Thinking · October 2026 · From the stage
What customers say and what they do
The written version of the Headline Stage session at eCommerce Expo, ExCeL London, 24 September 2026, which I gave with Joe Jaques of Jaques of London. The programme promised an answer to one question: which customer signals should guide action. This is that answer, set down so it can be argued with. The slides from the session are available as a PDF.
Everyone stand up, please
In 2017 I spent three days presenting to the store managers of Harveys and Bensons, in conference rooms in Glasgow, Bolton and London, in the post-lunch slot nobody wants. Four hundred people in total. I opened each session the same way: everyone stand up, please, and stay standing if you think ecommerce is stealing your sales.
Over the three days, 300 of the 400 stayed on their feet.
That was the loudest customer signal in the building, and no customer had been asked. It was 300 store managers agreeing with each other. It felt like insight because it was unanimous and heartfelt. It was an internal assumption wearing a customer badge, and it is the most common signal retailers act on.
Five kinds of signal
Most bad trading decisions I have seen come from mistaking one kind of customer signal for another. There are five.
SAY is what customers tell you when you ask: surveys, focus groups, the Net Promoter Score. SEARCH is what they look for, in your site search and on Google. DO is what they click, add to basket and abandon. BUY is what they repeatedly pay for. SHOUT is what they complain about: the review, the return, the email to the chief executive.
None of these is useless. Each answers a different question, and the trouble starts when one is asked to answer another’s.
SAY: Walmart asks a reasonable question
In 2009 Walmart wanted to compete with Target on store experience, so it asked customers whether they would like the stores to be less cluttered. Customers said yes. People generally do. Under Project Impact, Walmart cut around 15% of its range, on analyst estimates, and cleared Action Alley, the pallets of fast sellers down the main aisle. Bill Simon, then chief operating officer of Walmart US, described the aim as opening up the space for the customer.
What followed was eight straight quarters of US like-for-like decline to the first quarter of 2011, and the rollout was put on hold. The customers who said yes to less clutter were the same customers who came for cheap abundance. They answered the question they were asked. They were never asked what they would stop buying.
DO and BUY: the behaviour your operation lets you see
Tesco is the standard reference for acting on what customers do. In its 2024/25 results, Clubcard sales penetration in the UK stood at 84%. Tesco can see what most of its customers buy, how often and alongside what, and it builds its pricing and personalisation on that rather than on what customers say they intend.
IKEA’s UK business shows the less obvious half of the lesson. It wanted more customers to swipe their IKEA Family card at the till. It did not survey customers about the card. It worked on the co-workers: live leaderboards in store, sign-ups on the shop floor, rewards for the best performers. According to its 2024 DMA Awards entry (the entrant’s own figures, so read accordingly), swipes went from one in five transactions to one in three, monthly sign-ups doubled, and Family’s share of sales rose from 31% to 46%.
The point is uncomfortable. Before the change, four in five transactions carried no Family card. In all likelihood a good share of those shoppers were members nobody had asked to swipe. The customers were there all along; the operation had not been set up to count them. Your dashboard is the part of the customer your systems were built to notice.
When behaviour misleads too
Behaviour beats opinion. It is still not enough on its own.
In 2012 an employee at Bing suggested lengthening the headlines on its search ads. The idea was ranked low among hundreds and sat in the backlog for more than six months, until an engineer tested it because the code was cheap to write. Within hours an alert fired: the result was too good, and alerts like that usually mean a logging error. It was real. Revenue rose 12%, worth more than $100 million a year in the US alone, and it replicated. Note what the system did with the best idea in its history: it assumed it was a bug. That scepticism is the right reflex. Ronny Kohavi’s record of experiments at Microsoft is that roughly a third of well-designed tests improve the metric they were built to improve. The rest do nothing or make it worse.
A behavioural number tells you what happened. It does not tell you why. Conversion up can mean a better proposition, or that only the easiest customers are getting through. More items added to basket can mean intent, or shoppers using the basket to find out the delivery charge.
The cleanest recent example I know is my own test of an AI stylist in July. THG Ingenuity reported that MP Activewear shoppers who generated at least one image with its AI Stylist were nearly six times more likely to convert than the site average. That looks like a DO signal. It is closer to a SAY signal in disguise: a measure of which shoppers chose to use the feature, and those are the most committed shoppers on the site before the tool does anything. There was no control group. When I tried it with a photograph of a family member, it put a £5.99 tee on him and asked nothing about fit. The six-times figure measured intent and credited it to the tool. THG Ingenuity sells the tool, which is worth knowing too. The test, with the images, is in The chatbot paradox.
SHOUT: the metric on the board pack
Most board packs carry a Net Promoter Score, and most of them are asking the wrong people. Baehre and colleagues studied seven US sportswear brands over five years and found that NPS did predict sales growth, but in its brand-health form, sampled across all potential customers. The transactional survey sent to people who have just bought is a different instrument with the same name. A companion paper, on 193,220 responses, found that simple top-box measures of the same likelihood-to-recommend question performed better than NPS. Keep the score if the board likes it. Know which question it can answer.
Friday, 10:17
On stage we ran a short exercise with illustrative figures, not taken from any business I have worked in. Sales are down 18% this morning. Conversion is down 11%, traffic is flat, site search is up 22%, basket abandonment is up 14%, delivery complaints are up 31%, NPS is down three points, and a competitor is running 20% off today. What do you do?
Every answer is wrong, including doing nothing, because there is not enough information to act. The same numbers could be a stock problem, a broken delivery promise, a competitor’s price, a technical fault or a change in traffic quality. Trading is the discipline of deciding which signal is causal enough to deserve a decision. The hardest sentence on a Friday morning is telling the chief executive that you do not yet know, and what it will cost to find out.
The test
Before a signal earns a decision, five questions.
Who produced it: customers, a subset of customers who chose to respond, or colleagues? What kind of signal is it: SAY, SEARCH, DO, BUY or SHOUT, and is it being asked a question that belongs to another kind? When was it produced, and is it from before or after the thing you changed? Why might it have moved, and how many explanations fit the same number? So what: what would you do differently if it were true, and is that worth the cost of finding out?
Stay standing
At the end of each of those 2017 sessions, after presenting what online was actually doing for the stores (customers discovering products online and then visiting, those customers converting better in store, the investment in making stores easier to find, the “See inStore” tools and the early VR and AR work), I asked the same question again. Stay standing if you still think ecommerce is stealing your sales.
Over the three days, fewer than 15 stayed up.
The customers had not changed between the first question and the second. The evidence in the room had.
Sources
- Harveys and Bensons store managers roadshow, Glasgow, Bolton and London, 2017: first-hand.
- Walmart Project Impact: DailyFinance (AOL), 31 March 2010, quoting Walmart US COO Bill Simon; range reduction per analyst estimates; Walmart Q1 2011 results as reported via Blueocean, Perishable News, June 2011.
- Tesco PLC, Preliminary Results 2024/25, April 2025: Clubcard sales penetration UK 84%.
- IKEA Family: DMA Awards 2024, Gold, Retail and E-commerce, Nudging up Loyalty for IKEA, agency RAPP. Agency award entry; figures are the entrant’s own.
- Kohavi, R. and Thomke, S., The Surprising Power of Online Experiments, Harvard Business Review, September to October 2017.
- THG Ingenuity, AI Stylist press release, 29 July 2026 (vendor: sells the product measured); Barking Cat first-hand test, July 2026.
- Baehre, S., O’Dwyer, M., O’Malley, L. and Lee, N., Journal of the Academy of Marketing Science 50(1), 2022; Baehre, S., O’Dwyer, M., O’Malley, L. and Story, V., Journal of Business Research 149, 2022.
- Trading-room figures: illustrative.