(3) Feed | LinkedIn
By Vanessa Hung · August 19, 2025 · Curated by George's Blog
About 75% of the AI optimizations you make to your listings have zero measurable effect on sales rank or sales.
The other 25% can trigger dramatic spikes in market share.
That insight comes from a study by researchers using ACES, a controlled “mock” marketplace that let them randomize layout, positions, prices, ratings, reviews, and badges while testing Claude Sonnet 4, GPT-4.1, and Gemini 2.5 Flash. In that sandbox they could causally attribute how platform levers and listing attributes steer AI agents’ choices. Thanks to Nate Pegram for sharing this study.
👉 Key detail: these results come from ACES, not a live retail site. That is the point. By isolating the choice step, the study shows which levers can create those “25% moments” when AI agents are the shopper.
What has triggered the 25% spikes so far:
• Single, one-shot listing edits: In 25% of category–model pairs, a single description change produced large gains, like +21.8 and +23.6 depending on the product. This was not an iterative campaign, just one pass.
• Platform badges: A neutral item at roughly 10% share jumped past 20% and even 40% when it carried an “overall pick” style badge. A “sponsored” tag, with position held constant, often lowered selection.
• Position on the page: Moving the same product to the top row multiplied pick rates. One model showed a 5x lift from bottom right to top row.
•Price: Lower prices helped, but less than many expect. The impact changed by model and by the other signals on the page.
This gets so interesting because the study translated these levers into “price headroom.” In plain terms, a top row spot or an “overall pick” badge can be worth as much as charging a lot more and still getting chosen. In some cases, moving up one row or adding the badge rivaled the effect of a large price cut.
Most edits do nothing. Then one change flips an internal decision, and your share moves right away.
✏️ So what should you do if you sell on places like Walmart, where they built everything with Chatgpt?
• Treat optimization like testing an AI decision maker, not just pleasing human readers.
• Run tight, single-change experiments around badges, structured attributes, rating and review volume, price thresholds, and position levers you can influence through ads or merchandising.
• Look for step changes, not tiny trends, and expect results to shift when the platform updates its model.
In case you are an Amazon seller, that's a different story.
Amazon built its entire ecosystem, platform, and LLM. So, for the Amazon ecosystem, you have Rufus, which we kind of know what it looks for in optimization.
I will be talking about what Amazon says Rufus considers when recommending your listings.
If you’d like to read the full study, I’ve dropped the link in the comments.
#AIAgents #MarketplaceStrategy #Optimization #Ecommerce