There is no neutral list
Baymard’s long-running benchmark places cart abandonment at 70.19 percent. Among 344 top-grossing US and European sites, 65 percent of checkout experiences were “mediocre” or worse and only 2 percent were “good”; its combined testing estimates that a large site could potentially improve conversion by 35 percent through better checkout UX. Design is already a revenue variable.
Choice architecture is the environment in which people decide. The UK Competition and Markets Authority’s evidence review describes how ordering, defaults, framing, scarcity, and friction can affect consumers and competition. Generative answers create these elements dynamically, often without a designer reviewing the exact configuration.
Helpful structure and dark patterns use the same materials
Customers need reduction. No one benefits from a model dumping every possible alternative in arbitrary order. Grouping, ranking, and defaults can make complexity manageable. The ethical difference is not whether the design influences. It is whether the influence serves a legitimate customer goal, preserves material information, and allows meaningful correction or exit.
A recommendation can be justified with explicit criteria the customer can edit. A default can be reversible and safe. A shortlist can say what it excluded. A call to action can state the consequence. These designs influence behavior while increasing agency. Hidden sponsorship, asymmetric effort, or fabricated urgency do the opposite.
Accessibility changes choice
W3C guidance on clear content makes an important connection: understandable layout, familiar language, and small chunks reduce cognitive burden. When one option is clear and another is dense, the interface has altered choice even without changing price or quality. Complexity itself becomes a tax, often paid unequally by people with less time, language fluency, or domain knowledge.
This is why disclosure volume is not enough. A material limitation in a paragraph below the fold does not compete fairly with a vivid headline recommendation. Answer audits should consider prominence, sequence, and the mental work required to integrate the information, not merely whether the fact exists somewhere in the output.
Conversion is a real metric and an incomplete one
Karaca’s newly published client case makes the revenue link explicit. Its AIDA shopping assistant was shown to 20 percent of users — two million people — and reportedly doubled conversion versus search while reaching five times the rate of unaided sessions. The team also reported a 97.5 percent reduction in session cost before launch after guardrail and A/B-testing work. These are company-reported comparisons, not an independent randomized estimate.
The case shows why conversion is real and incomplete. A generated adviser influences judgments about need, fit, and risk before checkout. Use a metric chain: comprehension of alternatives, confidence calibrated to evidence, immediate action, later task success, regret, return or cancellation, and support cost. A one-session lift is not valuable if the answer moves customers into a plan they cannot successfully use.
Audit the choice architecture you already have
For a recommendation answer, create an architecture inventory: option order, default, labels, evidence prominence, edit controls, friction, and exit. Test the existing response against a transparent criteria-led version and, if appropriate, a customer-configurable version. Use identical underlying options and prices.
Measure choice, accurate understanding, time, confidence, and a delayed or simulated outcome such as successful configuration or preference stability. Include a conflict-of-interest condition to see whether the design makes sponsorship visible. Influence cannot be designed away. It can, however, be made inspectable and tied to a customer outcome the business is willing to defend.