Sentiment: listen and improve.

This is where people arrive and where they leave, and there is no flow here on purpose. Sentiment is not a message the brand sends. It is what the brand hears, from the people who left, from the people who stayed, and from the numbers that only tell the truth when they are counted in the right direction. Make sure to document what you need to improve, call it even, and move on. Ray Dalio's formula for it is three words long: pain plus reflection equals progress. The pain is free. The reflection has to be scheduled, and Deming scheduled it fifty years ago: plan, do, study, act, with the study step the one everyone skips.

The step: churn

Churn is Hirschman's exit: the departure that tells you something is wrong without telling you what. The flows in the six stages before this one exist to reduce it, and this stage exists to find out whether they did. That question is harder than it looks, because the reports are built to flatter, and Kahneman explains why we let them. What you see is all there is: the mind builds a confident story from the evidence in front of it and does not go looking for what is missing. A flow report is evidence in front of you. The held-back group that would disprove it does not exist until you create it.

Go back to the overview. A flow report counts the people who received a message and then ordered. At a pharmacy brand, abandoned checkout converted 3.22% of recipients and ranked first of four flows, and the suggestion in front of me was to make it the place to grow. The ranking measured how close each audience already was to buying. At an architectural products retailer, the email platform credited email with roughly 7 to 12 times the revenue Google Analytics gave it on last click, and the flows had only 165 clicks in the window, so most of that credit came from opens. A vendor's abandonment emails were credited with about ten orders; going through the buyers one at a time, 11 of 19 had bought without opening the email, and 18 of 19 had been on the site within a day. None of this was a lie. All of it was attributed, and attributed is not incremental.

Feynman's first principle is the only defence: you must not fool yourself, and you are the easiest person to fool. The only instrument that reads sentiment honestly at the flow level is the holdout: hold back a tenth of entrants, change nothing else, and pull orders for both groups from the store. The difference between the two order rates is what the flow adds. Everything else is a map of where buyers already were.

The split that changes the story

The same discipline applies to the whole business, not only the flows. A pharmacy and health brand came out of a hard spring with revenue going from 31.3% down to 12.4% up year over year between April and August. True. One split changed the story: everyday orders were still about 13% down, and bulk orders had carried the summer. By the first quarterly review, store revenue was up 12.5% year over year and revenue without bulk was up 0.3%, so the Q4 target was set on the second number. The topline alone would have told a true story that was practically wrong, and the brand would have spent the quarter congratulating itself.

That is what listening means at this stage. Not a survey. The one split in the data that the headline hides, found before anyone acts on the headline. Strathern's version of Goodhart's law explains why the headline hides it: once "revenue up" became the target, it stopped being a measure of anything the team could act on. The measure that still works is the one nobody has been asked to hit yet.

The number that could not be checked

At the same brand, the post-purchase plan had put the second order at day 73. I went looking for where the 73 came from: an average of averages over about 234 rows, from a segment that favoured fast repeat buyers, and that segment now held nobody, so the figure could not be reproduced. Across 2,477 customers who came back, the median gap was 111 days. The lesson I wrote down was not "73 was wrong". It was: a number that cannot be recomputed from the store's orders is not a number, it is a memory. Every model I build since has to reconcile to the store within about 2% before it is allowed to forecast, and the forecast is tested against what then happens: 4.9% off over 30 days, 2.7% off over 90, at the architectural products retailer. That is Deming's study step with a tolerance on it.

The objection: we review the numbers every month

Most brands do, and the review is the problem: it reads the reports, and the reports were counted backwards. A monthly meeting that asks "what did email do" and gets the platform's answer is Kahneman's confident story on a calendar. The review that works asks three different questions: which direction was this counted in, what is the split under it, and can the number be recomputed from the store. At the brands I run, the agents that pull the data have to pass every figure through a gate that asks exactly those questions before a person sees it, and the figures that fail are the ones worth the meeting.

Document it, call it even, move on

The reflection has a format. At the brands I run, every month ends with a short written record: what the numbers said, where they disagreed with each other, which split mattered, what changed because of it. It is the least glamorous page in the operating system and the one the next quarter depends on. Three habits make it work:

• Write the lesson, not the excuse. "The reminder fired at 89 and buyers came back at 111" is a lesson. "Reorder timing was suboptimal" is an excuse with a thesaurus.
• Date it, and put the number next to it. A lesson without a number cannot be checked next quarter, and a lesson without a date cannot be forgiven when it turns out to be wrong.
• Call it even. The point of the record is to stop carrying the mistake. Once it is written, the funnel has one less thing to teach you, and you can go back to Stage 1 and start winning people again.

What I check first

• Which direction the headline was counted in. Forwards from everyone who entered, or backwards from the people who finished.
• One holdout, on the flow with the best numbers. Because that is where the flattery is largest.
• The split under the topline. Everyday against bulk, new against returning, web against phone. The one that changes the story is the one to plan on.
• Whether the number can be recomputed. From the store's orders, by someone else, next month.
• Where last quarter's lessons are written. If the answer is nowhere, the funnel has been teaching an empty room.

If you take one thing from this stage: count forwards, split the headline, write it down. That is the flywheel. Seven stages, one job each, and a record of what each one taught you. The people who leave at this stage are the reason the next visitor, back at Stage 1, gets a better ten seconds.