New research finds that AI assistants widen product discovery while reviews remain the decisive signal at the moment of purchase. That combination is more uncomfortable than it sounds for brands. It means the top of the funnel is being rewritten by systems a brand cannot buy its way into, while the bottom of the funnel is still governed by user generated text a brand cannot fully control. The gap between those two facts is where generative visibility work now happens.
What the research actually shows
The finding is directional rather than precise: assistants expose shoppers to a broader set of options, and reviews then decide which option survives. Discovery expands because a conversational query can surface categories a keyword search would miss. Selection narrows because a shopper still wants evidence that other people were satisfied.
For a brand, that means two different problems with two different owners. Discovery is a content structure problem, solved by making specifications comparable and by being present in the sources an assistant reads. Selection is a reputation problem, solved by the volume and specificity of what customers write after delivery.
Discovery is not the same as consideration
Being mentioned in an answer is not the same as being considered. Assistants often list several options with a short rationale, and a brand that appears third with no distinguishing detail is functionally invisible. The detail is what makes a mention into a candidate.
Why reviews survive the shift
Reviews survive because they answer a question the assistant cannot: what happened after purchase. Return experience, durability, fit and service are all post transaction facts, and post transaction facts are exactly what a shopper worries about when the price is meaningfuls46|.
There is also a trust asymmetry. A brand describing its own product is an interested party. A hundred customers describing the same product are not. Assistants can summarize either, and shoppers still weight the second more heavily, which is why review content keeps its leverage even as the interface changes.
The specificity premium
Generic reviews have lost value. A review that says great quality adds nothing to a model that already knows the product exists. A review that says the medium fits a 32 inch waist but runs short in the sleeve gives the model something to cite and gives the shopper something to check.
Where the funnel now breaks
The funnel breaks at the handoff between assistant and merchant. A shopper who arrives from an AI answer often lands on a page that does not match the framing of the conversation, and the mismatch costs conversion. Pages built around specifications alone tend to fail here, because the conversation was about use cases.
Assistants are also inconsistent in how they cite. One may link a category page, another a single product, a third a review aggregator. Brands that only optimize one surface find their evidence missing from two of the three paths@@s47|.
Building evidence that survives summarization
Evidence survives summarization when it is structured, dated and attributable. Structured means the claim maps to a field rather than to a paragraph. Dated means the reader knows when the measurement was taken. Attributable means a link leads to the original material rather than to a restatement.
Most brand content fails at least one of the three. Statistics appear without a date, comparisons appear without a method, and claims appear without a source. Each omission is small, and together they make the content unusable as a citation, which is the practical definition of invisibility in an answer engine.
How to audit a page in ten minutes
Take the three most important claims on a page and ask where each number came from, when it was measured, and who would dispute it. If any answer is unclear, the page is not citable. The audit takes ten minutes and usually produces a short list of fixes that matter more than a redesign.
What to do this quarter
Three actions are worth prioritizing. First, rewrite product pages around use cases rather than around features, so an assistant summarizing a conversation finds matching language. Second, ask recent buyers for reviews that mention a specific scenario, which produces the specificity that models can cite.
Third, publish the method behind any proprietary number. A short methodology note is unglamorous and it is what makes a statistic quotable. Brands that skip it keep producing content that reads well to humans and disappears inside a summary. A methodology note also protects the brand internally, because it forces the team to agree on definitions before a number is published rather than after a journalist asks.
The broader point is that discovery and conversion are being separated by technology and reconnected by evidence. Whoever supplies the evidence controls the last step, and the last step is still where the money is made. Brands that treat those two stages as one campaign keep overinvesting in reach and underinvesting in the reviews and documentation that decide the final click.
