top 0 0
freq 170 393
0 Events per session Engagement rate Event count
count 3182 3182 3182
unique 2025 986 678
top 1 0 1
freq 133 393 115
Now, let’s convert the date column into an appropriate datetime format and group it for further analysis:
data['Date + hour (YYYYMMDDHH)'] = pd.to_datetime(data['Date + hour (YYYYMMDDHH)'], format='%Y%m%d%H')
data['Users'] = pd.to_numeric(data['Users'])
data['Sessions'] = pd.to_numeric(data['Sessions'])
# group data by date and sum up the users and sessions
grouped_data = data.groupby(data['Date + hour (YYYYMMDDHH)']).agg({'Users': 'sum', 'Sessions': 'sum'})The overall purpose of the above operation is to prepare and summarize the dataset for time series analysis, focusing on how user engagement (through sessions) varies by time. By converting data into appropriate types and grouping it by time, you can more easily perform operations like plotting time series graphs, calculating moving averages, or applying time series forecasting models.
Now, let’s analyze the total users and sessions over time:
import matplotlib.pyplot as plt
# plotting the aggregated users and sessions over time
plt.figure(figsize=(14, 7))
plt.plot(grouped_data.index, grouped_data['Users'], label='Users', color='blue')
plt.plot(grouped_data.index, grouped_data['Sessions'], label='Sessions', color='green')
plt.title('Total Users and Sessions Over Time')
plt.xlabel('Date and Hour')
plt.ylabel('Count')
plt.legend()
plt.grid(True)
plt.show()
From the graph, we can observe there are some fluctuations in the number of users and sessions, possibly indicating daily cycles or specific high-traffic periods. Both users and sessions appear to follow a similar trend, which is expected as more users generally mean more sessions. Some peaks might correspond to specific marketing activities, promotions, or events.
Now that we’ve analyzed the session trends, let’s move on to User Engagement Analysis. We will look into metrics like average engagement time per session, engagement rate, and events per session to evaluate how engaged users are when they visit the site:
# convert relevant columns to numeric for engagement analysis
data['Engaged sessions'] = pd.to_numeric(data['Engaged sessions'])
data['Average engagement time per session'] = pd.to_numeric(data['Average engagement time per session'])
data['Engaged sessions per user'] = pd.to_numeric(data['Engaged sessions per user'])
data['Events per session'] = pd.to_numeric(data['Events per session'])
data['Engagement rate'] = pd.to_numeric(data['Engagement rate'])
# group data by date and calculate mean for engagement metrics
engagement_metrics = data.groupby(data['Date + hour (YYYYMMDDHH)']).agg({
'Average engagement time per session': 'mean',
'Engaged sessions per user': 'mean',
'Events per session': 'mean',
'Engagement rate': 'mean'
})
# plotting engagement metrics
fig, ax = plt.subplots(4, 1, figsize=(14, 20), sharex=True)
ax[0].plot(engagement_metrics.index, engagement_metrics['Average engagement time per session'], label='Avg Engagement Time', color='purple')
ax[0].set_title('Average Engagement Time per Session')
ax[0].set_ylabel('Seconds')
ax[1].plot(engagement_metrics.index, engagement_metrics['Engaged sessions per user'], label='Engaged Sessions/User', color='orange')
ax[1].set_title('Engaged Sessions per User')
ax[1].set_ylabel('Ratio')
ax[2].plot(engagement_metrics.index, engagement_metrics['Events per session'], label='Events per Session', color='red')
ax[2].set_title('Events per Session')
ax[2].set_ylabel('Count')
ax[3].plot(engagement_metrics.index, engagement_metrics['Engagement rate'], label='Engagement Rate', color='green')
ax[3].set_title('Engagement Rate')
ax[3].set_ylabel('Rate')
ax[3].set_xlabel('Date and Hour')
for a in ax:
a.legend()
a.grid(True)
plt.tight_layout()
plt.show()
The user engagement analysis provides insights into how visitors interact with the website:
- Average Engagement Time per Session: The time spent per session shows fluctuations over the observed period. There are noticeable peaks, suggesting times when users were particularly engaged, potentially due to specific content releases or events.
- Engaged Sessions per User: This ratio fluctuates slightly but generally indicates that a good portion of sessions per user are engaged. Peaks in this metric could correspond to times when users find the content more relevant or engaging.
- Events per Session: The count of events per session remains relatively consistent but does show some variation. Peaks here could indicate more interactive content or features being used by visitors.
- Engagement Rate: The engagement rate over time shows how many sessions are considered engaged out of the total. There are some ups and downs which may relate to how different content resonates with users or how effective certain user acquisition channels are.
Now, let’s analyze the correlations between them:
fig, axes = plt.subplots(2, 2, figsize=(12, 10))
# plot 1: average engagement time vs events per session
axes[0, 0].scatter(data['Average engagement time per session'], data['Events per session'], color='blue')
axes[0, 0].set_title('Avg Engagement Time vs Events/Session')
axes[0, 0].set_xlabel('Average Engagement Time per Session')
axes[0, 0].set_ylabel('Events per Session')
axes[0, 0].grid(True) # enable grid
# plot 2: average engagement time vs engagement rate
axes[0, 1].scatter(data['Average engagement time per session'], data['Engagement rate'], color='red')
axes[0, 1].set_title('Avg Engagement Time vs Engagement Rate')
axes[0, 1].set_xlabel('Average Engagement Time per Session')
axes[0, 1].set_ylabel('Engagement Rate')
axes[0, 1].grid(True)
# plot 3: engaged sessions per user vs events per session
axes[1, 0].scatter(data['Engaged sessions per user'], data['Events per session'], color='green')
axes[1, 0].set_title('Engaged Sessions/User vs Events/Session')
axes[1, 0].set_xlabel('Engaged Sessions per User')
axes[1, 0].set_ylabel('Events per Session')
axes[1, 0].grid(True)
# plot 4: engaged sessions per user vs engagement rate
axes[1, 1].scatter(data['Engaged sessions per user'], data['Engagement rate'], color='purple')
axes[1, 1].set_title('Engaged Sessions/User vs Engagement Rate')
axes[1, 1].set_xlabel('Engaged Sessions per User')
axes[1, 1].set_ylabel('Engagement Rate')
axes[1, 1].grid(True)
plt.tight_layout()
plt.show()
Here’s what we can analyze from the above scatter plots:
- Average Engagement Time vs Events per Session: There appears to be a concentration of data points at lower average engagement times with a wide range of events per session. As the average engagement time increases, the number of events per session tends to cluster more narrowly around lower values.
- Average Engagement Time vs Engagement Rate: There is a clear trend where sessions with very low engagement times have a broad range of engagement rates, but as engagement time increases, the engagement rate converges towards higher values.
- Engaged Sessions per User vs Events per Session: Most data points cluster at lower values for both metrics, with few users having a high number of engaged sessions or events per session.
- Engaged Sessions per User vs Engagement Rate: There is a strong positive correlation between engaged sessions per user and engagement rate, especially noticeable at higher values of engaged sessions per user.
We will now proceed with the Channel Performance Analysis to assess how different marketing channels contribute to traffic and engagement. It will involve analyzing the session, user, and engagement data segmented by the marketing channel:
# group data by channel and aggregate necessary metrics
channel_performance = data.groupby('Session primary channel group (Default channel group)').agg({
'Users': 'sum',
'Sessions': 'sum',
'Engaged sessions': 'sum',
'Engagement rate': 'mean',
'Events per session': 'mean'
})
# normalize engagement rate and events per session for comparison
channel_performance['Normalized Engagement Rate'] = channel_performance['Engagement rate'] / channel_performance['Engagement rate'].max()
channel_performance['Normalized Events per Session'] = channel_performance['Events per session'] / channel_performance['Events per session'].max()
# plotting channel performance metrics
fig, ax = plt.subplots(3, 1, figsize=(12, 18))
# users and sessions by channel
ax[0].bar(channel_performance.index, channel_performance['Users'], label='Users', alpha=0.8)
ax[0].bar(channel_performance.index, channel_performance['Sessions'], label='Sessions', alpha=0.6)
ax[0].set_title('Users and Sessions by Channel')
ax[0].set_ylabel('Count')
ax[0].legend()
# normalized engagement rate by channel
ax[1].bar(channel_performance.index, channel_performance['Normalized Engagement Rate'], color='orange')
ax[1].set_title('Normalized Engagement Rate by Channel')
ax[1].set_ylabel('Normalized Rate')
# normalized events per session by channel
ax[2].bar(channel_performance.index, channel_performance['Normalized Events per Session'], color='green')
ax[2].set_title('Normalized Events per Session by Channel')
ax[2].set_ylabel('Normalized Count')
plt.tight_layout()
plt.show()
The data illustrates significant variations in performance across different channels, highlighting the strengths and weaknesses of each in driving traffic, engaging users, and encouraging interactions. The high performance of ‘Organic Search’ in driving traffic contrasts with its lower relative engagement and events metrics, suggesting quantity over quality of visits. In contrast, ‘Referral’ and ‘Organic Video’ channels, while not leading in volume, excel in engaging users deeply, pointing to potential areas for leveraging these strengths in marketing strategies.
Forecasting Website Traffic
Now, let’s see how we can forecast the website traffic for the next 24 hours. It will involve creating a time series model to forecast future values based on the observed session data. Let’s start this by plotting the autocorrelation and partial autocorrelation plots of our time series data, which is typically done to help identify the order of an autoregressive integrated moving average for time series models for forecasting:
from statsmodels.graphics.tsaplots import plot_acf, plot_pacf
time_series_data = grouped_data['Sessions'].asfreq('H').fillna(method='ffill')
seasonal_period = 24
differenced_series = time_series_data.diff().dropna()
# plot ACF and PACF of time series
fig, axes = plt.subplots(1, 2, figsize=(12, 4))
plot_acf(differenced_series, ax=axes[0])
plot_pacf(differenced_series, ax=axes[1])
plt.show()
Here’s how to interpret the above graph:
- PACF (Partial Autocorrelation Function): This plot helps determine the p parameter for the AR part of the model. You look for the lag after which most partial autocorrelations are not significantly different from zero. In our plot, the PACF shows a significant spike at lag 1 and then cuts off, suggesting an AR part of order 1. Therefore, p=1.
- ACF (Autocorrelation Function): This plot helps identify the q parameter for the MA part of the model. You look for the lag after which most autocorrelations are not significantly different from zero. The ACF plot in our case tails off gradually, but considering the first significant spike is essential. Since the spike at lag 1 is significant and there’s a gradual tailing off rather than a sharp cut-off, it suggests a potential MA component. However, the tailing-off nature complicates the exact determination of q, but a starting point of q=1 could be considered.
The other parameter is d, representing seasonality. In our case, as seasonality exists, we can choose the value of d as 1. Now, here’s how we can forecast the website’s traffic for the next 24 hours using the SARIMA model:
from statsmodels.tsa.statespace.sarimax import SARIMAX
time_series_data = grouped_data['Sessions'].asfreq('H').fillna(method='ffill')
seasonal_period = 24
sarima_model = SARIMAX(time_series_data,
order=(1, 1, 1),
seasonal_order=(1, 1, 1, seasonal_period))
sarima_model_fit = sarima_model.fit()
# forecast the next 24 hours using the SARIMA model
sarima_forecast = sarima_model_fit.forecast(steps=24)
# plotting the actual data and the SARIMA forecast
plt.figure(figsize=(14, 7))
plt.plot(time_series_data.index[-168:], time_series_data[-168:], label='Actual Sessions', color='blue') # last week data
plt.plot(pd.date_range(time_series_data.index[-1], periods=25, freq='H')[1:], sarima_forecast, label='Forecasted Sessions', color='red')
plt.title('Website Traffic Forecasting with SARIMA (Sessions)')
plt.xlabel('Date and Hour')
plt.ylabel('Sessions')
plt.legend()
plt.grid(True)
plt.show()
So this is how we can analyze the performance of a website and forecast its traffic using Python.
Summary
So, in this article, we conducted a comprehensive analysis of the website’s performance, based on:
- Session Analysis: Understanding traffic trends.
- User Engagement Analysis: Gauging the depth of user interaction.
- Channel Performance: Evaluating which channels are most effective.
- Website Traffic Forecasting: Predicting future traffic patterns.
I hope you liked this article on website performance analysis using Python. Feel free to ask valuable questions in the comments section below. You can follow me on Instagram for many more resources.
