Redbrain wins Google's Agentic Commerce Trailblazer award

On 6 October 2026 Google named Redbrain the Agentic Commerce Trailblazer at its CSS Spotlight Awards. It’s a new category this year, created to recognise the Comparison Shopping Service partners doing the most to prepare retailers for AI-led shopping.

We won it for one piece of work: a live test of AI-written product content, run across 520 retailers and more than 22 million products. This article explains what we tested, how we tested it, and what we found.

What is agentic commerce?

Shoppers increasingly ask an AI assistant to help them choose rather than scrolling a grid of results. “Which cordless vacuum is best for pet hair?” is a question, not a keyword, and the assistant answering it needs more than a product title and a price.

It needs the details a good shop assistant would know: what the product is for, what makes it different, and what people usually ask before they buy. Google Merchant Center now accepts that kind of information as structured attributes, including product highlights, short, scannable statements of what matters about a product, and questions and answers.

Most retailer feeds don’t carry any of it. Writing it by hand for thousands of products isn’t realistic, and that is the gap we set out to close.

What we built

Our feed optimiser uses AI to write the conversational content AI shopping experiences draw on, for each product:

• Product highlights, short, scannable statements of what matters about the product
• Questions and answers covering what shoppers ask before they buy

Every prompt is tested before it reaches a live feed

The live test below measures what shoppers do. Before any new prompt gets that far, it goes through our own evaluation tool, which generates content for real product data and scores it offline. It checks for:

• Hallucinations. Every fact the AI adds beyond the retailer’s feed is pulled out and checked on its own, including product attributes it might have invented, such as a material, a size or a feature the product doesn’t have, because we need to be able to trust the output at scale.
• Language contamination. We write in 15 languages, so every output is checked to confirm it is entirely in the language requested, with no English drifting into a German listing.
• Consistency. The same product is generated several times to make sure the output is stable, not a different listing on every run.
• An independent judge. An AI model from a different provider from the one writing the content compares each candidate against the current version, side by side. A model never grades its own work.

We’ll cover how the tool works in more detail in a future article.

How we tested it

A claim that AI content “works” is only as good as the test behind it, so we ran this one as a controlled experiment rather than a before-and-after comparison.

From June 2026, every product from 520 retailers was split into two equal halves. One half showed the AI-written content. The other half stayed exactly as the retailer supplied it, as a control.

For product highlights and Q&A, that means something against nothing: retailers don’t supply them at all, so the control products had none. Both halves ran in the same auctions, in the same markets, on the same days, so seasonality, promotions and competitor moves affected them equally.

By October the test had run on more than £1m in ad spend, with more than 11 million products on each side, across hundreds of retailers in fashion, electronics, beauty, home and sport, and in markets across Europe.

How we checked the results were real

A difference between two groups can be luck. Before trusting any number, we put it through four checks:

• A fair, fixed split. Each offer was assigned to its group by a hash of its ID, not by hand. The halves were the same size and comparable, and no offer could switch sides partway through the test.
• Sized before it started. A power calculation told us how many offers each side needed to detect a meaningful change in performance, so the test was large enough for a real effect to show and a fluke to wash out.
• Bayesian analysis. For click-through rate, conversion rate, impressions and daily performance, we simulated a million plausible outcomes for each group. That gives the probability that the AI content beats the original, and a credible range for how big the difference is, rather than a single number that could mislead.
• Bootstrap resampling. Pooling millions of offers can make a Bayesian model overconfident, so we also resampled the test offer by offer, 2,000 times. If a result only holds up in a handful of those redraws, we don’t report it as a win.

The same checks run retailer by retailer. That is how we know the overall gain is real, and also how we know some retailers didn’t share in it. A small catalogue can swing a long way on chance alone, so we treat a small retailer’s result as a pointer rather than a finding.

What we found

Revenue rose 4.5%. Across the whole test, products with AI-written content generated 4.5% more revenue than their unchanged twins.

The gain came from more valuable orders, not more of them. This was the result we didn’t expect. The number of orders barely moved, up 0.3%, but revenue per order rose 4.2%, which points to shoppers buying higher-value products. We set out to win more clicks. What the content actually did was change who was clicking.

We see three likely reasons, and they reinforce each other:

• Better search matching. Highlights and Q&A give Google more to match against. Products start appearing for longer, more specific searches, the kind people type when they know what they want and are close to buying, rather than broad browsing terms.
• Stronger signals for Google’s systems. Google’s bidding and ranking learn from the product data they’re given. Content that says clearly what a product is and who it’s for helps them put it in front of the shoppers most likely to value it, which fits the gain building over time rather than arriving on day one.
• A more committed audience. Highlights answer a shopper’s questions before the click. The people who still click have already compared and decided, so they arrive ready to buy, and more often buy the higher-value option.

It wasn’t a one-off: products with AI-written content earned more per order in 11 of the 15 weeks, and in every one of the last six. These are our working explanations for why rather than proven causes, and testing them is the next stage of this work.

Shoppers also saw and clicked the enriched products more. Impressions rose 2.0% and clicks 2.6%. Small percentages at this scale add up: that is millions of extra impressions and six figures’ worth of extra clicks during the test.

Results depended on what you sell. This is the finding we most want other retailers to hear. The overall figures average every kind of retailer together, and the retailers that came out ahead gained far more than the average suggests: across them, impressions and clicks both rose by close to 6%, more than double the test-wide figures. The gains were strongest for technical products like electronics, where shoppers compare features and highlights give them something new to compare. Luxury and premium fashion was the exception. These brands sell on a carefully built voice, and their own copy already does that job, so AI-written content didn’t add value for them.

The effect grew over time. The gap between the two halves widened the longer the test ran. We are still working out how much of that comes from Google’s systems learning to use the new content.

We are sharing these results as they are, because that is the point of running a controlled test. Richer product content, written by AI, gets products seen and clicked more on average. It is not a golden ticket. Whether it works for your products is something only your own results can tell you, which is why we keep a control running and keep watching it.

What comes next: a control that never switches off

The test didn’t end when the award was announced. As we roll AI-written content out more widely, we keep a randomised holdout in place: a share of every retailer’s products, chosen at random, that stays on its original content. It’s the same fair split the test used, run permanently.

That matters because the places this content shows up are multiplying. Google is putting product data to work across more of its channels, from Shopping to AI Max and its AI-led search experiences, and the same highlight might reach a shopper through any of them. Channel reports can’t tell you what the content itself is worth. A holdout can. As long as part of the catalogue stays untouched, we can measure the difference the content makes, wherever Google surfaces it, and keep proving it rather than assuming it.