Almost every DTC brand running on Shopify has at some point turned on "Sort by Best Selling" for their collection pages and felt good about it. It is an intuitive decision. Your best-selling products are your best-selling products. Put them first. Let the winners win.
The problem is that "best selling" is a historical aggregate metric. It tells you which products have sold the most units over some lookback window. It says nothing about which product is most likely to convert the specific person browsing your collection page right now, given how they arrived, what their session behavior looks like, and what their implicit purchase intent signals are.
These two things, historical sales rank and real-time conversion probability, diverge more than most brands realize.
The Aggregation Trap
When you rank by units sold, you are implicitly assuming that your past buyers and your current visitors share the same preferences and intent profiles. In many DTC businesses, that assumption holds reasonably well for returning customers who came back to reorder something they already love. It holds very poorly for new visitors arriving from diverse acquisition channels.
Consider a footwear brand with a men's running shoes collection. Their historical best seller is a cushioned long-distance trainer. It has sold well for two years, has strong reviews, and consistently drives reorders from a loyal running community. So it ranks first in the collection.
Now consider two different visitors landing on that collection page in the same hour. The first arrives from a Google search for "trail running shoes under $120." The second arrives from a Meta ad that featured lifestyle imagery of the brand's lightweight racing flat. The optimal first product to show each of those visitors is almost certainly different, and almost certainly neither one is the cushioned long-distance trainer that dominates the aggregate sales rank.
The search visitor has explicit intent for trail running. The social visitor was primed on the racing flat aesthetic. Serving them the same static grid, ranked by aggregate best-sellers, leaves both of them scanning for relevance instead of immediately seeing a product that matches their arrival context.
Where Units-Sold Ranking Goes Wrong in the Data
The divergence between sales rank and session-level conversion probability tends to be largest in a few specific situations.
The first is when a brand has seasonal hero products. A sunscreen brand with a heavily promoted summer SPF 50 product may have that product sitting at rank one based on last 90 days sales, well into fall. That product converts beautifully when the visitor intent profile matches summer sun exposure. When fall visitors with moisturizer intent arrive, the rank-one placement of a summer SPF creates a relevance mismatch that adds friction to the session.
The second is when paid acquisition channels bring in visitor populations with different category orientations than the organic or direct traffic base that built the historical sales rank. This is the situation we encounter most commonly in practice. A brand that grew organically in one niche, say minimalist hiking gear, and then scales via Meta ads targeting adventure travel audiences, will find that the Meta traffic converts differently than the organic hiking community. The product that converts the organic visitor best is not the one that closes the Meta visitor.
The third is when newer products have strong conversion rates but insufficient sales volume to rank well historically. A product launched three months ago may have a session-to-add-to-cart rate of 9% for visitors who see it in position 1 or 2. But because it has only sold 300 units, it ranks below five other products that have been in the catalog for two years. The catalog hides a strong converter behind the weight of history.
Conversion Rate Per Session vs. Units Sold: A Better Lens
A more useful metric for collection page ranking than units sold is product conversion rate by traffic segment. For each major traffic source category, compute which products have the highest session-to-add-to-cart rate when they appear in the top 3 grid positions. That tells you what actually converts for each visitor type, rather than what has historically sold to the aggregate population.
This analysis often produces lists that look quite different from your historical best-sellers. Products that are moderate overall sellers but strong converters for specific traffic types become visible. Products that look like best-sellers in aggregate but are being driven by one concentrated channel cohort become less universally dominant.
We are not suggesting that units-sold data is useless in a merchandising model. Historical sales data carries real signal: it reflects market validation, price acceptance, return rate patterns, and product quality at scale. The point is that it should be one input in a ranking model, not the sole ordering criterion, and it should be weighted against real-time session-level conversion signals rather than applied statically across all visitor types.
A Practical Illustration: Supplements Brand, Two Traffic Channels
A growing supplements brand with a catalog of about 45 SKUs had their collection page sorted by 60-day units sold. Their rank-one product was a protein powder that had a strong reorder base among gym-focused returning customers. It was a legitimate best-seller by volume.
When they segmented session data by acquisition channel, a different picture emerged. For organic search traffic arriving via fitness and performance keywords, the protein powder was indeed the top converter in position 1. But for Meta ad traffic arriving from creative that featured a daily wellness routine theme, the rank-one converter was an omega-3 and vitamin D bundle, which ranked 8th in the default collection. Sessions where that bundle appeared in position 1 or 2 had significantly higher add-to-cart rates for the Meta cohort than sessions where they encountered the protein powder first.
The static bestseller grid was well-optimized for one channel. It was underperforming for the other channel in a way that only became visible once the data was segmented by traffic source rather than pooled.
The Direction This Points
The logical endpoint of this analysis is dynamic collection page ranking: different visitors see different product orderings based on real-time session signals. That is what Cartlyzer does at the inference layer, using clickstream behavior from the first 60 to 90 seconds of a session to rerank the grid toward the products most likely to convert that specific visitor.
But even without dynamic ranking, the analysis is worth running as a diagnostic. Pull your conversion rates by traffic segment and compare them to your collection page default order. If there is a material divergence between what converts by channel and what ranks by units sold, you have a static merchandising problem that a simple collection variant or scheduled rule can address, even before any real-time system is in place.
Your best sellers are real. They represent what the market has validated. What they do not represent is what the visitor in front of you right now came to buy.