Answers/Original research/I Tested 40+ YouTube Titles Against vidIQ's Scoring Model — One Clause Moved the Score 24 Points

Original research

I Tested 40+ YouTube Titles Against vidIQ's Scoring Model — One Clause Moved the Score 24 Points

By Boboy Watson·Last verified August 31, 2026·1,133 words·Original research

Most advice about YouTube titles is assertion. Someone tells you to use curiosity gaps, or numbers, or emotional words, and there is no way to check whether it is true or whether it is just what worked for them once.

I had access to a scoring model and time, so I ran structured tests instead — including proper controls, where a single clause changed and everything else stayed identical. This is what came back.

Method, and its limits

Titles were drafted around four recurring subjects on my channel and submitted to vidIQ's AI Coach title-scoring tool in batches, with the exact numeric score recorded for each. Over forty titles across several sessions. Where I was testing a specific variable, I held every other clause word-for-word identical and changed only the one under test.

The honest limitation, stated upfront: these are model scores, not view counts. A 97 is a prediction made by a tool, not evidence that a video performed. Every session also included a deliberately weak control variant to confirm the model was responding to substance rather than pattern-matching my phrasing — it consistently was.

Read what follows as findings about what the model rewards. That is genuinely useful when drafting, and it is not the same claim as proven view lift.

Finding 1 — One clause moved the score 24 points

This was the result that reframed everything else.

I took a title with a three-part structure and held the first and third parts identical, changing only the middle consequence clause:

Same opening. Same closing. Same subject. A 24-point spread from one clause.

The ranking is not random. It tracks a single dimension almost perfectly:

• 97 — specific, financial, personal, visceral. Names actual damage.
• 95 — a soft metaphor. Damage implied but unnamed.
• 82 — dramatic but vague. "Everything" specifies nothing.
• 73 — a systemic claim. Sociology rather than something that happened to a person.

The lesson is that the further a clause moves from a named thing that happened to one person, the worse it does — and abstraction costs more than weak phrasing does. "We found out the hard way" is a cliché and still beat the generational-poverty framing by 22 points, because at least it stays personal.

The formula was never the lever. Specificity was. I spent sessions optimising structure when a single vague clause was costing more than any structural choice gained.

Finding 2 — The best structure depends on the subject

I had assumed one architecture would win everywhere. Tested head to head, on the same day, that turned out to be wrong.

The pattern underneath: a subject that resolves wants a resolution, and a subject that does not resolve is weakened by pretending it does.

Money problems have an answer — you learned something and now you can pass it on — so the three-beat structure ending in a resolution fits. Loss does not have an answer, and appending a hopeful third beat to it scored worse than simply stating the loss and stopping.

Forcing the money structure onto the grief subject capped it well below its ceiling, and vice versa. Matching structure to subject was worth more than polishing either structure.

Finding 3 — Widening the audience only helps when the subject is already universal

I write from a specific place — Puna, on Hawaii's Big Island — and I assumed naming it was always an asset.

Tested directly: on a money subject, replacing the place name with a collective "us" scored the same or better (97 versus 96). On a grief and place subject, the place name was clearly stronger.

Money is nationally relatable, so naming a small place narrows an audience that did not need narrowing. Loss rooted in a specific community is the opposite — the specificity is the substance, and generalising it removes the thing that made it land.

Finding 4 — Scale words are not free

Adding an explicit scale word to the final beat — "everyone", "every family", "a whole island" — reliably lifted mid-tier titles.

But bolting one onto an already-strong title that ended in personal commitment dropped it from 95 to 93. That title's strength was the smallness of the ending. Making it bigger made it vaguer.

So scale is a fix for a weak ending, not an upgrade to a strong one. Test the closing clause both ways rather than assuming more reach is better.

Finding 5 — 97 was a hard ceiling

After hitting 97 three times I tried to break into the high nineties deliberately: six word-level variants across the two best titles — tense changes, added specificity, synonym swaps, reframings.

Every single variant tied or scored lower. None exceeded 97.

The tool's own diagnostic was that the format category itself was capped, that further word-level surgery was being penalised rather than rewarded, and that a higher score would require a structurally different hook — a question, a curiosity gap, a number-led opening — rather than more editing.

That is a genuinely useful negative result. It means that once a title in a known-good structure hits its ceiling, additional revision has negative expected value. The next gain comes from trying a different structure, not from another pass at this one.

What I actually changed

• Draft the middle clause first. It carries more score than the structure around it.
• Never ship a vague or metaphorical consequence. Name the specific damage.
• Never write the sociological version. "Kept my family poor for generations" is an essay thesis. "Nearly wrecked my family's finances" is something that happened.
• Pick the structure from the subject, not from habit.
• Stop editing at the ceiling. Change the structure or ship it.

An operational warning

One practical note for anyone testing this way: the AI Coach would auto-continue into follow-up work I had not asked for — including generating thumbnails, which consumed real credits. Roughly thirty credits went on unrequested images in one session before I caught it.

Say "titles only, no images" explicitly at the start, and watch for it proposing next steps as though you had already agreed.

This is one channel's data against one scoring model. The specificity principle should generalise because it is about concreteness rather than phrasing; the specific templates were tuned to my subjects and may not transfer. Test rather than copy.