One picture, one purchase

A photograph doing two jobs

The creative constraint was the research constraint. We were not hired to make the lookbook smaller. The question was whether a click on the picture and a click on the buy control were one intention or two. If they were one, the edge control was a placement problem and a larger button on the image might have been enough. If they were two, the picture could stay huge and the purchase had to be given its own place on the grid.

The commercial fear was that a shop row would cheapen the season. That fear is about brand, and it is legitimate. It is also testable. Browsing can stay high while buying moves. If buying does not move, the row is not worth the compromise. Leah said that in the room before the variant was built.

Each look, not the page

The lookbook is a sequence of stories. A page-level add rate hides which look failed. We tied clicks, scroll depth, and add-to-cart to the look in view. An add that happened after three looks was attributed to the look whose control was used, not smeared across the season. Scroll depth told us whether the edge control had even entered the viewport.

Returning customers were reviewed separately from first visits. A returning customer already knows the house. If the miss lived only with them, or only with strangers, the design would have been different. It lived in both, which meant the control’s position was the common cause, not familiarity.

A zoom is not an add

Clicks that only enlarged the picture were coded as image events and were forbidden from counting as purchase intent. The old edge control sat so close to the image that a sloppy click boundary would have mixed them. We drew the boundary in the log before we drew it in the layout. Pinch-zooms and double-taps on the photograph joined the image class.

The first release and the following release were kept as separate samples. Pooling them would have let a one-week spike borrow the credibility of the next season. Bots and internal stylist previews were removed. Stylists add everything. They are not the customer.

A choice model with two kinds of click

The model estimated the chance of an add given where attention landed. Image clicks and edge clicks entered as separate features, along with whether the price had been in view. The result was blunt. Image attention did not raise the chance of an add once you knew the edge control had been available. The picture was the product and the distraction. It did its first job completely and its second job not at all.

The heatmap is the same sentence in cells. Almost all clicks fall on the large image. The narrow control at the right edge is nearly cold. A choice model that treats those clicks as one feature would have concluded that the lookbook was “engaging.” Engagement was true and it was not the purchase.

Hot picture, cold control

Read the heatmap as a miss, not as a success. Heat on a photograph in a lookbook is expected. Cold on the only action is the finding. The tension on the blueprint sits on Look 04 for that reason. The call to action under it is quiet, because quiet is what we want a control to be when it is finally in the right place: available, not begged for.

One product, one price, one row

The blueprint does not draw a coat. The look is an empty grid region, labeled Look 04, full width, carrying the heat. Under it, on the same unit: Wool coat, ink, £420, and Add to cart. Before, price and action sat in the margin, off the grid of the photograph. After, the image kept its width. The purchase stopped competing with the story by stopping pretending to be part of the story’s edge.

A shop, not a poster

Elise could live with one product, one price, one button if the result looked like a shop and not a ruined poster. We did not shrink the story. We stopped asking the edge of the picture to be the buy button. The model had already said those clicks were not the same event.

One release, then the next

The gridded look ran against the wide one for a single release. Add to cart from the lookbook was primary. Traffic to the look was watched and was not the goal. A design that made people leave the story faster would have shown up as a fall in look views. It did not have to. Browsing was allowed to stay.

Add to cart rose 18%. The following release, built the same way and not pooled into the first, repeated the rate. The heatmap had made the miss visible. The second release made it a result, not a one-week spike. The wide edge control did not come back.