Frequently Asked Questions
What is the goal of CreatorML?
Learn from your data and other channel data what kinds of thumbnails and titles work, and have a safe way to see how changing your titles and thumbnails will affect your click-through rate.
Several big channels are already using it before uploading to take their thumbnail and titles, and optimize them to get the highest Estimated CTR% possible, before uploading to YouTube.
How does it work?
Get accurate CTR estimates for your videos using the power of machine learning.
Supply your video's title, thumbnail, duration, and YouTube channel to instantly receive an estimated Click-Through Rate percent (CTR).
This tool uses a machine learning model trained on over 28-day CTR data snapshots from over 1000 different videos, titles, and thumbnail combinations. Using that information, the model detects patterns in the thumbnails and titles that correlate well with high click-through rates.
The model uses computer vision and natural language processing in order to obtain estimates that are accurate within half a standard deviation of the true click-through rate as shown in YouTube studio. The model can be fine-tuned for specific channels; scroll down for more information on this.
Below, we describe the best ways to use this tool.
Try different titles
Changing the title will have a moderate effect on CTR estimate.
Once you've uploaded your thumbnail and title, you now have a baseline CTR. Next, you can try changing the title. Try different keywords, phrasings, and title lengths. For example, shorter titles (<40 characters) tend to get a higher CTR.
- It currently only works for *English language* titles.
- It is *case-insensitive*. For example, “THIS TITLE” and “this title” will be treated the same.
Try different thumbnails
Changing the video Thumbnail has a large impact on estimated CTR.
Statistically, the biggest thing that impacts CTR is the thumbnails. The model analyzes the thumbnail for colors, brightness, texture, and the presence of objects (computers, phones, cars) and people in the image. Additionally, it extracts text from the thumbnail, and considers it in the CTR estimation.
CreatorML will automatically generate a *Thumbnail Attention Heatmap* that indicates the things in the thumbnail that correlate strongly with a higher CTR. Warmer (red) areas, are elements of the thumbnail that draw viewer attention and lead to an increase in the Estimated CTR.For example:
Originally uploaded thumbnail.
Thumbnail with Attention Heatmap overlay.
Notice how the thumbnail text and logos get most of the attention in this thumbnail.
See how CTR changes as views increase
Expand on the CTR single point estimate, and see how it changes with duration and views.
Once you get your estimate, you can scroll down to see a chart of how Click-Through Rate Percent changes as the number of views increases. It is common knowledge that the YouTube algorithm factors both CTR and Average View Duration (AVD) when evaluating a video's performance, and whether to increase the number of people to show it to (impression count). Higher CTR videos will tend to have more views, and vice versa, though viral videos tend to start with a much higher click-through rate until it is shown to a broader audience.
How do I get more accurate CTR predictions for my YouTube channel?
By default, CreatorML uses a Shared CTR Predictor which finds patterns across 20,000 video titles and thumbnails that earn the viewer’s click. But you can get up to 60% more accurate estimates with a Custom CTR Predictor that learns from the other channels, and is fine-tuned to your channel’s audience viewing behavior.
For every Custom CTR Predictor created, the predictions for every CTR Predictor on the platform become more accurate. This is because the underlying machine learning algorithms benefit from more data to learn from.
All CreatorML needs to build a Custom CTR Predictor is a snapshot of the last 28 days of CTR data from YouTube studio. The instructions for exporting and uploading the data can be found here: https://creatorml.com/get-ctr-predictor
Making a Custom CTR Predictor is still a manual process at the moment, so it takes me a few days, but we're working on automating it over time. You will receive a Custom CTR Predictor Secret you can use to make predictions with it.
Does video duration affect CTR?
Yes, the duration (aka length) of a video has a big impact on CTR. For example, generally a 10:00 video will have a lower CTR than a 6:00 video, but will have a higher CTR than a 30:00 video.
When viewers see the thumbnail they also see the duration too. Longer videos tend to be clicked less because they are longer and are more of a commitment to watch. Have you ever not clicked a video because it was too long?
In some ways, you can think of the video duration as the attention price-tag of a YouTube video.
Who made CreatorML?
My name is Charles Weill, and I'm an engineer who used to do Machine Learning at Google Research. For the past 5 years, I worked on some of Google's biggest Machine Learning projects.
Now I make YouTube videos about Machine Learning, and build AI-powered software tools to help YouTubers get the most out of our YouTube channels. You can follow me on Twitter as I build this product in public, and share regular updates and discoveries.
How much does a CreatorML subscription cost?
CreatorML is free for a limited usage. Each CTR prediction uses 1 credit. When you run out of credits, you will need to upgrade to a paid plan to get more credits. Paid plans come with other benefits, such as receiving refreshed Custom CTR Predictors every month.
Visit the CreatorML Pricing page to learn more.
How do I cancel my subscription?
You can cancel your subscription on the Manage Subscription page.
All you need is the License Key you received in your inbox when you purchased your subscription.