What 300 YouTube Titles Taught Me About CTR (Data from Building TitleScore)

Background

I built TitleScore — a tool that scores YouTube titles 0–100 using a rubric fed into the Claude API. Since launching, I've had the chance to see a lot of real titles come through. Patterns emerge fast.

This post is about what those patterns actually look like — the structural mistakes that reliably tank scores, and what the high-scoring alternatives have in common.

The Scoring Dimensions

TitleScore evaluates titles across five dimensions:

• Curiosity gap — does the title withhold something the viewer needs to click to get?
• Front-load strength — are the first 3–4 words doing real work?
• Emotional stakes — is something at risk, or is this just information delivery?
• Specificity — numbers, names, concrete outcomes vs. vague generalities
• Action orientation — does the title sell the click, or describe the content?

Each dimension scores 0–10. The weighted total becomes the 0–100 score.

The Patterns

Pattern 1: Describing Instead of Selling

This is the most common low-score pattern. Titles like:

• "How I Film My Videos" → score: 22
• "My Morning Routine" → score: 18
• "What I Eat in a Day" → score: 31

These describe the content accurately. They fail to make a case for why you should click. Compare:

• "The Camera Setup That Doubled My Watch Time (Under $400)" → score: 79
• "I Tracked Every Meal for 90 Days — Here's What Actually Changed" → score: 74

Same underlying content. Completely different framing.

Pattern 2: Passive Front Loading

The first three words carry outsized weight. Titles that open with "A Look At," "All About," "Let's Talk," or "In This Video" reliably score under 30 on front-load strength. These are filler openers — they delay the interesting part.

High-scoring openers tend to be:

• Numbers: "5 Mistakes That...", "I Spent $10K On..."
• Verbs: "Stop Doing This If...", "I Quit My Job To..."
• Named subjects: "[Person] Just Changed...", "The [Specific Thing] That..."

Pattern 3: Zero Stakes

Title stakes don't have to be dramatic. They just have to make it clear that something matters in this video. The best stakes patterns:

• Money: "I Lost $2,000 Doing This"
• Time: "I Wasted 3 Years Before Learning This"
• Social proof reversal: "Why I Stopped Following [Common Advice]"
• Transformation: "The One Change That Fixed My [Specific Problem]"

Titles without any stakes signal tend to score 15–25 points lower across the board.

Building the Rubric Into a Prompt

The technical challenge was getting Claude to apply a consistent rubric rather than generating vibes-based feedback. The approach that worked:

1. Define each dimension with explicit scoring criteria and examples in the system prompt
2. Require structured JSON output with per-dimension scores AND reasoning for each
3. Temperature = 0 for consistency
4. Include a few-shot example showing a low-score title and a high-score title with full breakdowns

The few-shot examples made the biggest difference. Without them, Claude's scores were accurate but the reasoning was generic. With them, it mirrors the specific vocabulary of the rubric.

Try It

gettitlescore.com — free, no account required. Paste a title, get the breakdown.

If the scoring feels off for your niche, I'd genuinely want to know — calibration across different content categories is an ongoing project.