New research finds that AI tools substantially improve how shoppers discover products, but that reviews remain the factor that closes the sale. The finding is less surprising than it looks. Discovery and conviction are different problems, and the technology that solved the first has not solved the second (Modern Retail research).The gap between those two steps is where most of the commercial value now sits.
Discovery is no longer the constraint
For most of the last decade, the binding constraint in ecommerce was finding the right product among too many options. Search, filters and recommendation feeds gradually reduced that friction, and AI assistants now compress it further by accepting natural language and returning a shortlist. When discovery becomes cheap, the marginal value of another discovery feature falls sharply.
Why shorter sessions mislead dashboards
The consequence is a shift in where traffic comes from. A shopper who arrives via an AI answer has already been pre-filtered, which means the session starts later in the funnel. Sessions get shorter and intent gets narrower, and the metrics that made sense for browse-heavy traffic start to mislead. Teams that keep optimising for time on site will read a healthier funnel as a weaker one.
Reviews carry the conversion weight
Reviews work because they answer the question the product page cannot: what actually happened when someone like me bought this. That question is resistant to automation, because it depends on lived experience rather than specification. An AI assistant can compare materials and dimensions, but it cannot tell a shopper whether the item survived a winter or fit as expected (consumer trust signals).
The weight of reviews also explains why AI-generated product copy has not moved conversion much. Copy improves comprehension, but comprehension was rarely the blocker once a shopper reached the page. What blocks the purchase is residual doubt, and doubt is resolved by evidence of other people's outcomes, not by better adjectives.
What a useful review corpus contains
Three elements matter more than star ratings: specificity about use case, disclosure of duration of ownership, and detail about failure. A review that says the item was fine adds little. A review that says it held up for eight months of daily use in a humid climate, then the seam failed, gives a shopper something to weigh. Structured prompts collect this far better than an open text box.
The review corpus is now an operating asset
If reviews are the conversion lever, they belong in operations rather than in marketing. That means monitoring review velocity by SKU, flagging categories where review volume is thinning, and treating a drop in review count as a leading indicator of conversion decline. Most teams discover this only after a product has already lost momentum.
The same corpus serves two channels
There is also an exposure dimension. Reviews are among the most frequently cited content by AI assistants answering product questions, which means a thin review base does not just hurt on-site conversion, it also weakens how the brand appears in generated answers. The same corpus serves two channels, and investing in one improves the other at no extra cost.
How brands should audit their review signal
A practical audit has four steps. Count reviews per SKU against category median, then check recency, since a five-year-old corpus signals an abandoned product. Then check length distribution, because short reviews carry little decision value. Finally check whether reviews mention the specific concerns that drive returns in that category, which is where the highest-value content sits.
Fixing gaps is slower than it appears. Soliciting reviews after delivery raises volume but often produces low-detail responses; prompting for specific use cases produces fewer but more useful entries. The trade-off should be decided per category, since high-consideration goods benefit from depth while commodity goods benefit from volume, and applying one policy everywhere wastes effort on the wrong metric.
