Client story · Halden Goods · 2026

Checkout stopped asking early

Halden’s checkout showed delivery dates before the order was trusted. A leave classifier and twelve shop-aloud sessions put the exit on sequence, not on delivery. Waiting to show dates cut leaves 22% and raised payment 16%.

5 min read

−22%leaves before payment

  1. Capture
  2. Model
  3. Insight
  4. Interface
  5. Test
  6. Hold

Summary

Halden sells furniture and kitchen goods. Checkout revealed a delivery-date grid before the order summary had settled. People left while they were still checking the items. The funnel called it a delivery problem, because the last thing on the screen was a calendar. Delivery is expensive, and the team wanted more date options, not fewer.

For every checkout that did not pay we stored the step enter, the step leave, and the last element under the pointer. A classifier scored the chance of a leave in the next step from the elements visible, not from the campaign that started the session. The date grid was the last thing seen, but the sessions still contained item-check behavior. Twelve shop-aloud sessions said it in the room: they were counting chairs, and a calendar covered the count. The heatmap is bright on the right-hand date grid and only warm on the summary.

The summary stayed. A confirm control reveals the dates. Dates were not removed. They wait. Over four weeks, across furniture and kitchen, leaves before payment fell 22%. Sessions that paid rose 16%. The gain came from sequence, not from a new offer.

A delivery leak that was not about delivery

Priya Nand, in ecommerce, had a reasonable theory. The dates were the leak, because delivery is the expensive and uncertain part of a furniture order. The proposed fix was more date options. More options would have been the right fix if people were leaving because no date worked. The sessions did not say that. They said people were still reading the order when the dates arrived.

The question was whether a leave in the next step could be scored from what was visible, and whether that score blamed the calendar or the moment the calendar appeared. If the calendar itself was toxic, we would have changed dates. If the overlap was toxic, we would have changed sequence and left the dates alone.

The last element under the pointer

Checkout was already a sequence of steps. We did not need a new analytics product. We needed the step boundary, the elements on screen at the boundary, and the last element under the pointer for checkouts that did not pay. Paid sessions were kept as the contrast, not as the training pool for “what leavers look at,” because leavers and payers do not look at the same moment.

Shop-aloud sessions were twelve people with a real basket, asked to say what they were checking. They are not a sample for a rate. They are a sample for a sentence. The sentence was about counting items. The rate work stayed in the logs.

Leaves, payers, and the item-check that survived

A session that paid was removed from the leave class even if it had hesitated. Hesitation that recovers is not an exit. Item-check clicks were kept even when the last pixel was the date grid. Dropping them would have made the calendar look like the only behavior in the session, which is exactly the funnel’s mistake.

Refreshes and payment-error returns were stitched back onto the original checkout so a person was not counted as two leavers and one payer. Catalog was kept as a stratum, furniture against kitchen, because a result that lived in only one catalog would not have been a checkout rule. Bot-like bursts with no pointer movement were dropped. They do not have a last element.

A classifier of the next step

The classifier scored the chance of a leave in the next step from the elements visible at the current step. Campaign, channel, and the ad that started the session were available and were not used as the explanation. They shift who arrives. They do not decide what the screen does to them once they are in checkout. Using them would have produced a media recommendation dressed as a UX finding.

Visible elements carried it. When the date grid and the summary were on screen together, leave risk jumped, and the pointer’s last rest was the grid. The sessions still contained item-check clicks in the seconds before. Delivery was the screen. Doubt about the order was the cause. The heatmap shows the split in space: the right side, the dates, is hottest. The summary on the left is only warm.

Bright on the grid, thin on the items

The checkout heatmap is the argument against “people are fascinated by delivery.” Fascination and obstruction look the same in a last-click report. They do not look the same next to item-check clicks and a sentence from a person counting chairs. The right-hand grid outruns the summary. That is the tension the blueprint still marks, under the order, waiting.

HeatmapAttention in checkout. The right-hand date grid outruns the summary on the left.

The order stays. The dates wait.

The screen is an order, not a delivery picker. Oak sideboard, one item, £640, and Confirm items. The date grid is on the drawing as a dashed region labeled after confirm, and it carries the heat so we do not forget why it moved. Before, the date grid and the summary arrived together, and on a phone the grid covered the items. After, the summary stays. Nothing else was added. Dates were not cut. They wait until the items are confirmed.

Halden Goods
ShopDelivery
Order
Oak sideboard
1 × Oak sideboard£640
Confirm items
DatesAfter confirm
Order
  • Quiet
  • Low
  • Warm
  • Hot
  • Tension

Order blueprint for Halden Goods. The sideboard summary stays in view. Tension sits on the date grid, which waits until the items are confirmed.

Twelve people, one calendar

Priya heard the same line the classifier had scored. Twelve people said they were counting chairs while a calendar covered the count. Leah’s rule for the variant was sequential: if payment does not rise with the drop in leaves, we do not keep it. A screen that stops exits by also stopping purchases is not a resolution.

Sequence, tested on both catalogs

Summary-first ran against the old sequence for four weeks, furniture and kitchen both. Leaves before payment were primary. Completed payment was the hold. We did not add a discount, a new delivery promise, or a rewritten product description. If payment moved, it had to move because the order stayed readable.

Leaves before payment fell 22%. Sessions that paid rose 16%. Both catalogs agreed. The conversion gain came from sequence, not from a new offer. The model named the step. The test proved the step.

The test for Halden Goods: control, variant, sample, primary result, and the measure that had to agree.
ControlDates and summary together
VariantSummary, then dates
Sample4 weeks, both catalogs
Primary−22% leaves before payment
Held+16% sessions that paid