If you sell through distributors, resellers, retailers or marketplaces, the richest customer dataset you own is a shipping record. It tells you that a pallet left a warehouse, that a reseller reordered on Tuesday, that a SKU is moving faster in the Midwest than in the Southeast. Teams call this customer data. It is not customer data. It is logistics data with a customer-shaped hole in the middle.
That hole is not a rounding error. In May 2026 alone, US merchant wholesalers sold $817.4 billion of goods, according to the Census Bureau Monthly Wholesale Trade report released on 8 July 2026 (release CB26-107). Every one of those transactions has a person or a business at the end of it, and for the manufacturer upstream, almost none of them have a name.
Answer first
Sell-through data proves that units crossed a boundary. It does not identify a customer, a reason, or an alternative. A sell-through record is generated by an inventory system, not by a buyer. It can tell you what moved, where, when, and how much with excellent precision. It cannot tell you who bought it, what they compared it against, what job they hired it for, what they would have bought if you had been out of stock, or whether they will buy again. Those five facts are not missing because your reporting is immature. They are missing because they were never captured by the event that created the record.
The practical consequence: channel data is a superb operational signal and a poor research instrument. Treating it as the latter is how a company ends up with a decade of sales history and no idea why anyone buys.
Sell-in, sell-through, sell-out: three different events
Most channel arguments are really vocabulary arguments. Three separate events get collapsed into one word, and each proves something different.
| Record | The event it captures | What it proves | What it cannot show |
|---|---|---|---|
| Sell-in | You ship to the distributor | A purchase order was placed by a partner | Whether anyone downstream wanted it |
| Sell-through | The distributor ships to a dealer or retailer | Goods moved one tier closer to a buyer | Whether the goods left the shelf |
| Sell-out | The end customer buys | A real purchase happened | Who bought, why, and what else they considered |
Sell-in is the number most manufacturers actually have, because it is generated by their own ERP. It is also the number furthest from the customer. A strong sell-in quarter can mean demand is rising, or it can mean a partner is building stock ahead of a price increase. The record looks identical in both cases.
The gap between these tiers is measurable, and it is large. The same Census release put wholesale inventories at $941.8 billion at the end of May 2026 against that $817.4 billion of monthly sales, an inventories-to-sales ratio of 1.15. Roughly 35 days of goods were sitting in the channel between the manufacturer and anybody who wanted them. Every unit in that 35-day buffer has been sold once, from your point of view, and zero times from the customer's point of view.
That buffer also moves. The same ratio was 1.31 in May 2025, a fall of about 12% in a single year. Nothing about that shift tells you anything about customer preference; it is a statement about how much stock the channel chose to carry.
This is not a new observation
It is the same problem Hewlett-Packard documented in the article that named the phenomenon. In The Bullwhip Effect in Supply Chains (MIT Sloan Management Review, 38(3):93-102, 1997), Hau Lee, V. Padmanabhan and Seungjin Whang describe the position plainly: "In the past, without being able to see the sales of its products at the distribution channel stage, HP had to rely on the sales orders from the resellers to make product forecasts, plan capacity, control inventory, and schedule production."
Read that sentence carefully. HP was not short of data. HP had every reseller order. What HP lacked was visibility past the reseller, so the reseller order became a stand-in for customer demand. Nearly thirty years later, that substitution is still the default in most channel businesses, and it is still the same substitution.
The four questions the record structurally cannot answer
There is a clean test for whether a question can be answered from channel data: ask whether the answer was present at the moment the record was written. If it was not, no amount of analysis will recover it.
1. Who bought it? The record identifies the account that ordered, which is your partner, not your customer. Warranty registrations, rebate claims and product registrations capture a slice of end customers, but a self-selected one: the people motivated enough to fill in a form. That is a biased sample of your buyers, and it is biased in the direction of enthusiasm.
2. Why did they choose you? Purchase is a single observable outcome of a decision process that included a consideration set, a price comparison, a recommendation, and possibly a specification written by someone who is not the buyer. The record contains the outcome and none of the process.
3. What did they reject? This is the most valuable and most completely absent field. Your channel data contains no row for the customer who walked in, looked at your product, and left with a competitor. The dealer knows. Your database does not.
4. What happens next? Repeat purchase, satisfaction and referral all occur downstream of the boundary. A second order from the same distributor is not evidence of a happy customer; it is evidence of a distributor restocking.
If those four questions matter to your roadmap, pricing or positioning, they need an instrument that asks a person. Our guide to customer research for manufacturing and industrial B2B walks through how those instruments get placed in industrial channels, and the logistics and supply chain equivalent covers shipper, carrier and end-customer research where the same boundary problem applies.
What channel data is genuinely excellent for
It would be dishonest to argue that sell-through data is weak evidence in general. It is outstanding evidence about a specific set of things, and confusing the two is the actual error.
Channel data reliably measures velocity (how fast units move through a location), distribution (how many locations carry you at all), geographic variation, stock position, and mix. Those are real, decision-grade facts. A regional velocity gap is a genuine finding. It is simply not an explanation, and the temptation is always to supply one from imagination.
The disciplined pattern is to use channel data to locate the anomaly and a research instrument to explain it. Note that this is also why simply collecting more channel data does not close the gap: as we show in why adding distributor data makes your customer picture worse, order data amplifies at every tier it crosses. Velocity data tells you the SKU is underperforming in one region by 30%. It cannot tell you whether that is a shelf-position problem, a price-perception problem, a competing local brand, or a dealer who stopped recommending you. Those four causes imply four different fixes, and choosing between them by inspection of the sales chart is guesswork with a spreadsheet attached.
The partner is not the same respondent as the customer
The obvious next move is to ask the partner, since the partner meets the customer. That is worth doing, and it is a different study with a different subject. Measuring how your partners feel about your programme, margins and support is partner satisfaction research, and it answers questions about the partner relationship. It does not substitute for end-customer evidence, because the partner is reporting a filtered recollection of somebody else. When you genuinely cannot reach end users and must work through an intermediary, the proxy research playbook sets out how to do it and, more importantly, what the resulting evidence can and cannot support.
How to put a customer back into the record
The structural fix is not more channel reporting. It is attaching a direct instrument to the moments where an end customer is briefly reachable: at registration, at onboarding, at first support contact, at renewal, inside the packaging, on the dealer forecourt, in the app.
This is where the economics of AI-moderated research change the calculation. Historically, running qualitative research at the end of a channel was prohibitive: you would need a recruiter to find dealer customers, a moderator to run sessions, and weeks of analysis, for a study that might yield thirty conversations. The cost per answer was high enough that most manufacturers simply did not do it, and defaulted to the sales chart.
Koji removes both the recruiter and the moderator from that equation. A single interview link goes out through whatever touchpoint you control, and the AI interviewer runs the conversation, probes the interesting answer, and moves on. If no touchpoint of yours reaches the people you need, you do not need an audience of your own either: describe the market, the demographics and the screening you want, approve a live per-respondent quote in credits, and Koji recruits the respondents for you. Because the interviews are AI-moderated, running 500 of them costs what running 20 used to, and there is no moderator whose phrasing drifts between session 3 and session 300.
Two features matter specifically for channel work:
- Structured questions alongside open conversation. Koji supports six question types (open_ended, scale, single_choice, multiple_choice, ranking and yes_no), so a single instrument can capture the clean quantitative fields your channel data lacks (which brands were in the consideration set, ranked purchase drivers, a satisfaction scale) while still probing the reasoning in their own words. You get the countable comparison and the explanation from the same respondent.
- Automatic thematic analysis. Channel research generates messy, region-specific language. Koji clusters themes across every interview and produces a one-click report, so a study across four territories does not become a four-week coding project.
The result is that the anomaly you spotted in the velocity data on Monday can have an explanation with quotes attached by Thursday, instead of entering a research backlog. That is the difference between a sales chart that raises questions and an evidence base that answers them.
One caveat worth planning for before you start: capture identity and consent at the moment of the interview, because the customers your channel owns cannot be re-contacted afterwards.
Start free with 10 credits and no card, and no subscription after that. Interviews start as low as €1 per qualified interview, and you pay only for what your study actually uses. Only conversations that score 3 or higher on the quality gate consume credits, so a pilot at the end of a distribution channel is not billed for the noise.
Frequently asked questions
What is sell-through data?
Sell-through data records goods moving from a distributor or retailer to the next tier down, typically expressed as units or value per period per location. It sits between sell-in (your shipment to the partner) and sell-out (the actual end-customer purchase). It proves that inventory moved and is used for velocity, replenishment and forecasting.
What is the difference between sell-in, sell-through and sell-out?
Sell-in is what you ship to your channel partner. Sell-through is what the partner ships onward. Sell-out is the final purchase by an end customer. Only sell-out corresponds to a buying decision by a person, and most manufacturers have the least visibility into exactly that tier.
Why can I not use channel sales data for customer research?
Because the record is generated by an inventory transaction rather than by a buyer, so the fields a researcher needs were never captured. Identity, consideration set, reason for choice, rejected alternatives and intent to repurchase are absent at the moment of writing, and cannot be recovered by analysis afterwards.
How much product is sitting in the channel at any time?
For US merchant wholesalers in May 2026 the inventories-to-sales ratio was 1.15, against $941.8 billion of inventories and $817.4 billion of monthly sales, which is roughly 35 days of stock. The ratio was 1.31 a year earlier, so the buffer itself changes materially year to year.
Are warranty registrations a good substitute for customer data?
They are useful but self-selected. Only customers motivated enough to complete a registration appear, which skews the sample toward engaged and satisfied buyers and under-represents exactly the disappointed customers whose reasoning would be most valuable.
How do I research customers I reach only through a distributor?
Place a direct instrument at a touchpoint you control (packaging, registration, onboarding, support, app or an interview link the partner distributes) and capture consent yourself at that point. AI-moderated interviews make this economical at the volumes a channel produces, and capturing consent directly is what preserves your ability to follow up later.
