Investigation history

What is worth doing for YouTube Channel

One opportunity survived scrutiny, and it reframes the objective: the highest-leverage thing to do with a YouTube channel in 2026 is to engineer it as a citation source for AI answer engines rather than as an audience-growth funnel. The supporting evidence converges from three independent datasets. BrightEdge (May 2024–Sep 2025) finds YouTube is the single most-cited domain in Google AI Overviews at 29.5% share and cited ~200x more than any other video platform. Ahrefs' 75,000-brand analysis finds YouTube mentions correlate with AI visibility at r=0.737 — stronger than backlinks or domain authority. Otterly's ~100M-citation study supplies the actionable mechanism: citation frequency has near-zero correlation with views/likes/subscribers (r≈-0.03), 94% of citations go to long-form, and 78% of timestamped videos are cited repeatedly across 2–5 chapters. That combination is unusual: a distribution surface where structure beats audience size, so a small or brand-new channel can compete. Practically: how-to/tutorial and review/comparison query clusters, question-format chapter titles, clean human-checked transcripts, VideoObject schema on embed pages, and citations — not views — as the success metric. I deliberately tried to break this and found real limits, which is why the score is 65 and not higher. In Otterly's dataset all video/social sources are only ~5.5% of citations, so "200x" describes dominance of a small slice. The payoff is concentrated on Perplexity (38.7%) and Google AI surfaces, while ChatGPT cites YouTube just 4.4% and Gemini/Copilot ~0% — plausibly because Google self-preferences its own property, an advantage it can revoke. AI Overview citations click at roughly position-6 levels and 76.1% of cited URLs already rank top-10 organically, so citation may follow authority more than create it, and the r=0.737 figure is likely confounded by brand size. Two hypotheses died. The "AI-slop crackdown creates a compliance-service gap" thesis failed because YouTube explicitly called the July 2025 policy a minor clarification of long-standing rules against mass-produced/repetitive content — no large newly-demonetized cohort materialised. And the sharper commercial variant of the surviving thesis — that a productized video-AEO service is *unserved* — was weakened to 40%: dojoAI, iPullRank and others already publish detailed YouTube AEO playbooks, Otterly already sells YouTube citation tracking, the tracking category is crowded at $20–$3,000/mo with practitioners calling it snake oil, and I found no evidence of anyone actually paying for video-specific AEO. So the tactic is well-evidenced; the business around it is not yet. Net judgment: do this for your own channel or your own brand now, where the cost is marginal (chapters and transcripts on video you were making anyway) and the timing edge is real but closing. Be far more cautious about building a company on it until you observe a paying buyer — demand there is inferred from a 20x rise in "answer engine optimization" searches, not from transactions.

10 searches · 9 reads · 17 steps · $3.15

Find things worth doing in local AI

**One opportunity survived; two starting hypotheses did not.** I entered "local AI" expecting the gap to be at the runtime or model-quality layer. It wasn't — those layers are crowded and improving fast. The investigation converged instead on an unguarded seam in the artifact distribution chain. **What died.** The hypothesis that small local models' tool-calling unreliability blocks agentic use was **rejected**: llama.cpp auto-converts JSON Schema to GBNF grammar and vLLM/SGLang ship equivalents, so the mechanical remedy is already everywhere. Residual failures are per-model notation gaps (e.g. Gemma's non-JSON fc notation), which are a symptom of the artifact problem, not a standalone opportunity. The hypothesis that task-scoped quantization regression testing is a missing layer was **weakened to 0.30**: Unsloth runs 150+ KL-divergence sweeps per model family with published artifacts, a Jan 2026 arXiv paper does unified per-GGUF-format evaluation, and promptfoo/DeepEval already support Ollama-backed custom evals. Real need, credible incumbents, incremental. **What survived (0.72, adversarially checked).** The chat template — executable Jinja that runs on every inference call, between the user and the model — is validated by nobody. This produces one root cause with two faces. Quality: a leading publisher re-uploaded a single popular model **six times** to fix a `[::-1]` bug, an LM Studio/llama.cpp divergence, broken tool calling and mismatched BOS tokens that silently disabled speculative decoding — the original defect found by a Reddit user, and the publisher stating they believe rivals' quants are still broken. Security: a demonstrated backdoor vector, 90%→15% accuracy when triggered, dormant on benign input, hijacking agent tool-use across 3,868 episodes, generalising over 18 models, 7 families and 4 engines — invisible to prompt-injection guardrails because it renders *before* user input, and invisible to hub scanners because it is valid Jinja, not binary malware. **Why I believe the neglectedness claim.** Four falsification attempts found only adjacent coverage: JFrog matches keywords (`os`, `eval`, `__class__`) for code-execution SSTI on a CVE patched in 2024; hub scanning covers malware/pickle/secrets and still let 100+ malicious models through; Sigstore model-signing attests publisher identity rather than template fidelity, at 244 stars; promptfoo's own vulnerability database **prescribes** "automated analysis of `tokenizer.chat_template` to detect anomalous conditional logic" and then advises users to manually eyeball template strings. A prescription with no product behind it is the strongest neglectedness signal I found. **The honest caveats.** The oracle is weak — upstream templates are themselves often buggy at release (the Qwen3 defect originated upstream), and re-quantizers legitimately patch them, so "diverges from canonical" ≠ "malicious"; false-positive triage is the core product risk. There is no confirmed in-the-wild template backdoor, so security urgency is anticipatory. Monetisation is unclear: this most naturally lives as open-source infrastructure or gets absorbed by an existing scanner or the hub itself. And the person closest to the problem — a quant publisher — would reasonably argue the real fix is upstream authors shipping tested templates plus runtimes converging on one Jinja implementation. It clears the bar anyway because the dual-use mechanism pays off on quality hygiene even if no attack ever materialises, and because feasibility is unusually high: templates are small text files, analysis needs no GPU, and an ecosystem-wide scan is cheap enough for one person to run as a launch artifact.

14 searches · 11 reads · 21 steps · $5.03