Sweaty Betty used AI to write Google Shopping highlights and reported 80% more yoga revenue

12 min read

A product feed can be technically complete and still say very little. A pair of yoga shorts may have a colour, size, material and price, yet omit that they were designed to stay comfortable through a warm class. The website holds the richer story. The shopping feed can arrive at Google wearing only its care label.

That gap sat at the centre of a product feed experiment Sweaty Betty ran in the second quarter of 2026. The British activewear brand sent its catalogue to six channels across three markets. Its own product pages carried useful detail, and the syndicated feed described products with far less depth. Writing and maintaining highlights item by item would have turned a data problem into a permanent copy desk.

Sweaty Betty tested whether AI-generated highlights could make its Google Shopping feed more useful at catalogue scale. It kept matched products on the normal feed and gave the treatment group AI-generated descriptions of product uses and benefits. Athos Commerce, which managed the feed and published the case, reports 80% more revenue for 161 yoga treatment products than for the control, and 43% more for running products. The published case does not disclose the test duration, advertising spend or criteria used to match the products.

The catalogue had more knowledge than the feed

Sweaty Betty began in Notting Hill, London, in 1998 and sells women’s activewear for yoga, running and training. According to the case study published by Athos Commerce, the brand moved its product feed operation to Athos in January 2026. The platform syndicated data to Google, Meta, TikTok, AWin, Braze and StoryStream across the United Kingdom, Ireland and the United States.

The feed worked as more than a pipe. Custom rules handled fragmented stock, excluded archived sale products, mapped categories and attached labels used to segment campaigns. With that setup, feed quality was part of acquisition. A missing attribute could change whether a product entered an auction, which query matched it and what the shopper understood before clicking.

Google Shopping’s Product Highlights field gave the teams a specific place to intervene. Instead of adding a campaign or changing a bid, the Athos Commerce Channel Assistant generated short, product specific uses and benefits from data flowing through the feed. The AI product highlights were meant to say what an item is for, in language a search system and a shopper could both use.

The acquisition opportunity was in the existing catalogue. More informative product highlights might help Google match products to relevant searches and help shoppers assess them before clicking. The test therefore focused on a feed attribute rather than a new campaign, with matched products retained as a comparison group.

Matched products kept a piece of the old feed

The experiment compared product sets, with people left out of the unit of comparison. Within the yoga and running segments, Group A stayed unchanged, and some of those control products still carried manual, rules based highlights from earlier work. Group B received AI generated product highlights through Channel Assistant.

Athos describes this as a controlled A/B split across matched products. It then compared the groups two ways. The first was the straight difference after launch. The second considered each group’s performance before the change, which can reduce the risk that the treatment simply held stronger products from the start.

That second view matters because this was a product level test and a user level randomised test would behave differently. A legging does not switch variants when the next shopper arrives.

Products differ in price, stock, reviews, colour, discounting and existing demand. Matching can make two groups more comparable, though the public account does not say which variables were used, how close the matches were or whether products were assigned at random after matching.

The choice of segments was deliberate. Running had been used around the London Marathon, an event that can create a sharp demand spike. Yoga was selected after Wimbledon and a sponsored padel tournament, partly because it is less exposed to weather and event cycles. Testing two demand patterns gave Sweaty Betty a sharper question than one blended account total. Would the feed change travel across segments, or only catch a seasonal wave?

The yoga rail moved furthest

Athos Commerce reports that the yoga treatment held 161 products. Against the held back control, those products generated 62% more impressions, 72% more clicks and 80% more revenue. Impression share was 52% higher and return on ad spend rose 40%.

Running pointed the same way, with a smaller reported revenue gap. Treatment products generated 79% more clicks, 43% more revenue and 25% higher impression share than the control. Fraser Hodson, Sweaty Betty’s paid search manager, is quoted saying the results gave the company a reason to widen its use of AI generated product data.

The order of the metrics sketches a plausible mechanism. Richer highlights may have made more products eligible or relevant for more queries, which lifted impressions and impression share. More useful pre click descriptions may then have earned more clicks, and revenue rose further downstream. That is the path the intervention was designed to move.

The case publishes no orders, conversion rate, average order value or gross margin. The 80% yoga revenue difference could reflect more traffic, higher conversion, differences in product value or a combination. Athos also reports improved ROAS, but does not disclose spend by group or absolute revenue.

The holdout is stronger than the headline

The most transferable feature of this story sits beneath the number 80. The generation layer lived inside the product feed and could be switched on for a defined set while a matched set stayed untouched. Many catalogue changes get rolled out everywhere and judged against the previous month, which mixes the intervention with demand, promotions, budget changes and the weather. Weather matters a great deal in activewear.

A matched product holdout has flaws, and it still gives a growth team a counterfactual without building its own model. The same syndication path carries treatment and control, and the same account measures both. A before and after adjustment can account for some pre existing difference. Any retailer on a feed platform that preserves item level assignment and joins product exposure to commercial outcomes can copy the method.

For the technically minded, the point is that product data belongs to the acquisition system. The AI wrote a structured field that ranking systems consume across a live commerce pipeline, and a campaign slogan was never part of the job. That makes data lineage important. Teams need to know which source attributes produced each highlight, when the text changed, which destinations received it and whether a later feed rule replaced it.

It also creates an evaluation problem that sits between content quality and causal measurement. Check generated claims for product accuracy before launch, then measure commercial performance with a holdout after launch. One test asks whether the text is true, and the other asks whether the truth, stated more usefully, changes revenue.

Several useful denominators are missing

Athos Commerce supplied the system, ran the feed operation and published the case. Sweaty Betty’s named manager backs the commercial result in Athos’s account, and the retailer has not published the dataset or its own analysis. The source gives no duration for either segment test, and it also leaves out the number of control products, traffic, queries, budget allocation and market split.

There is no confidence interval, p value, Bayesian probability or pre registered stopping rule. The description says the yoga treatment gradually pulled away, and an outside reader cannot tell whether the test ran to a target sample or stopped when the line looked persuasive. The baseline adjusted calculation that Athos says it used is not open to inspection either.

The control adds a wrinkle. Some products carried manual, rules based highlights while others may have had none. That keeps the comparison commercially realistic, since the challenger had to beat the system it would replace. It also makes the mechanism harder to isolate. The result cannot say whether AI generation beat skilled manual copy, blank fields, inconsistent rules or a mixture of all three.

Athos reports large revenue differences between the matched product groups after introducing AI-generated highlights. Without the assignment method, underlying results or statistical uncertainty, the 80% yoga figure cannot be treated as a proven causal lift or a forecast for another retailer.

The next test should rotate the products

A stronger follow up would publish the matching variables and rotate assignment. Products could be paired on prior impressions, clicks, conversion, price, discount, stock depth, reviews and category. Within each pair, one product would be randomly assigned to treatment, and after a fixed period the assignments could switch, provided seasonality and stock were stable enough to make the crossover readable.

Revenue per product stays the primary business metric. Spend, impressions, clickthrough, conversion, average order value and gross margin explain the route, while product disapprovals, inaccurate claims, return rate and feed processing errors act as guardrails. Reporting absolute values beside relative lift keeps a big percentage tied to its commercial base.

A component test could separate generation from ordinary feed repair. One group keeps the current feed. A second receives rule based enrichment from verified website attributes. A third receives AI generated highlights. The comparison would show whether the model creates new value or mostly fills fields that disciplined catalogue operations should have completed anyway.

Sweaty Betty tested an existing acquisition constraint rather than launching another campaign. Athos reports higher revenue for products with AI-generated highlights, but the unpublished matching criteria, spend and conversion data limit how confidently that difference can be attributed to the text.

The next useful result would be revenue and gross profit per product, with spend and conversion data beside them. That would show whether richer Shopping highlights brought in more profitable demand, and whether AI generation outperformed simpler feed enrichment.

Experiment detail	What the reporting shows
Constraint	The syndicated feed described products in less depth than Sweaty Betty’s own product pages, and manual product level rules did not scale.
Intervention	Matched product sets compared the normal feed with AI generated Google Shopping product highlights.
Audience and period	Yoga and running segments in the second quarter of 2026. The yoga treatment held 161 products. Test duration and market were not disclosed.
Reported result	Athos Commerce reports 80% more yoga revenue and 43% more running revenue than the respective controls.
Limits	No control size, matching criteria, spend, absolute revenue, assignment method or statistical uncertainty is public.

Sources and disclosure

• Athos Commerce’s Sweaty Betty case study
• Sweaty Betty’s official company history
• Sweaty Betty paid search manager Fraser Hodson on LinkedIn

Athos Commerce managed the feed, supplied the AI product highlights and published the case. Sweaty Betty’s paid search manager is quoted by name, and neither company has released the underlying product assignments, raw results or statistical analysis.

FAQ

Compare more measured experiments in the Signal Index, browse the Growth Signal archive, or use the benchmark tools in the IC Lab.