TikTok Captions Don't Matter — We Analyzed 40 Viral Posts to Find What Actually Does

We built a content scoring tool. To make it accurate, we had to collect real data — viral posts, low-performing posts, and everything in between — and figure out what actually separates a 44M-view TikTok from a 300-view one.

We scored 40 posts across TikTok, Instagram Reels, Instagram captions, Facebook, and LinkedIn. We transcribed real videos. We ran backtests. Here's what we found.

Finding #1: TikTok captions are irrelevant

This was the biggest surprise. When we scored TikTok captions the same way we scored Instagram captions, our model was only 30% accurate. We kept predicting low performance on videos that hit millions of views.

The reason: TikTok captions are a label, not content. Nobody reads them before deciding to watch. The video autoplays. The first 3 seconds of audio — the verbal hook — is what drives watch time, shares, and reach.

Once we switched to scoring the transcript (the actual words spoken in the first 10-30 seconds), accuracy jumped to 88%.

Finding #2: There are 5 TikTok archetypes that actually go viral

After transcribing and analyzing 10+ viral TikToks, a clear pattern emerged. Nearly every high-performing video fell into one of five archetypes:

The lowest-performing videos we analyzed shared a different pattern: they opened with process instead of conflict. "So today I'm going to show you..." is a death sentence on TikTok.

Finding #3: The death patterns are specific and avoidable

We identified the exact phrases that consistently appeared in low-performing TikTok content:

What do these have in common? They're all process-first. They announce what's about to happen instead of creating a reason to keep watching. The audience hasn't been given anything to want yet.

These openers are all result-first or conflict-first. They create a gap the viewer needs to close. The watch time follows automatically.

Finding #4: Instagram Reels and TikTok are completely different platforms

This matters for agencies managing content across both. The viral archetypes that work on TikTok don't automatically translate to Reels — and using the wrong format kills performance on both.

Dimension	TikTok	Instagram Reels
Energy	Chaotic, raw, unpolished	Cinematic, produced, intentional
Best opener	Conflict or hot take	Result or transformation
Trust signal	Personal story + authenticity	Visual quality + journey narrative
CTA style	Implicit ("follow for part 2")	Explicit save/share request
Death pattern	Process-first opener	Chronological backstory

The highest-performing Reels we analyzed opened with the result first — then went back to explain how it happened. The lowest-performing started from the beginning and worked forward chronologically. Audiences don't have patience for buildup anymore.

Finding #5: Facebook algorithm actively suppresses spam signals

Facebook and Instagram have very different spam detection. Facebook aggressively suppresses posts with:

• All-caps phrases (4+ capitalized words)
• Excessive exclamation marks (3+ clusters)
• Clickbait framing ("doctors don't want you to know")
• Promo spam openers ("🔥 FLASH SALE 🔥 50% OFF!!")

In our dataset, a flash sale post with these signals scored high on value markers (discount language, urgency, CTA) but got suppressed in real distribution. Our early scorer missed this entirely until we hard-capped scores for posts with 4+ spam signals.

Facebook's highest-performing content is short, emotionally resonant, and debate-worthy. "Agree or disagree: most people overcomplicate their morning routine." That format drives more reach than any promotion ever will.

Finding #6: Series content plays by different rules

One thing that broke our initial model: series content. "Welcome back to Automating My Life with Python, day 30" scored poorly on hook strength — because we were penalizing "generic openers." But videos with that format routinely hit hundreds of thousands of views.

The reason: series content isn't trying to hook new viewers, it's serving an existing audience. The engagement mechanic is loyalty, not curiosity. Save rate matters more than like rate. Depth and specificity matter more than a compelling opener.

Once we added series detection (looking for "Part X", "Day X of", "Welcome back to", "Episode X"), we stopped penalizing a format that's working exactly as intended.

The scoring accuracy results

After incorporating all of these findings into a scoring engine, here's where we landed on a 40-post backtest:

Platform	Before tuning	After tuning
LinkedIn	80%	80% ✅
Instagram captions	60%	100% ✅
Instagram Reels	—	100% ✅
TikTok	30%	88% ✅
Facebook	30%	~80% ✅

The TikTok jump from 30% to 88% came entirely from switching to transcript-based scoring. The Instagram caption jump from 60% to 100% came from adding spam detection and a content density minimum. LinkedIn was already the most consistent — the platform's native format (storytelling + structured framework + CTA) maps cleanly to rule-based detection.

What this means for your content

Most content fails before anyone reads the caption. The platform decides in the first 3 seconds whether to push your content or bury it — and that decision is based on early engagement signals driven entirely by your hook.

The good news: hooks are learnable. The patterns are specific. The death patterns are avoidable. You don't need to go viral by accident — you can engineer content that has a legitimate shot.