Information Theory, Asymmetric Risk, and the CPG Growth Blind Spot

The Stanley Quencher nearly died on the shelf before becoming a $750 million breakout. Oatly skipped the shelf entirely. Both were reacting to the same blind spot. What follows is a mathematical account of that flaw, built from stationarity assumptions, Shannon entropy, and the feedback loops that make POS data lie.

Prologue: The Brand That Almost Missed Its Own Breakout

In 2019, by every metric a category-management system tracks, the 40oz Stanley Quencher was a failure. The product had been sitting in Stanley's catalog for years. An oversized, $45 stainless-steel cup designed for construction workers and outdoorsmen, marooned on shelves full of cheaper plastic alternatives. Turnover was slow. Days-of-supply was bloated. GMROI (gross margin return on inventory investment, basically "is this product earning its shelf?") was below the company's threshold. The brand had quietly stopped marketing it. Stanley's replenishment algorithms, doing what they were built to do, responded by trimming orders, lengthening reorder windows, and reallocating shelf space to products with better signals.

Four years later, that same SKU would anchor a brand reborn. Stanley revenue went from roughly $73 million in 2019 to over $750 million by 2023, on the back of a women-led word-of-mouth wave that started with a few bloggers in Utah and eventually broke through into mainstream social-media virality. None of that growth was visible to the brand's forecasting systems in time to capitalize on it.

The tumbler that should have been quietly euthanized in the next category review became one of the most successful consumer product launches of the decade.

The uncomfortable part is that the systems were doing their job. Stanley's near-miss wasn't a software bug. It was the predictable output of a mathematical framework the entire CPG industry runs on, and that framework has one fatal property: it cannot tell the difference between the early signal of a generational breakout and the random noise of a stagnant SKU.

Call this the Optimization Paradox. In the chase for lean operations, we have built supply chains that behave like low-pass filters. (A low-pass filter is the audio-engineering trick where you cut the treble and let only the deep bass through.) Apply it to demand and you get the same effect. Sharp, fast spikes get treated as noise. Only the slow-moving baseline reaches the people making decisions. For a stable brand selling 100,000 units a month with 2% standard deviation, that is a feature. For a brand on the cusp of a regime shift, it is a slow-motion catastrophe.

1. The Stationarity Assumption

To understand why Stanley's systems failed, look at what they assumed.

Almost every inventory model in production today rests on an assumption called stationarity. Formally, a demand process $\{D_t\}$ is stationary if its statistical properties (its mean $\mu$, its variance $\sigma^2$, and its autocorrelation structure) do not change over time:

$$D_t \sim \mathcal{N}(\mu, \sigma^2)$$

Translated: demand at time $t$ is normally distributed around a fixed mean $\mu$ with a fixed variance $\sigma^2$. Both numbers are constants. Day-to-day variation is just random noise wobbling around an unmoving center.

The assumption collapses the messy future into two numbers and a distribution. It lets you compute the Reorder Point (ROP), the inventory level that, when crossed, triggers a new order, using the textbook formula:

$$ROP = (\bar{d} \cdot L) + (z \cdot \sigma_L)$$

Where $\bar{d}$ is your average daily demand, $L$ is your lead time in days, $\sigma_L$ is the standard deviation of demand over that lead time, and $z$ is the service-level multiplier ($z = 1.65$ buys you 95% in-stock probability; $z = 2.33$ buys you 99%).

The first term, $\bar{d} \cdot L$, is the inventory you need to cover average demand during the lead time. The second term, $z \cdot \sigma_L$, is your safety stock, the buffer against randomness. Together, those two terms are the entirety of how the modern supply chain conceives of the future.

The formula assumes the mean is known and stable. Anything that deviates from the mean gets treated as $\sigma$, randomness to buffer against, never as signal to act on. There is no place in this equation for the possibility that $\mu$ itself is moving.

Which is, of course, what happened to Stanley. The 2019 system saw the Quencher's mean as low and flat. When demand started climbing, first via a small community of Utah-based bloggers and later through broader word-of-mouth, the system did not interpret it as "$\mu$ is shifting upward." It interpreted it as values in the high tail of the existing distribution. The textbook response to a tail observation is not to increase the forecast. It is to expect regression to the mean. So the system, behaving with perfect statistical discipline, would have ordered less in response to the spike. Spikes are by definition supposed to revert.

Statisticians call demand whose underlying distribution shifts over time non-stationary. Markets that are creating new categories (energy drinks in 2010, oat milk in 2018, hydration vessels in 2019, GLP-1-friendly snacks in 2024) are non-stationary by definition. The act of category creation is a shift in $\mu$. And the standard inventory models in deployed use are not built to register that shift.

To be fair to the math, more sophisticated tools do exist. Bayesian forecasters with hierarchical priors update their beliefs as new data arrives. State-space models like Kalman filters and dynamic linear models explicitly allow time-varying parameters. Change-point detection is its own subfield of statistics. The math, broadly, is up to the task.

The problem is that almost no deployed retail replenishment system uses any of this. The vast majority of CPG replenishment in 2026 still runs on moving averages, exponential smoothing, and rules of thumb that assume a stable distribution. And even where the better tools are deployed, they still need historical data on the new regime to parameterize themselves correctly, which by definition you do not have during the early phase of an unprecedented breakout. And even when the parameterization works, the bullwhip lag still applies, because the data has to physically climb the supply chain before it reaches any model.

The mean was shifting. The model was anchored to a past that no longer existed. In statistical terms, this isn't a forecasting error. It is mathematical censorship.

2. The Variance Amplification Ratio

The censorship gets worse as it travels upstream. Stanley's retailer doesn't see the demand signal once and react. The signal climbs a ladder: shopper to store, store to distributor, distributor to brand, brand to manufacturer, manufacturer to component supplier. Each rung has its own forecasting model, its own safety stock policy, its own batching cadence. At each rung, the signal gets louder and less informative.

This is the bullwhip effect, and it is well-measured. Lee, Padmanabhan, and Whang's seminal 1997 paper in Management Science formalized it as a Variance Amplification Ratio:

$$\text{Bullwhip} = \frac{\text{Var}(q)}{\text{Var}(D)}$$

Where $q$ is the order quantity placed upstream and $D$ is actual customer demand. Ratio of 1 means orders track demand perfectly. Ratio of 4 means orders fluctuate four times as violently as the demand that caused them. Empirical work on real CPG supply chains routinely measures ratios of 5 to 20.

Chen, Drezner, Ryan, and Simchi-Levi (also Management Science, 2000) later derived a clean closed-form lower bound for the specific case of moving-average forecasting with lead time. For a system with moving-average window $h$ and lead time $L$:

$$\frac{\text{Var}(q)}{\text{Var}(D)} \geq 1 + \frac{2L}{h} + \frac{2L^2}{h^2}$$

Two variables drive everything.

$L$ is your lead time. The gap between deciding to order and actually having product on hand. Domestic sourcing might give you a 4-week lead time. Sourcing components from China can stretch you to 8 to 12 weeks. Notice where it sits in the formula: in the numerator of both correction terms, and in the second term it shows up squared. Lead time does not make bullwhip worse linearly; it makes it worse on a curve. Doubling your lead time can roughly quadruple your variance amplification.

$h$ is your smoothing window, the part most operators have never had explained to them clearly. It is the number of past periods you average together to forecast the future. A 4-week moving average means $h = 4$. A 26-week moving average (a common default) means $h = 26$. The bigger you make $h$, the more past data folds into your forecast, which makes things feel more stable: any single weird week gets diluted by the normal weeks around it. A 26-week average barely twitches when one week comes in hot. A 4-week average reacts much faster but lives with more noise.

Naturally, then, managers want to crank $h$ up. The formula seems to reward them: $h$ in the denominator, bigger $h$ shrinks the bullwhip ratio. The trap is that a bigger smoothing window also means a slower forecast. Average 26 weeks of data and an emerging trend in the last two weeks barely registers. By the time the trend shows up in your numbers, the breakout has already happened and you have already missed it. You bought stability by paying with lag.

There is no clean win in this formula. You can have a stable forecast or a fast forecast. You cannot have both. A deeper finding from this body of literature, originating with Lee et al. and reinforced ever since: even when every actor in the chain behaves rationally, the bullwhip still emerges. It is not a behavioral failure. It is a structural property of the channel. The supply chain manufactures variance the way a refinery manufactures heat, as an unavoidable byproduct of running.

3. The Asymmetry of Ruin

So far we have two findings. The system can't see growth. Any error it does make gets amplified upstream. The next question: what does each error actually cost?

The textbook answer comes from the Newsvendor Model and its Critical Ratio:

$$CR = \frac{C_u}{C_u + C_o}$$

Where $C_u$ is the cost of understocking (what you lose if you don't have enough) and $C_o$ is the cost of overstocking (what you lose if you have too much). The ratio gives you your optimal service level. If $C_u = C_o$, your CR is $0.5$ and you should aim for a 50% in-stock rate. If $C_u$ towers over $C_o$, your CR climbs toward 1 and you should aim for near-perfect availability.

In standard textbooks, $C_u$ is calculated as the lost margin on the unit you didn't sell. Make $1 in profit per can, lose $1 in a stockout. Simple.

This is dangerously incomplete for a high-growth CPG brand. The stockout is not a one-shot transactional loss. It cascades into a chain of correlated downstream losses that can run an order of magnitude or two larger than the lost margin.

Take a brand like Liquid Death and itemize what $C_u$ actually contains.

The first piece is the Halo Effect Loss. ICSC and other retail-analytics groups consistently document a halo effect in which physical retail presence drives a meaningful lift in online sales for the same brand within the same trade area. The lift varies by category and methodology but is routinely measured in the high single digits to low double digits. So the shelf is not just a point of sale. It is also a geo-targeted impression you have effectively pre-paid for through your slotting fees. When the shelf goes empty, that impression stream goes dark, and you start paying full freight to re-acquire those customers through paid digital.

The second piece is the Substitution Loss. A shopper who came for Liquid Death and found the shelf empty rarely walks out empty-handed. They buy a competitor. Critically, that substitution is not symmetric: a non-trivial fraction of substituters become loyal to the competitor and never come back. What seems like a lost $1.99 sale translates to a lost customer lifetime value, which for premium beverage brands typically lives in the tens to low hundreds of dollars.

The third is the Velocity Rank Loss. Category managers evaluate brands on sales per linear foot per week, a velocity measure. Stockouts produce zero-sale days that pull your trailing-12-week velocity down. Plainly, the stockout you had in March can cost you shelf space in November.

The fourth, and the most lethal, is Algorithmic Demand Suppression. Retailer auto-replenishment systems learn from POS data. Empty shelf, zero POS, algorithm reads it as low demand and orders less next time. This is the Sriracha Trap, and Section 6 has its own treatment. For now, just note that today's stockout directly causes a smaller order tomorrow, which causes another stockout the day after.

What looks like a $1.99 stockout on the dashboard is actually a five-step chain reaction that ends in a category review meeting eight months later, with someone you've never met deciding your facings get cut.

When you stack all four honestly, $C_u$ for a high-growth brand probably lives somewhere between 10× and 50× the lost margin, depending on category, brand life-cycle stage, and how heavily you weight the long-tail effects. (To be transparent: this is an analytical estimate, not a published empirical figure. The literature documents the directional cascade rather than putting a single multiplier on it.) Meanwhile $C_o$, the cost of overstocking, is roughly the holding cost on the unsold unit. Maybe 2 to 5% of unit cost per month.

Plug an illustrative version into the Critical Ratio. If $C_u = 30$ and $C_o = 0.5$:

$$CR = \frac{30}{30 + 0.5} \approx 0.984$$

The point of the math isn't the specific 98.4% target. The point is the structural implication, the reality that the optimal policy involves ordering far more aggressively, far earlier, on far weaker signals than any traditional replenishment system would tolerate.

The textbook version of $C_u$, the lost-margin version, is the one every replenishment system uses. Which is why every replenishment system systematically under-orders at exactly the most important moments in a brand's life.

4. Behavioral Calculus

If the system under-orders, who over-orders? Historically, the answer has been: the human in the store. Behavioral economists call it overconfidence.

The polite version of the academic finding is that store managers are biased. The blunter version is that store managers see things the algorithm cannot, and their decisions look "biased" only because we are comparing them to a model that assumes the world is more stable than it is.

A 2025 paper in Sustainability on retailer overconfidence and the bullwhip effect splits "overconfidence" into two distinct cognitive parameters worth knowing precisely.

Bias 1: Overprecision ($\alpha$)

Overprecision is the belief that your information is more accurate than it actually is. Picture a weather forecaster who feels 90% confident in a forecast the data only supports at 60%. They are not lying. They genuinely feel certainty the world has not earned them.

The math captures this by warping how the retailer perceives uncertainty. The actual noise in the demand signal has a variance of $\sigma^2$. A retailer with overprecision $\alpha$ perceives that noise as smaller, specifically as $(1-\alpha)\sigma^2$. If $\alpha = 0.3$, the retailer believes their data is 30% tighter than reality. They underestimate uncertainty.

Bias 2: Overestimation ($\beta$)

Overestimation is the belief that the size of demand, or the size of your appropriate response to it, should be bigger than the math suggests. When the retailer sees a small uptick, $\beta$ acts as a multiplier. If the math says "order 100 more units," a retailer with $\beta > 1$ orders 130. They are not just responding to demand. They are amplifying the response.

Plug both biases into the bullwhip equation and the paper derives a closed-form expression for how chaotic upstream orders become as a function of these biases:

$$BWE_{order} = \frac{2\beta^2 \rho^2 (1 - \rho^{L+1})^2}{(1-\rho)^2} + \frac{2\beta\rho(1-\rho^{L+1})}{1-\rho} + 1$$

Where $\rho$ is the autocorrelation of demand, a number between 0 and 1 that captures how much today's demand depends on yesterday's. Set $\rho$ aside. Watch $\beta$.

$\beta$ shows up squared in the dominant term. So small increases in retailer overestimation produce accelerating increases in bullwhip. A manager with $\beta = 1.2$ is not 20% more chaotic than a perfectly rational manager. Because of the squaring, they are closer to 44% more chaotic. The paper's policy recommendation flows naturally: train managers to recognize their biases. Get them to behave more like the algorithm. Drive $\beta$ toward 1. Smooth the signal.

That is the conventional reading. In any market that is actually growing, it is backwards.

The Inversion

The bullwhip-from-overconfidence finding assumes a stationary world. It assumes $\mu$ is sitting still, and any over-reaction to a spike is wasteful because the spike will revert. True for mature, boring categories.

For the world we have been describing, though (Stanley in 2019, Oatly in 2017, Liquid Death in 2020), $\mu$ is moving, the algorithm cannot see it, and the only entity in the system capable of detecting the shift is the human staring at the actual shelf.

When that human triple-orders, what looks like overestimation is them integrating high-resolution local information the algorithm has no access to: who is in the parking lot, what they are reaching for, how fast the shelf empties on Tuesdays, what the woman at the register said when the last unit sold out. The algorithm has access to a POS feed that is, in many channels, a week or more behind real time. The store manager has access to the actual store.

The algorithm is estimating $D_t$, the level of demand. The manager is estimating $\frac{dD}{dt}$, the rate of change of demand. The "bias" we usually criticize them for is, mathematically, a heuristic derivative. They are sensing the slope of the demand curve, not just its position. In any non-stationary market, the slope is what matters.

The bullwhip-from-overconfidence research is correct in stationary markets. It is a guide to what to suppress when growth is over and stability matters. During growth, the very thing the research warns against is the only thing that works.

5. Information Theory

To formalize why the manager's override carries more information than the algorithm's order, we need a different mathematical framework: information theory, which Claude Shannon invented in 1948.

Information is surprise.

If I tell you something you already knew, or could have easily predicted, you have learned nothing. Information content: zero. If I tell you something you did not expect, you have learned a lot. Information content: high.

Shannon turned that intuition into math. The information content of an event $x$ with probability $p(x)$ is:

$$I(x) = -\log_2 p(x)$$

The unlikelier the event, the larger this number. The unit is "bits." Rare events carry more information; routine events carry essentially none.

Apply this to a supply chain. Two signals arrive at a brand's data warehouse on a typical Tuesday morning.

Signal A. Target's auto-replenishment system places an order for 2,400 units of SKU 47291. The order was generated by a deterministic algorithm running on a recent window of POS data the brand also has access to. The brand could have computed the same order itself. The probability of this exact order, given the POS history, is essentially 1, so the information content is essentially zero. Signal A tells the brand nothing it could not already infer.

Signal B. A store manager at a Walmart in Bentonville logs into the inventory system at 6:47 AM, manually overrides the algorithmic suggestion, and requests three times the recommended quantity. Managers override algorithmic orders rarely. A 3× override is rarer still. The probability is very low, so the information content is high. Many bits.

In Shannon's strict sense, Signal B is hundreds or thousands of times more informative than Signal A.

This is the principle Oatly intuited when they entered the U.S. market by skipping the grocery aisle entirely and seeding their product into high-end coffee shops. The grocery shelf is a low-entropy environment. Almost every order from a Whole Foods replenishment system is predictable from POS data, which means the resulting signals are information-poor.

A barista is a different kind of agent. When a barista tastes a new oat milk, decides she likes it, commits her cafe's budget and shelf space to ordering 20 cases, and tells her customers about it the next morning, she is producing a high-entropy event. Her decision is rare, costly to her, and based on local information no algorithm can access (the texture, the steam profile, the way it pairs with espresso). It contains many bits.

Oatly's founders weren't doing influencer marketing in any conventional sense. They were doing information arbitrage. They located their product in the part of the channel where every transaction carried maximum bits.

The general principle: in a supply chain, information is concentrated where prediction is hardest. The dominant data infrastructures of CPG (POS dashboards, syndicated scanner data, replenishment feeds) are built to capture routine, high-volume, algorithmically predictable signals at scale. They are, almost by design, structured to discard the anomalous ones.

This is the second mathematical censorship in the chain. The first, in Section 1, censored signal by treating it as noise. The second censors signal by failing to collect it at all.

6. The Sriracha Trap

Of the failure modes we have catalogued, one stands above the others in lethality. It is the moment when the supply chain stops merely being slow and starts actively deceiving itself. Statisticians call it left-censoring. CPG operators know it as the Sriracha Trap.

A POS system records a sale only when a sale occurs. When a product is in stock and nobody buys it, POS records a zero. When a product is out of stock and a customer wanted to buy it, POS also records a zero. From the algorithm's perspective, the two zeros are identical. They are not distinguishable in the data.

This is what statisticians mean by censored data. The underlying variable (demand) is not observable. Only a censored function of it (sales) is. And the censoring is one-sided. Sales can never exceed demand. The data is systematically biased downward whenever supply is the binding constraint.

The Huy Fong Sriracha shortage of 2022 to 2023 is the most-cited recent case. As Huy Fong's chili pepper supply collapsed and shelves emptied across thousands of stores, the resulting zero-sales records created exactly the self-reinforcing downward pressure on algorithmic forecasts that censored-data theory predicts. Operators in the category have widely reported that even after production resumed, retail-level demand forecasts and shelf allocations took a long time to recover. The empty-shelf interval had taught the systems, falsely, that demand had collapsed.

Formalize the dynamic as a feedback loop:

This is a negative feedback loop pointed in the wrong direction. Feedback is supposed to make systems converge toward truth. Censored data does the opposite: it makes them converge away from it. Mathematicians call this a degenerate equilibrium, a stable state that is also wrong.

Without that signal, the brand quietly disappears from the assortment over the next two category review cycles, killed not by lack of demand but by an algorithm that mistook its own success-induced shortage for failure.

7. The Precision Problem

Everything so far argues for listening to the human over the algorithm. Taken alone, that is dangerous advice.

If you read Section 5 literally and decided to act on every rare event in your channel, you would not detect more breakouts. You would manufacture your own bullwhip. A 3× manager override at 6:47 AM on a Tuesday could be a regime shift. It could also be a manager covering for a delivery that didn't show up. A one-off promotional pull. A buyer's hunch about an unrelated category. A data-entry mistake. Across a national retail footprint, the number of high-entropy events generated by purely idiosyncratic causes vastly exceeds the number generated by genuine regime shifts.

This is the precision problem. Most rare signals are just rare noise. And acting aggressively on noise is not free; it is the same upstream variance amplification the rest of this essay is structured around criticizing.

So how do you tell the difference? The solution lies in a layered set of consistency checks, each cheap on its own, devastating to noise when applied together.

Spatial correlation. A real regime shift is rarely confined to one store. Five managers across a contiguous trade area overriding the same week is a different event than one manager doing it alone. Geographic clustering separates idiosyncratic behavior from coordinated demand response. This is, not coincidentally, the pattern Stanley showed early on. The Buy Guide following wasn't one person in one town. It was a cluster.

Temporal persistence. Real regime shifts compound over time. A one-off override that never repeats is noise. An override that recurs the following week, intensifies the week after, and triggers similar behavior at neighboring stores is signal.

Cross-channel confirmation. When the same brand is being searched more often on Google in a region, mentioned more often on social, and manually over-ordered at retail, the triangulation gives you confidence the underlying demand is real and not a single-channel artifact.

Magnitude relative to baseline. A 3× override on a sleepy SKU during a quiet period means something different from a 3× override on a known-promotional SKU during a known-promotional week. Domain priors matter. The same nominal signal carries different information content depending on context.

Survivorship. Real regime shifts survive multiple validation rounds. Noise doesn't. Wait two weeks, look at the cluster again, and if it's still there and still growing, the prior on "this is real" climbs significantly.

What you end up building, taking all of this seriously, is not a system that "listens to high-entropy signals." It is a system that listens to high-entropy signals, runs each candidate through several layers of validation, and only escalates to action the candidates that survive. The precision discipline is what separates a useful detector from a noise-amplifying false-positive machine.

This sharpens what we should expect from a real solution. The Detection Lag isn't solved by anyone who can flag manual overrides; anyone with a SQL query can flag manual overrides. It is solved by whoever can flag the manual overrides that are correlated, persistent, cross-confirmed, contextually surprising, and survivorship-tested, and then act on them fast enough to matter.

8. The Volatility Tax

Up to this point, the consequences of the Detection Lag have all been operational: lost sales, depressed velocity, a self-reinforcing trap of censored data. But the damage does not stop at the warehouse. It propagates into the financial structure of the firm itself, in ways that show up in the firm's cost elasticity.

Cost elasticity measures how much a firm's costs change when its demand changes. A firm with high cost elasticity has mostly variable costs, which scale up and down with sales. A firm with low cost elasticity has mostly fixed costs, which are sticky and barely move. The trade-off is fundamental. Variable costs offer flexibility but are typically more expensive per unit (you pay a markup to a 3PL, a co-packer, a temp agency). Fixed costs offer per-unit efficiency but are dangerous to carry if your demand is volatile.

A 2024 paper in Contemporary Accounting Research on bullwhip and cost structure found that the bullwhip effect acts as a magnifying glass for demand uncertainty, and it forces firms into specific cost structures based on where they sit in the supply chain. The paper's findings, across thousands of firms:

• Manufacturers facing high bullwhip-amplified demand uncertainty shift toward low cost elasticity. They double down on fixed-cost capacity and automation. Why? Because manufacturers face the maximum amplification (they sit at the end of the whip), and the only economic defense against violently swinging order quantities is to absorb them with capacity that is already paid for.
• Retailers and CPG brands facing high bullwhip-amplified demand uncertainty shift toward high cost elasticity. They avoid fixed costs and lean on variable third-party arrangements. Why? Because they face demand they cannot predict, and committing to fixed costs in the face of unpredictable demand is what kills companies.

The economic consequence for a CPG brand: the bullwhip effect forces you to pay a flexibility premium. You cannot invest in the high-efficiency, high-fixed-cost manufacturing that would lower your COGS and expand your margins, because the demand signal you receive is too distorted to commit against. You pay the 3PL markup, the co-packer markup, the temp-agency markup. Not because they are economically optimal, but because they are the only way to survive the variance the channel is creating for you.

Call this the Volatility Tax. It is invisible in any single quarter. It just shows up as a quiet drag on gross margin relative to your category benchmarks. The Contemporary Accounting Research paper documents the structural finding without putting a single number on the magnitude. The directional claim, however, is solid: brands operating in highly bullwhipped channels structurally pay more per unit for capacity, logistics, and labor than they would in a clean-demand environment.

The bullwhip effect is not, as it is usually described, an "operations problem." It is a margin compression mechanism. It is the reason your finance team's gross margin model never quite ties out. And it is why every CPG founder you talk to eventually says some version of: "if we could just see demand cleanly, we could change our entire cost structure."

9. The Rebuild

So, what do you actually do?

The instinct of most operators is to ask for "better data." Higher-fidelity POS, more frequent pulls, prettier dashboards. That instinct is wrong. None of those things solve any of the problems we have catalogued. POS is a low-entropy, high-latency, left-censored signal. Doubling the pull frequency does not change its information content; it just gives you the same wrong signal twice as fast.

The fix is something the industry used to have and quietly dismantled: the territory manager.

What got cut when we cut field labor

Twenty years ago, every serious CPG brand had territory managers. Humans who drove between stores in their assigned region, walked the aisles, talked to store managers, fixed shelf problems, pushed for displays, gathered field intelligence, and reported back on what was actually happening at the point of sale. They were the brand's eyes and ears in the channel. They were also, in the language of this essay, the brand's high-bandwidth signal-collection layer. Every conversation a territory manager had with a store manager was a high-entropy event. Every override they noticed, every shelf configuration they corrected, every "we keep selling out of this on Saturdays" comment was a bit of information no POS feed in the world could replicate.

The job got cut. It got cut because it didn't scale, because the unit economics in a low-margin world stopped supporting field labor, and because POS data and replenishment algorithms appeared to make it redundant. The cuts looked rational on a spreadsheet. They were a catastrophe disguised as an efficiency. The brand traded a high-entropy human signal for a low-entropy algorithmic one and called it progress.

What an actual rebuild has to do

Any honest fix needs three components.

Detection. Some layer that plugs into a brand's transactional and customer data and runs analytics designed to surface store-level opportunities the existing tools cannot see. This is where the precision discipline of Section 7 lives. Not just flagging anomalies, but filtering high-entropy signals through the layered consistency checks (spatial clustering, temporal persistence, cross-channel confirmation, contextual magnitude, survivorship), and only surfacing opportunities that the math says are likely to be real. Where is velocity higher than the algorithm thinks? Which retailers are underserved relative to their local demand? Which accounts have stopped reordering and look like silent churn? The output isn't a dashboard with a thousand candidate alerts; it is a much smaller list of opportunities that have already survived several rounds of validation.

Action. Detection without action is a dashboard, and dashboards do not move product. The system has to go further. Once an opportunity is surfaced, something has to actually contact the store-level decision-maker who has authority to act. Targeted, contextual, designed to convert into a commitment to order. This is the layer most existing data products skip. They surface insights and assume someone else will act on them. In practice, the act of contacting the channel is what turns observation into participation. Every outreach is, in Shannon terms, a high-entropy event the system is deliberately generating, forcing the channel to produce signals it would otherwise never have produced.

Learning. Every contact generates a response. Stores commit to orders, or they don't. They report inventory levels, placement issues, what is selling and what is not. They tell you why they aren't ready to buy yet, or that a competitor's display went up next door last week. All of it has to flow back into the analytics, improving contact data, sharpening merchandising intelligence, tuning the targeting of the next round. This is the closed loop that locks the system onto the actual state of the channel rather than the lagged, inferred version of it.

The economics that killed the territory manager aren't coming back. But the function — a high-bandwidth presence in the channel, processing signals that POS data structurally cannot capture — is not optional. What form it takes in 2026 is the open question.

Coda: Timing as a Hedge

There is a line of thinking in modern finance that goes: the best hedge against being wrong is being early. It is not a rigorous claim, but it points at something true. When you are early, the cost of incremental information is low and its value is high. When you are late, the cost is high and the value has already gone to a competitor.

The CPG industry has spent twenty years optimizing for being right. It has built remarkably sophisticated infrastructure for predicting demand, smoothing variance, minimizing inventory, and squeezing working capital. The infrastructure works exactly as designed. Its design is the problem.

In a stationary market, optimizing for rightness is correct. In a non-stationary market (the only kind in which growth is possible), optimizing for earliness is correct. Safety stock is a hedge against being wrong about the level of demand. Timing is a hedge against being wrong about the existence of demand. The first protects against $\sigma$. The second protects against shifts in $\mu$. They are not substitutes, and in growth regimes the second is dramatically more valuable.

The algorithm will tell you what happened two weeks ago with high precision. The shelf is telling you what happens next month. Someone has to be standing there to hear it.

Stanley got lucky. The TikTok wave was loud enough that even a deaf system eventually heard it. Most brands do not get that lucky. The math says they don't have to be.