Deep Reads: A Label Is Not an Argument
“Most companies invoking the word flywheel this year have the data and not the loop.”
0:00 / 9:38
Deep Reads: A Label Is Not an Argument
This week's briefing follows one question: what has to be true underneath a word for the word to mean anything? Three September reads take three words the insights industry says constantly — flywheel, job, creative impact — and each one turns into a test you can run against a claim instead of a label you can repeat. None of the three is a new study. All three are arguments from people who watch this field closely, which is what makes them useful as tests and unreliable as evidence.
Test the loop, not the label
Greenbook published The Flywheel Part 1: What Actually Compounds in the Insights Industry on September 18, 2026. The piece first ran on the Insight Innovation Ventures Substack and was written for investors sizing up AI-era companies; Greenbook reprints it with an editor's note that frames it as a ten-question audit. It is the first of three parts. 1
The essay's starting claim is that every insights company competing on "better AI" is competing for an advantage that expires in months, because models commoditize. What lasts longer is a loop: usage generates proprietary signal, the signal improves the judgment, better judgment produces better decisions, and better decisions generate more usage. The essay is blunt about what that does not include. A panel is not automatically a flywheel. A dashboard is not automatically a flywheel. An AI feature bolted onto a legacy platform is not automatically a flywheel.
Most of the argument goes into separating three loops the industry mixes into one. The data loop is the easiest to build and the easiest to overstate: a panel can collect a million new observations a month and still not get better at making client decisions, unless something downstream is learning from them. The product or workflow loop closes inside the product, because usage generates correction signal and query patterns. The decision-outcome loop is the rarest, because it needs something almost no insights company has — reliable access to what happened after the recommendation was delivered. Did the client act on it, what did they do instead, what result followed, and was the recommendation the cause of that result or would it have happened anyway.
That is where the essay puts the bottleneck. The rarest asset in the category is not raw proprietary data; it is validated decision-outcome data, tied back to a governed decision context. 1 The model layer is also the wrong place to look for the advantage: frontier-model access commoditizes, a fine-tuned domain model is not a moat, retrieval over a generic corpus is not a moat, and an interface that lets agents query a system directly is a channel rather than a compounding loop.
The counterfactual test is the one to remember: if a well-funded competitor had your model and your public data tomorrow, would this company still be improving faster? If not, the essay says, there is a feature, and features get copied.
There is also a contract argument worth carrying into procurement. A supplier with genuinely unique data can destroy its own loop in the first contract it signs, by bundling training rights in with basic access and converting a compounding asset into a one-time payment. The suggested default is to separate inference access, analysis rights, fine-tuning rights, foundation-model training rights, and persistent-derivative rights into distinct, deliberately priced categories — and never to sell irrecoverable rights cheaply, blindly, or as part of a contract built for a different era.
Read it as an investor's framework reprinted for the industry rather than as an independent study, and it earns its place as a buying checklist. If your team buys research or tools, the audit is a set of questions to ask before the word "flywheel" settles an argument. If your team sells them, the harder question is the same one turned inward: which loop can you close that a competitor cannot see, buy, or reproduce?
A job needs energy for progress
Jobs-To-Be-Done is nearly thirty years old, and the second read asks whether AI makes it more relevant than ever. Jennifer Axen, who runs an insights and experience strategy consultancy, put that question to two people who shaped the method: Taddy Hall, who worked with Clayton Christensen to develop and validate the theory and later co-authored Competing Against Luck, and Chris Spiek, who spent years developing and teaching JTBD alongside Bob Moesta at the Re-Wired Group. 2 The article is an e-newsletter methodology piece built from those two interviews, not new research.
Hall's starting point is causality. The industry is well supplied with tools for who people are, what they think, and how they behave, and those do not tell you what caused a particular choice. What predicts whether somebody pulls a product into their life, he says, is the extent to which it enables desired progress and fulfils an aspiration in a given circumstance. The operator word there is progress: JTBD looks for circumstances where somebody has enough motivation to change how they currently solve a problem.
That is why the interview is described as an investigation rather than a conversation. The researcher reconstructs one real decision in detail — what was happening, what had stopped working, when the idea of change first appeared, which alternatives were considered, who else was involved, and what finally tipped it. Hall describes this as recovering the raw data behind the numbers, and the raw data, he says, has the shape of a story. He is equally clear that asking someone why they acted and treating the answer as truth does not work: people recall what happened and how they felt far better than they explain their own motives. "Respondents don't give you the insights. That's our job. They give you the raw material of insights."
The piece also names a failure mode. Hall calls it "job washing": taking an existing need, observation, or persona, putting JTBD language around it, and declaring the job identified. The test it fails is energy. Most people need to exercise more, save more, or sleep more, and that does not mean they will do any of it tomorrow. The awkward "to be done" phrasing is doing work — it signals that someone is spending energy to resolve a trade-off or pursue an aspiration.
Spiek supplies the boundary. JTBD does not replace behavioural data or quantitative research; its value is explaining what caused behaviour to change, including small changes such as a customer suddenly adopting a feature they had long ignored. Hall adds the innovation use: the method can reveal the full portfolio of things people value, which is how a category changes the basis of competition instead of competing on the same narrow set of benefits.
Then comes the AI question. Spiek already uses AI to analyse JTBD interviews, identify forces and timelines, and get feedback on his own interviewing, and AI has accelerated product development at the same time. That produces his new problem: "I can build anything. I can ship anything very, very quickly. I've got to figure out what to go build." As building gets cheaper, knowing what is worth building gets more valuable. Hall's limit is the sharper one: "These are large language models. They're not deep emotion models." Tone, hesitation, enthusiasm, anxiety, and contradiction are often exactly where the insight lives, and those are the parts that do not survive a summary.
Two audiences, two verdicts
The third read starts with a number. Research by INSEAD, Jellyfish and WARC put 480 winners of the Creative Effectiveness Lions Award in front of human judges and language models, and found zero correlation between how the two audiences ranked them. 3 Jack Miles, a director at the London insight agency Shape Insight, uses that finding and the rest of WARC's post-Cannes reading of creative effectiveness to list where insight teams have to change. It is an industry-voices column rather than an empirical study, and parts of the research it cites sit behind WARC's registration wall.
The first consequence is a second audience for creative work. Generative engine optimization — structuring content so AI search engines and chatbots can find and cite it — is now part of how work gets seen, and a campaign can move people while staying invisible to the machines that increasingly mediate attention. Segmentation and target-customer work has to hold for both audiences.
The same research found two things driving creative impact for human and machine audiences alike: distinctive brand assets — the colours, shapes, taglines, and sounds that let people recognise a brand without its name — and category entry points. Miles points at the gap on our side of the table: an insights team will at best know how to interpret the standard grid of those assets, and very little research innovation has gone into creating or measuring them, including subtle ones such as colour and sound.
He also argues for joining things that are usually measured apart: brand and product research, which often sit under separate budgets, and brand and customer experience, where the brand is the promise and the experience is the reality. He quotes Coinbase's chief marketing officer, Catherine Ferdon, saying that brand is product, product is brand, and that the two are working in service of each other. The practical version is to stop treating customer experience as a satisfaction score and start linking friction, touchpoints, and delivery back to brand metrics.
The last item on the list is the smallest and the strangest. Research in System1's e-book The Creative Dividend found that the most surprising ads deliver 4.5 times more fame effects than the dullest ones, outperforming even happy ads on that measure — and a moment of surprise lasts roughly 1/125th of a second. 3 That is far shorter than the moments insight teams are used to studying, which is the point: the to-do list ends up being less about new tools than about what gets measured, and how finely.
What the claim has to earn
Three reads, three words, and the same move in each one. A flywheel claim should show you the loop that closes: if nobody knows what happened after the recommendation, the loop is open. A job should show you the energy for progress: if it is only a need, the language is decoration. A creative claim should show you which audience it landed with, because human and machine audiences have just demonstrated how differently they judge.
One thing all three reads share is worth stating plainly. None of them is a study. They are arguments from people inside this industry, which is what makes them useful as tests and unreliable as evidence.
Before you repeat one of these words in a meeting, ask what would have to be true for it to hold — then ask who would know. The label is cheap. The test is the work.
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.