Berrypicking Still Beats Single-Answer Search in the AI Era
Marcia Bates’s 1989 search research remains the essential blueprint for designing effective AI interfaces.
1The Berrypicking Model
Marcia Bates argued in 1989 that search is not a static query-response loop but an evolving process. Users collect information in pieces, similar to picking berries in a forest. This behavior contradicts the classic retrieval model, which assumes a single query yields one perfect result, a flaw still prevalent in modern AI chat interfaces.
2AI and Evolving Needs
Large language models often struggle when tasks arrive in pieces. A 2025 study of 200,000 conversations showed that top models performed 39 percent worse when tasks were fragmented. These systems guess early and fail to recover, punishing users whose questions are still forming rather than supporting the natural, iterative nature of human information gathering.
3The Search Tactic Framework
Bates identified 29 specific search tactics, such as pinpointing, broadening, and tracing, to help users navigate information. Currently, users must improvise these moves through trial-and-error prompting. Designers should expose these tactics as interface buttons, allowing users to refine their searches without needing to rephrase complex prompts or manually copy-paste previous responses.
4Levels of Automation
Bates proposed a five-level scale for search automation, ranging from no help to fully autonomous research agents. Most AI products currently jump to the highest level of automation, which often reduces user trust and control. Designers should offer variable automation levels, letting users choose when to delegate tasks and when to retain manual control.
5The Cascade of Interactions
Search quality depends on a cascade of layers, including infrastructure, metadata, and indexing. A fluent AI answer built on poor data remains fundamentally flawed. Organizations must invest in semantic layers and clean metadata, as Gartner forecasts that 60 percent of AI projects lacking proper data practices will be abandoned by 2026.
6Browsing as Exploration
Browsing is a series of glimpses where users examine, select, or discard information. Unlike a single chat answer, which forces a linear path, effective interfaces must allow users to see the broader context. Research shows that AI summaries can reduce click-through rates, potentially limiting a user’s ability to discover diverse, relevant sources.
7Supporting Search History
To support the berrypicker, interfaces must display question history clearly. Users should be able to edit earlier prompts without restarting the entire conversation. This prevents the system from losing context and allows for the natural refinement of a search strategy, acknowledging that the initial query is rarely the final, most accurate one.
8Refinement Control Buttons
Instead of forcing users to describe their search moves in natural language, interfaces should provide dedicated buttons for narrowing, excluding, or broadening results. This approach mirrors the faceted search interfaces that 91 percent of students preferred in historical studies, providing faster and more precise navigation than a plain, text-only chat prompt window.
9Managing Automation Bias
Users often defer to machine-generated answers even when they are incorrect, a phenomenon known as automation bias. By allowing users to set the level of automation per task, designers can mitigate this risk. Providing transparency into how the system reaches its conclusions helps build trust and encourages more critical engagement with AI results.
10Semantic Data Foundations
A semantic layer defines how terms relate, ensuring that synonyms like 'client' and 'account holder' resolve to the same concept. Bates proposed cluster thesauri in 1986 to achieve this. Without such foundations, models guess at definitions, leading to inconsistent outputs that fail to meet the rigorous requirements of professional or technical search environments.
11The Value of Absence
Sometimes, the most important information is what cannot be found. An empty result set signals to a user that no conflicting clauses or earlier cases exist. Systems that always provide a confident answer hide this signal. Effective design must allow the absence of data to be visible, informing the user’s next strategic move.
12Designing for the Future
The technology between a person and information changes, but human search behavior remains constant. Teams that succeed will design for the berrypicker, prioritizing control, history, and browsing over the single-answer paradigm. The ultimate test for any AI assistant is whether it supports the user’s evolving need or forces them into a rigid, static interaction.