8,758
Videos
Analyzed
Analyzed
An empirical look at 8,758 AI-themed videos posted to TikTok and Instagram between February and May 2026. We examine which structural choices - duration, pacing, vocabulary, hook, platform - correlate with engagement, and which do not.
Videos opening with a self-introduction ("Hi, my name is…", "Hello, I'm…") record the lowest mean engagement rate of any hook category we classified: 4.33% vs. 5.17% across all hooks. The instinctive opener that founders use on stage is the same one the algorithm appears to penalise on short-form video.
n = 235 videos · 4.33% ER · −16% vs baselineThe relationship across pace buckets is monotonic. Videos delivered at 210+ words per minute record a mean ER of 5.57%; videos under 90 wpm record 3.52%. A measured, deliberate cadence - what we might call the "boardroom voice" - appears to underperform a faster, conversational one on this dataset.
n = 2,447 (fast) vs 68 (slow) · +58% ER gapVideos with captions under 50 characters record a mean ER of 6.69%. Captions over 500 characters record 4.12%. The relationship is monotonic across all five buckets we tested. One interpretation: long captions read as paid or promotional, even when they are not; concise captions read as native.
n = 533 (short) vs 1,472 (long) · +62% ER gapMarch recorded the highest video volume; April recorded the highest total views (574M), suggesting some of the March content compounded into April distribution. May figures are partial (data through May 15).
The Oriane Shadow Reach Layer: 483 videos in this dataset mention major AI labs (OpenAI, Anthropic, Google AI) only in spoken audio - never in a caption or hashtag. Those 483 videos accounted for 70.9 million views, none of which would surface in a text-based listening tool such as Brandwatch, Meltwater or Talkwalker. The shadow layer is not the headline of this report, but it is the methodological note worth flagging.
For OpenAI / ChatGPT specifically: 191 spoken-only videos accounting for 36.2M views. For Anthropic / Claude: 172 spoken-only videos · 24.0M views. Text-based competitive monitoring would miss roughly a third of the surface area on either brand.
The top 1% of videos in this dataset (n=65) post a mean engagement rate of 34.4% against a dataset baseline of 5.17%. They share three structural traits in common.
The platform mix is the largest single separator: 86% of top-1% videos sit on TikTok versus 44% of the rest. Pace, jargon density, and word count are nearly identical between cohorts. The data suggests that what founders are saying matters less than where they choose to say it.
We classified the opening words of every English-language video in the dataset into nine hook archetypes. The performance gap between the best- and worst-performing hook is roughly 35 percentage points in mean ER. The most common opener used by founders ("self-introduction") sits in the bottom third.
Hooks framed around the viewer ("you") outperform hooks framed around the founder ("I built…", "Hi, my name is…") by 35–39% in mean ER. "The problem is…" - a common framing in pitch coaching - sits in the bottom third. The data suggests that what works in a pitch room does not transfer directly to short-form video.
Real openings, real performance, pulled directly from the dataset's outliers.
The pattern across the highest-performing openers is consistent: they make a claim that contradicts a conventional assumption, name the viewer rather than the founder, or promise something immediate. Founders trained for credentialed introductions appear to be operating with a different frame than the platform rewards.
Among the three opener phrasings we tested directly: "Watch this" records the highest mean ER (7.64%), ahead of "The problem is" (5.46%) and "We built" (5.72%). The gap between the best and worst of these three is roughly 40%.
Three structural variables - video duration, delivery speed (words per minute), and content theme - account for a substantial share of the engagement variance in this dataset. We examine each in turn.
The "shorter is always better" heuristic does not hold in this dataset. The 90s–3min band records the strongest mean ER (6.05%); sub-30-second clips underperform (4.19% at 15-30s). One reading: audiences for technical content appear to self-select for longer-form engagement.
For reference: conversational English averages around 130 wpm and audiobooks target roughly 150. The top-performing AI founder videos in this dataset run at 200–235 wpm - closer to a podcast guest in flow than a keynote register. A measured cadence with deliberate pauses appears to underperform on short-form video, though we cannot rule out that "fast pace" is itself a proxy for higher production energy.
A note on the "AI agents" theme: it has the lowest mean ER of any major theme we tested in this dataset (4.31%). One interpretation is audience fatigue around abstract framing; another is that "agents" content tends to be more conceptual and less product-specific, and concrete demos appear to engage more reliably. We flag this as a finding worth replicating before drawing conclusions.
If we aggregate the top-decile traits into a single composite, the highest-engagement video in this dataset is on TikTok, runs roughly 110 seconds, opens with a contrarian or instructive line ("stop doing X", "watch this"), is delivered at 210+ words per minute, uses a caption under 50 characters, and closes without an explicit call to action.
Both signals point in the same direction: content that looks and reads less like marketing performs better. There is an obvious causal story (audiences and algorithms both prefer non-promotional content) but also a confounding one (creators who happen to use shorter captions and no CTAs may simply be making different kinds of content). The data shows the pattern; the interpretation is contestable.
We tagged the spoken transcript of every English-language video for technical jargon - agentic, multimodal, transformer, embeddings, inference, RAG, fine-tuning, latency, and similar terms - and counted occurrences. The relationship between jargon density and engagement is non-monotonic: a small amount is neutral or mildly positive; higher volumes are associated with lower engagement.
One technical term outperforms zero (5.42% vs 5.15%). Two or three already sit below baseline. From four upwards, the relationship turns sharply negative. A useful working principle drawn from the data: introduce at most one technical concept per video, and translate it explicitly. Each additional term beyond the first correlates with measurably lower engagement.
Videos that contain explicit translation phrases ("think of it like…", "in plain English…", "this basically means…") outperform videos that don't by roughly 18% in mean ER. This is one of the larger effect sizes we recorded for a single linguistic variable.
"AI washing" is real but solvable. The data shows audiences don't punish founders for using AI terminology - they punish founders for using AI terminology without explaining it. Say "agentic" once, then say what it means. Say it five times, lose the audience.
When AI founders make videos, they don't operate in a vacuum - they reference the platforms shaping the conversation. Here is the share of voice across the dataset.
Two labs dominate the named-mention conversation in this dataset: Anthropic/Claude (636 videos) and OpenAI/ChatGPT (624). xAI/Grok records 51 mentions despite high mainstream visibility for its leadership. The substantive AI conversation on TikTok and Instagram, by volume of mentions, is largely a conversation about two companies.
Shadow reach in this dataset: 191 of the 624 OpenAI / ChatGPT mentions appear only in spoken audio - the brand is named in the video but never written in the caption, hashtag, or @mention. For Anthropic / Claude: 172 of 636 mentions are spoken-only.
For a brand running competitive monitoring exclusively via caption-based listening tools, this represents roughly a third of the actual mention surface area in the dataset. The implication is methodological: any market-share or share-of-voice estimate based on text scraping alone will be a meaningful undercount for AI brands.
The dataset splits roughly 55 / 45 between Instagram and TikTok. The engagement-rate gap between the two platforms is the largest single performance delta in this analysis - material enough that conclusions about content type or hook archetype need to be read in conjunction with platform.
TikTok now rewards substantial duration for AI/founder content. The 90s–5min range records the strongest mean ER on the platform - 7-12x longer than the 15-second clips TikTok built its reputation on. The platform's audience for technical content does not appear to behave like a short-attention-span audience.
Instagram's engagement curve by duration is flatter than TikTok's but skews slightly more toward longer formats. The Reels-as-short-form-clone era appears to be over for technical content in this dataset - sub-15-second clips meaningfully underperform 90s+ videos on both platforms.
The median duration among top-1% TikTok videos is 100 seconds. The most-engaging pattern we observed is a fast opening hook, a concrete demonstration of 60–90 seconds, and a close without an explicit CTA. Talking-head footage layered with screen recording is the dominant winning format.
TikTok mean ER 7.32% · 86% of top 1% are on TikTokInstagram accounts for 40% more total views than TikTok (1.2B vs 869M) on a similar content volume, but at roughly half the mean engagement rate. The data is consistent with using Instagram as a top-of-funnel surface - for awareness and broader distribution - while running deeper engagement and community-building on TikTok.
IG: 1.2B views · 3.42% mean ERThe platforms behave less like substitutes than like distinct channels with distinct functions. TikTok appears to be the better channel for comments, saves, and audience depth; Instagram appears to be the better channel for total reach. The same video, cross-posted without adaptation, will likely underperform its potential on at least one of the two.
3,764 unique creators appear in this dataset. The follower-tier breakdown shows a non-monotonic relationship: mid-tier accounts (100K–1M followers) outperform macro accounts (1M+) on mean engagement, and a meaningful share of the highest-engagement videos come from accounts with under 50K followers.
Mean engagement decreases monotonically with follower size above 10K, with the sub-1K tier behaving anomalously (likely a distribution effect - very small accounts struggle to reach a meaningful audience at all). The 100K–1M tier records the strongest mean ER above the 10K floor, which on this evidence is the band worth examining most closely for partnerships.
| Creator | Platform | Videos | Total Views | Followers | Avg ER |
|---|---|---|---|---|---|
| @amazonwebservices | IG | 9 | 49.3M | 786K | 0.80% |
| @dope.motions | IG | 4 | 22.4M | 779K | 2.71% |
| @tiffintech | IG | 13 | 12.9M | 626K | 5.97% |
| @nytimes | IG | 4 | 12.9M | 20.0M | 5.06% |
| @ray_fu | TT | 24 | 12.3M | 126K | 7.32% |
| @mytechceo | IG | 7 | 10.2M | 252K | 4.62% |
| @vaibhavsisinty | IG | 25 | 9.9M | 1.6M | 3.50% |
| @thevamshikurapati | IG | 15 | 9.6M | 912K | 4.48% |
| @pikacodes | TT | 2 | 7.1M | 192K | 3.26% |
| @manojsaru | IG | 2 | 6.2M | 2.1M | 9.24% |
| @socho.abhi | IG | 9 | 5.6M | 307K | 6.50% |
| @brycent | IG | 65 | 6.0M | 97K | 3.95% |
The top of the reach leaderboard is largely populated by accounts that cover AI and startup content (creators, news brands, AI commentators) rather than by founders themselves. Founder accounts tend to appear lower on the reach list but record higher mean engagement rates, consistent with the broader tier finding.
High engagement, mid-size following, multiple videos in our window. These are the creators a startup should be paying attention to before everyone else does.
| Creator | Platform | Videos | Total Views | Followers | Avg ER |
|---|---|---|---|---|---|
| @realestrhxtt | TT | 3 | 1.1M | 132K | 40.77% |
| @adivunsolicited | TT | 3 | 328K | 221K | 26.19% |
| @thedreydossier | TT | 3 | 291K | 484K | 19.34% |
| @thotline_news | TT | 4 | 138K | 98K | 19.12% |
| @chriswinfield | IG | 14 | 471K | 330K | 18.69% |
| @hemu_rahman | TT | 12 | 2.3M | 118K | 18.03% |
| @davejorgenson | TT | 3 | 652K | 175K | 16.14% |
| @jthefrog0 | TT | 6 | 4.0M | 327K | 15.34% |
| @heyroyall | TT | 3 | 2.6M | 187K | 14.15% |
| @longliveai | IG | 17 | 11.8M | 130K | 12.84% |
All but one of the top-10 mid-tier creators in this view are on TikTok. Engagement rates of 15-40% are unusually high and likely reflect a tight content-audience fit rather than something replicable through paid amplification - these accounts have built their audiences around AI/founder commentary specifically.
The AI brands that will define 2026 are not the ones with the biggest social agencies. They are the ones that figured out which 100-creator partnership map gets them inside the loops where AI is actually being talked about.
Each item below corresponds to a specific finding earlier in the report. The framing is observational, not prescriptive: these are the patterns we found in the data, with the caveat that any single video may succeed by ignoring all of them.
Videos opening with "Hi, my name is…" or similar self-identification record a mean ER of 4.33% - 16% below the dataset average and 26% below the best-performing hook archetype (direct address). Founders who recognise their viewers can read their account handle and bio appear to use that screen space for something else: a contrarian claim, a piece of news, or a direct address to the viewer.
Self-intro hooks: 4.33% ER · Direct address: 5.83% ER · +35%The specific phrase "Watch this" appears as an opener in 312 videos and records a mean ER of 7.64% - roughly 48% above the dataset average. The variant "We built…" records 5.72%. A reasonable interpretation: "watch this" promises immediate value to the viewer; "we built" centres the creator. The first frame appears to be where this distinction matters most.
"Watch this": 7.64% ER vs "We built": 5.72% ER · +34%The relationship between words-per-minute and ER is monotonic across the six buckets we tested. Videos delivered at 210+ wpm record 5.57% mean ER; videos under 90 wpm record 3.52%. The top 1% of videos average 196 wpm - closer to podcast tempo than to a keynote delivery. Whether this reflects audience preference, algorithmic preference, or a correlated variable (energy, editing) is unclear from the data.
210+ wpm: 5.57% ER · Under 90 wpm: 3.52% ER · +58%On both platforms, videos in the 90s–3min duration band record higher mean ER than sub-30-second clips. On TikTok, the 3–5min band performs best (8.78% ER). The "shorter is always better" heuristic does not hold in this dataset for AI/founder content; one plausible reading is that audiences for technical material self-select for longer engagement.
90s–3min: 6.05% ER · 15-30s: 4.19% ER · +44%Videos using exactly one technical term ("agentic", "RAG", "multimodal", "fine-tuning") slightly outperform videos with none (5.42% vs 5.15%). Each additional term beyond the first correlates with a decrease in mean ER, reaching 2.37% at seven or more terms. A useful proxy: introduce the term once, then translate it with a phrase like "think of it like…" - videos using two or more such translation phrases record 6.02% ER vs 5.10% for videos with none.
1 term + translation language: optimal · 7+ terms: 2.37% ERVideos without an explicit CTA record 5.24% mean ER; videos with one record 4.90%. The phrase "link in bio" carries the largest individual penalty (4.08% ER, n=412). The causal direction is contestable - content that includes a CTA may also tend to be more transactional in tone overall - but the correlation across phrasings is consistent.
No CTA: 5.24% ER · "Link in bio": 4.08% ER · −22%TikTok records 2.1x the mean engagement rate of Instagram (7.32% vs 3.42%) in this dataset. Instagram, however, accounts for 40% more total views (1.2B vs 869M) across a similar volume of content. For founders, the implication is operational: the two platforms behave less like substitutes than like distinct channels with different functions - engagement on one, awareness on the other.
TT ER 7.32% · IG ER 3.42% · IG views +40% over TTAggregated across 8,758 videos, the patterns above point in one consistent direction. The format rewards directness, brevity in caption, depth in delivery, and an unscripted register. None of these are revelations on their own; the contribution of the dataset is that all four show up together, with reasonable sample sizes, across two platforms and three months of content.