Русский
English
A small thumbnail research project can reveal the visual conventions of your niche and identify ideas worth testing on your own channel. It cannot tell you the click-through rate of another creator's video or prove that a design choice caused its views. Treat the output as a shortlist of hypotheses, with a record of how you found them.
You need a spreadsheet, a consistent sampling rule and a narrow question. A manageable project completed carefully is more useful than hundreds of loosely collected screenshots.
Define one question before collecting examples
“What thumbnails work?” leaves too much open to interpretation. Ask something you can describe and record:
- Do standout used-car reviews show the price on the thumbnail?
- How often do cooking tutorials use no thumbnail text?
- In renovation videos, is the finished room or the creator's face the main subject?
Write the question above the spreadsheet before looking at results. This reduces the temptation to redefine the project around whichever pattern seems exciting afterward. Decide whether you are studying search-oriented tutorials, entertainment, interviews or another specific format.
Build a comparable sample
Choose channels in the same language and topic area. A practical starting plan might include 15–30 channels and 10–20 comparable videos per channel, but these are workload suggestions, not statistical requirements. Include creators near your own scale as well as a few established ones. A sample consisting entirely of famous personalities tells you relatively little about packaging for unfamiliar viewers.
Keep long-form videos separate from Shorts and livestreams. Define a publication window and record each video's age when you collect the data. A video published last week has not had the same opportunity to accumulate views as a video published six months ago. Where possible, compare similar ages and formats within each channel.
Record the URL, title, publication date, collection date and public view count. Also note obvious differences such as a topical event, a collaboration or a substantially different video length. Public information will not expose every cause of performance; record uncertainty rather than filling gaps with guesses.
Use an index without mistaking it for a causal measure
One useful descriptive measure is the video's views divided by the median views of comparable sampled videos on that channel. An index of 1 means the median level; 3 means three times that level. Do not calculate it when the median is zero.
Choose a “standout” threshold before coding, such as an index of 2. The threshold is your analysis convention, not a YouTube benchmark. Normalizing by the channel's median helps with scale, but it does not remove differences in age, topic, promotion or audience. Keep the underlying counts visible.
Create a short coding guide
Choose a small set of features connected to your question. Define what each label means so you do not change the rules halfway through.
| Feature | Possible values | Example coding rule |
|---|---|---|
| Face | None, small, large | Large means at least half the image height |
| Expression | Neutral, happy, surprised, doubtful, other | Code the main face; mark unclear cases |
| Text | None, 1–3 words, 4 or more | Count all visible words; count a number as one item |
| Number | Present, absent | Include prices, durations and quantities |
| Title repetition | None, partial, exact | Compare image lettering with the title |
| Main subject | Person, object, place, diagram | Choose the dominant subject at thumbnail size |
| Background | Light, dark, visually busy | Judge the reduced image, not the full-size file |
| Visual approach | Before/after, comparison, reaction, close-up, other | Record one principal approach |
The labels are tools for consistency, not universal definitions of design. If your question is about prices, you may not need to code every expression or background color. Extra columns add work and make incomplete data more likely.
Code without looking at performance
Hide the view-count and index columns while reviewing thumbnails. Otherwise, a popular video can unconsciously look more “clear” or “expressive” because you already know the result.
Ask another person to code a small overlapping set, perhaps 20 examples. Discuss disagreements and tighten the definitions before continuing. AI can assist with initial labels, but review ambiguous expressions, text and main-subject judgments yourself. A consistent human check matters more than a fast but unstable classification.
Compare patterns cautiously
Split the sample into your predefined standout group and the remaining videos. For each feature, report both the share and the count: “3 of 8” and “30 of 80” have the same percentage but very different support. Look across channels, not just across individual videos.
- A pattern concentrated in one creator may reflect that creator's brand.
- A pattern concentrated in one hot topic may reflect demand for the topic.
- A tiny difference in a small sample is not a practical design rule.
- A feature used by nearly everyone describes a convention; without meaningful alternatives, it does not establish superiority.
Save examples that challenge your preferred explanation as well as those that support it. They help turn “faces work” into a more useful question about when a face adds context.
Test a hypothesis on your own videos
Turn one observation into a specific design question: “For this used-car review, does showing the price improve the package compared with leaving it out?” Prepare variants that keep the subject and message comparable. If you change several elements together, record that you are testing a concept, not isolating one feature.
YouTube's native desktop A/B test can compare up to three eligible variants and evaluates watch time. It requires access to advanced features and is not available for Shorts.
If native testing is unavailable, compare your own results cautiously across matched formats and traffic sources. Sequential replacements and different videos are observational comparisons, not equivalent experiments.
Keep a one-page research summary
- The original question and collection date.
- The sampling window, included formats and exclusions.
- The index definition and coding guide.
- Findings with counts and known confounders.
- Two or three testable ideas.
- Your own test results added later.
In Bloggers.tools, Professional mode provides three concepts and lets you edit the selected concept before generating an image. Use that description to specify the hypothesis you want to explore. For broader channel questions, consider the channel audit.
Frequently asked questions
Can I find another channel's thumbnail CTR?
It is not available in public YouTube view counts. The owner or someone with analytics access can share it. Without registered impressions, you cannot calculate it from views.
How many videos are enough?
There is no universally sufficient number. Choose a scope you can code reliably, show the group sizes and avoid treating a small standout group as proof. More examples do not fix a biased sample.
How often should I repeat the research?
Repeat when the format, competitors or audience changes enough to affect a real decision. A six-to-twelve-month review can be a useful planning habit, not a scientific expiry date. Keep the previous snapshot for comparison.
Can AI do all the coding?
It can prepare labels, but ambiguous imagery and expression categories still need verification. Use the same coding guide for machine and human work and document corrections.
Platform references: native A/B testing and interpreting impressions CTR.