Turning 42 Public Sales Videos into a Competitive Intelligence Package

I have built or rebuilt solo competitive-intelligence functions three times, twice in mobile robotics and warehouse automation. In my industry, information untainted by my competitors’ own marketing spin is unusually difficult to come by: nobody reviews their warehouse robots on G2, systems run from a few million to $20 or $50 million, and competitor pricing is almost never public. You can't buy a competitor's system to see how it works, you can’t mystery-shop when the entire buyer pool is mega-brands everyone has heard of, and you can't sit in on their demos.

One of our sales engineers sent me a link: a UK sales rep for AutoStore, a major competitor, had built a YouTube channel for his prospects. Forty-two short videos across three playlists: how to design an AutoStore, a line-by-line walk through the ROI case, and a series rebutting competitor attacks. Each ran three to five minutes and had about a hundred views. It was clearly meant to keep deals warm between calls, and it was good selling. He’s great at his job and kudos to him for being proactive in creating his own resources to help his buyers. But the channel was also his whole pitch, recorded.

The tab sat open in my browser for about five months.

The videos added up to about three hours of total runtime, but analyzing them properly is not three hours. Watching with a notebook and a fine-toothed comb, rewatching, cross-referencing claims across videos and verifying numbers is three or four days at least. A team of one can't disappear for most of a week on a speculative source, so the project kept losing to everything else.

The real problem was the ratio of effort to uncertain reward, which kills projects like this before they start. I needed a way to extract intelligence from the whole channel, not a summary of it, in time I could actually spare, without trusting anything I hadn't verified.

I treated the AI like an enthusiastic, tireless and extremely naive junior analyst: fast, useful, and never to be believed without checking.

• Transcripts, in batches. I pulled YouTube's auto-generated transcripts for all 42 videos and fed them to the AI four to six at a time, organised roughly by playlist. I skimmed each batch first so I could aim the prompt: if five videos were about pricing, go deep on pricing and don't hunt for robot specs.
• Structured extraction, not summarization. If you ask for a summary you get a summary. My prompts asked the AI to hunt for specific information: cost per unit, quoted exactly, no paraphrase, no inference; claims repeated across videos; contradictions between videos; how the rep answers competitor attacks.
• Cross-validation as the method. Outputs went into a knowledge base alongside the transcripts. After each batch I asked whether the new material confirmed or contradicted earlier batches. A claim that appears almost word for word in different videos, recorded months apart, is a trained talking point, and that is a window into sales strategy.
• Synthesis last. After nine batched sessions I ran synthesis passes to roll everything up into a single report.
• The judgement stayed mine. The AI flagged that the rep says 80% of customers expand within three years. It did not decide what that meant. I made that inference, and I decided which findings mattered in a live deal.

What broke:

• Rate limits. Transcripts are token-heavy, and I kept hitting the limit on a $20 plan. Out of frustration, when I hit my limits on Claude, I started running the same batches through NotebookLM, then asked each tool what the other had missed or got wrong. The workaround became the better method, because I was running two independent AI analyses in parallel.
• Hallucinated quotes. The AI produced plausible quotes, attributed to specific videos, that the rep never said. I caught them because I had skimmed the transcripts first: when a quote sounded too good, I went back to the video. Every key specific was checked against the source.

• YouTube auto-generated transcripts (42 public videos)
• Claude for batched extraction and analysis
• NotebookLM (since renamed Gemini Notebook) for parallel analysis and cross-critique
• A knowledge base holding transcripts and outputs together, for cross-referencing
• Manual verification against the original videos
• The finished report was stored in our CI platform and distributed to the senior leadership team

The method is tool-agnostic. The prompts matter more than the model.

About four hours of my active time, spread over a few days, turned a five-month-old browser tab into a 30 to 40 page intelligence report and a set of sales tools.

What the chain of findings showed:

• Pricing. A line-by-line breakdown for a hypothetical, but typical, mid-sized system in the €5 million range: hardware, implementation, risk, floor preparation, ongoing costs. In this industry that essentially never comes from public sources.
• An expansion signal. "80% of customers expand within three years" appeared four times across three playlists. My read: if 80% expand, 80% of initial systems are undersized, and the ROI story, built on an 18-month payback, includes none of that expansion capex.
• A concentration risk. 47% of the ROI in the independent study the rep walks through comes from footprint reduction. For a customer who isn't space-constrained, about half the value proposition disappears, and the rest is common to any goods-to-person system.
• The rep revising his own study downward, on camera. Labour cost per picker from €45,000 to £25,000–30,000 in the UK; peak-preparation savings from about €500,000 to €400,000. High by 20%, in places by as much as a third.
• A second competitor for free. One whole playlist contained point-by-point responses to an attack document that appeared to come from another competitor. Strong rebuttals marked real strengths, vague ones marked possible weak spots, and concessions marked confirmed ones. The defensive response in the videos revealed at least as much as the attack.

Those became an objection-handling playbook (his script, where he is strong, where he struggles), discovery questions built from the competitor's own claims, counter-messaging that quotes their own facts back rather than FUD, a vulnerability matrix, and battlecard updates.

The speed helped, but the real value was that the analysis happened at all. And you don’t need to hope you randomly find one plucky sales rep‘s YouTube channel either. Every competitive practitioner has some pile of unstructured information where, before these AI tools existed, the juice just wasn’t worth the squeeze. For example: say your competitor’s CEO does podcast interviews and speaks at conferences. When that CEO tells the same story once to a room full of lawyers, and later to a room full of investors, and later to a room full of engineers, what changes? What points do they add or leave out? What gets emphasized differently?