The problem
Store sales went into the Google Shopping bid goal on January 3, 2026. By spring, Shopping ads were converting to ecommerce orders at about half the prior year's rate. The year-end view looked fine: paid search sales finished flat on 15% less spend. Two questions were open. What did the change cost ecommerce? Did bidding on store sales produce more total sales?
The method
- Measure the rollout as one change. It had three parts: store sales into the bid goal on January 3, a switch to a product feed built around store SKUs on February 2, and Shopping ROAS targets raised the same day from about 2 to about 9 to match store-scale values. The cost covers all three.
- Hold traffic and spend constant. A Poisson model on every campaign-week, with campaign and week fixed effects, clicks and cost per click as controls, and each campaign's own prior-year goal and target as the counterfactual. Weekly and daily versions agree within 0.5%.
- Test the break date. The February 2 step against 53 placebo dates with no known change. The real step was twice the largest placebo.
- Follow the lost orders. One row per order with every touchpoint, to see whether Brand search, non-brand search, email, or stores picked them up.
- Check the store side on equal terms. Same Shopping clicks, both years at the prior year's same-SKU prices, with the company-wide shift to pricier products removed.
- Find where the conversion went. Funnel step by funnel step, then a product-mix decomposition on the 119 landing products with at least 200 visits in both years.
What it showed
| Cost of the rollout, January to September 2026 | Measured |
|---|---|
| Ecommerce orders lost, share of what the site would have had | 5.2% |
| 95% interval | 2.9% to 8.0% |
| Site ecommerce conversion, without the rollout vs. actual | 2.25% → 2.13% |
| Shopping-ad shopper conversion, without vs. actual | 1.54% → 1.02% |
| Visits per Shopping click | unchanged |
No other channel picked up the lost orders. Ecommerce orders from journeys that touched a Shopping ad fell 40%. Brand search, non-brand search, and email lost orders on those same journeys. Orders with no Shopping touch fell too. There was no sign of recapture anywhere.
Stores did not make it back. Store orders per 1,000 Shopping clicks fell 13%, and only 4 of 16 comparable weeks were up. The store ticket rose, because Google spent the budget on premium beds. Measured like for like, omnichannel sales per Shopping click fell 12.7%. Ecommerce lost more than stores gained.
The loss was add to cart, on different products. Shopping visitors' add-to-cart rate fell from 8.5% to 4.1%. Once a shopper had a cart, the order rate held at 12%. Other paid search landings over the same weeks lost 13% on add to cart, not half. About half the drop is which products Google showed: premium beds went from 4% to 40% of Shopping visits, while everyday-low-price and web-priced beds fell from 67% to 18%. That fits store sales carrying the most bid value on premium beds. The other half is the same products converting worse. That half is not price, device, new vs. returning visitors, or the site, and its cause is not yet proven.
What changed
- The readout went on the analytics hub on October 16 with five recommendations. Put Value and Promo products back on a cart-weighted goal. Keep store value for Premium and Luxury but down-weight it with a value rule. Stop the daily target changes from the pacing rule, which moved each Shopping portfolio's target 167 to 274 times since the portfolios began. Settle store incrementality with a geo split test. When cutting Shopping spend, raise the target instead of the budget.
- A new Shopping test campaign launched October 7 was meant to bid on ecommerce sales. Within a week the analysis showed it bidding on the same store-heavy goal as every other Shopping campaign.
Technical notes
Model. Poisson pseudo-maximum likelihood on campaign × week, January to September 2026 against the same weeks of 2025, with campaign and week fixed effects. Levers: store sales in the campaign's bid goal, log target ROAS by tactic, and the feed swap as its own term. Controls: log clicks and log cost per click. Week effects carry company-wide conversion, basket size, attribution capture, season, and holidays. Intervals come from draws on two-way clustered coefficients, with a campaign cluster bootstrap as a check.
Net cost. Shopping last-touch orders lost, plus orders lost in other channels on journeys that touched a Shopping ad, plus Brand, non-brand, and local-inventory settings changes (not significant). Scaled to ERP ecommerce orders on settled data, after phone and chat orders finished linking. Four weeks of partial analytics load in March are added back at each tactic's clicks.
Placebo. A step dummy at each of 53 dates more than four weeks from any known Shopping change: largest absolute effect 0.27 on the log scale against 0.54 for the real step.
Store side. Web-first store orders stitched to the click, same weeks both years, the earlier year lifted to the later year's match rate. A weekly paired difference was negative with a 95% interval below zero. Whether any store dollars were incremental cannot be settled without a geo holdout that removes Shopping ads from shoppers near stores.
Like-for-like sales. Both years valued at the earlier year's same-SKU, same-channel prices, with the rise in dollars per order inside campaigns whose goal did not change removed from the later year.
Caveat. Observational. The goal, the feed, and the targets changed within a month of each other, and the targets were set partly by an automated pacing rule. The cost is stated for the rollout as a whole.
Related: brand search, where more budget could not buy more clicks, and online-to-offline attribution. Or back to all work.