In 1989 Marcia Bates showed that people search by gathering, shifting, and revising as they go. Her ideas explain what AI still misses and what is still relevant from research done 47 years ago.
Much of what product teams are learning about artificial intelligence search, an information scientist wrote down before the web existed. Marcia J. Bates spent her career at UCLA studying how people search and her 1989 paper, The Design of Browsing and Berrypicking Techniques for the Online Search Interface, argued that real searches wander, change shape, and collect value in pieces.
She aimed it at library databases that assumed one query and one best result; AI turns this on its head.
The chat box makes the same assumption, only faster.
Bates is still relevant in the age of artificial intelligence, arguably more than ever. Large language models changed what a system can answer but they didn’t change how a person goes about asking, which still means starting vague, learning while reading, and changing direction midway.
I had the pleasure of meeting her at the IA Summit in Chicago in 2018 — she most politely asked if we could have lunch and I brought a few others. It became a great conversation where I still learned concepts that I apply today, and they are incredibility relevant to what we’re doing with AI back when my company was only applying machine learning.
I’ve used her concepts to frame search and artificial intelligence summarization in legal tech software. Anyone designing search or summarization should be thanking her.
Five of her ideas hold up, each tied to a decision you’re making now.
Support Evolving Searches
Bates built the berrypicking paper as an attack on the classic retrieval model, which treats a search as one query matched once against a database to produce one best set of results.
Her objection fit in one line: “typical search queries are not static, but rather evolve.”
People begin with one reference or a narrow slice of a bigger topic, read what comes back, and let each find reshape the question.
She named the bit-at-a-time collecting after picking huckleberries scattered through a forest. Marti Hearst later made the model a standard reference for interface designers in Models of the Information Seeking Process.
The need itself shifts as the person learns, and that separates berrypicking from search methods that look similar from the outside:
- Known-item lookup. You know what you want, such as a fact or a file, and one query finds it or doesn’t. You can accomplish this usually in a single term.
- Iterative refinement. You rephrase to hit a fixed target, the job Gerard Salton’s relevance feedback automated. The words change, but the need doesn’t.
- Browsing. You glimpse, sample, and pick within one place. Bates treated it as one move inside a larger search, not the same thing.
- Berrypicking. The need evolves, and the answer arrives in pieces from many sources rather than as one final set.
That last difference is why a successful search is so hard to define:
- The classic model let researchers score one result set for precision and recall.
- A berrypicking search has no single set to score, and Bates called knowing when to stop one of the fundamental problems of search strategy.
- A thumbs-up on one chat answer measures the last berry, not the picking.
Sometimes what matters most is what isn’t found.
An empty patch tells a berrypicker something, whether it’s no earlier case on point, no conflicting clause in a contract, or no research on a question. A system that always produces a confident answer hides that signal, because it never lets the absence show.
The chat window rewards the user who already knows exactly what they want, which is the user who needs the least help.
Language models repeat the classic assumption.
In LLMs Get Lost in Multi-Turn Conversation, a 2025 study of more than 200,000 simulated conversations, every top model tested did worse when a task arrived in pieces, with an average drop of 39 percent across six tasks. The models guessed early and didn’t recover.
The design job is to support the person whose question is still forming. Show what has shifted since the first prompt, and let people revise early assumptions without starting over.
Bates even sketched the tooling in 1989: a temporary store where searchers could flag findings as they went.
Expose Search Tactics
A decade earlier, Bates published Information Search Tactics, which won her field’s best paper award the following year. A tactic, in her definition, is a move made to further a search.
She named 29 of them in four groups, including these:
- Check. Compare the current search against the original request.
- Pinpoint. Narrow to the most precise terms.
- Parallel. Broaden with synonyms and related terms.
- Trace. Mine what you’ve already found for new terms.
- Record. Track the trails you followed and the ones you meant to revisit.
Her premise was that naming a tactic makes it easier to reach for. Reference librarians used many of these moves without names, which left beginners no easy way to learn them.
The web repeated the problem.
People had to learn to search like Google, stripping questions down to two or three keywords and adding quotes or a minus sign when results went wrong. They’re doing the same with chat now, working out by trial and error which phrasings get a useful answer.
A blank prompt box ships with none of them, so people improvise.
Nielsen Norman Group documented one improvisation in Accordion Editing and Apple Picking: Early Generative-AI User Behaviors.
Users copied part of an earlier response into a new prompt by hand, because the interface gave them no other way to point at it. It’s berrypicking with copy and paste as the only tool.
That’s a search behavior, because chat and search are closer than they look. In both, a person puts a need into words, the system matches it against stored information, and the person judges what comes back and tries again.
Retrieval-augmented assistants run a search behind the scenes and summarize the results, so the biggest difference is the shape of the reply: a list of links in one, a paragraph in the other.
Everything Bates learned about searching also applies to chatting in spades.
Missing controls matter most for experts, who read a bad result set as a clue rather than a verdict. They tighten, exclude, widen, and re-sort in small passes until the set matches what they meant, and they decide when it’s done. In Investigating the Querying and Browsing Behavior of Advanced Search Engine Users, Ryen White and Dan Morris found marked differences between people who used query operators and those who didn’t, in how they queried, clicked, browsed, and fared.
Faceted search is the interface version of those moves.
In a study reported in Hearst’s chapter Integrating Navigation with Search, 91 percent of 32 art history students preferred a faceted image interface to a plain one, even though it was slower.
A chat reply hides exactly those facets and filters, when most of Bates’s 29 tactics could become buttons nobody has to phrase.
Choose the Automation Level
In 1990, Bates asked the question every agent team is now answering under deadline: Where Should the Person Stop and the Information Search Interface Start? Her answer was a scale.
- At level 0 the system offers no help.
- Level 1 explains possible moves
- Level 2 carries them out on command
- Level 3 watches the search and suggests next steps
- Level 4 runs the whole search, either explaining its choices or just returning results.
In product terms, the levels are easy to spot:
- Level 1 is a help panel that lists search operators
- Level 2 is a filter that runs when you click it
- Level 3 is a “did you mean” correction or a suggested follow-up question.
- Level 4 is the research agent that disappears for several minutes and comes back with a finished report.
She argued that research had fixated on level 4, while many searchers wanted the speed of automation and control of strategy. Jakob Nielsen called generative tools AI: First New UI Paradigm in 60 Years in 2023, describing a reversal of the locus of control in which users state the outcome and the computer picks the method.
Bates had mapped that reversal decades earlier, and put most of the useful ground short of the far edge.
Most AI products jump straight to level 4 and call the result an agent. That’s not right.
Skipping the middle skips the levels where trust gets built. People defer to the machine even when it’s wrong, which I document in the Automation Bias entry on UX Guidelines, a reference site I maintain. Pick the level per task, and let people move the dial.
Fix the Content Layers
Her 2002 paper The Cascade of Interactions in the Digital Library Interface lists the layers under any search experience: infrastructure, content, metadata, indexing, and the search engine, with the interface and the person at the top.
Decisions at each layer ripple upward, and good interface design can’t rescue weak choices further down.
Swap “digital library” for “retrieval-augmented assistant” and little changes. Above the model is the chat window your team keeps polishing. Below it are document stores, chunking rules, and permissions someone set up years ago for another purpose. Good metadata is what sets context for the model, telling it what a document is, who it applies to, and whether it’s still current. The demo runs on 20 clean documents, and production meets a decade of duplicate policies and unlabeled drafts.
A semantic layer goes a step further.
It defines what terms mean and how they relate, so “customer,” “client,” and “account holder” resolve to the same thing, and the model works from shared definitions instead of guessing.
Bates proposed something close to this in 1986: a cluster thesaurus that gathered every variant of a concept under one entry, so a search on any term could reach them all.
A fluent answer built on a bad index is still a bad answer.
Gartner predicted in Lack of AI-Ready Data Puts AI Projects at Risk, a February 2025 release, that through 2026 organizations will abandon 60 percent of artificial intelligence projects that lack ready data. That’s a forecast, but the same release reported that 63 percent of organizations lacked the right data practices or weren’t sure they had them.
This is the uncomfortable part for designers. The quality of a search result, or of a generative answer, is the user experience, and no polish around it makes up for a weak one.
Your influence has to reach down the cascade into labeling and indexing, and Bates argued the specialists for every layer had to design together.
Design for Browsing
Bates returned to browsing in 2007 with What Is Browsing — Really? A Model Drawing From Behavioural Science Research. Browsing, she argued, is a series of glimpses: you take in a field of view, select something, examine it, then keep it or drop it, and repeat.
She traced it to the exploratory behavior animals use to find food and shelter, which is why nobody needs to be taught how.
Similarly, I’ve been through 65 airports, partly to see how cities tell people where things are. The good terminals let you take in the whole hall at once, spot the sign you need, and change course without asking anyone. That’s the opposite of waiting at the information desk for one answer.
Generative search is quietly moving people toward the desk.
Pew Research Center tracked the browsing of 900 U.S. adults in March 2025 and reported that Google Users Are Less Likely to Click on Links When an AI Summary Appears in the Results. Visits with a summary led to a result click 8 percent of the time, against 15 percent without one, and sessions ended outright on 26 percent of summary pages versus 16 percent elsewhere.
Some of that is the summary doing its job. Some is a person who never glimpsed the source they didn’t know to ask for. So keep the scene around the answer, with neighboring sources and related threads, so the glimpse still has somewhere to land.
Conclusion
None of these ideas came from studying artificial intelligence. They came from watching people in libraries, card catalogs, and early online databases, and that’s exactly why they hold.
The technology between a person and the information keeps changing. The person mostly doesn’t.
Read together, Bates’s five ideas make one argument. Search is something people do over time, using moves they choose, at a level of control they want, on top of layers they never see. Their eyes do more of the work than any query. Every piece of that is a design decision, and most current products make it by default.
She also left what I’d call the berrypicking test for any assistant on your roadmap. Does it expect the question to change, or punish the change? Does it name the moves, or make people invent them? Can the person choose how much the system decides, and does anyone own the content underneath? Is there still room to look around?
The single answer is impressive. It’s also the classic model she spent decades arguing against, with better grammar. The teams that do best with this generation of tools will design for the berrypicker, the searcher whose question hasn’t settled yet.
Action Items
- Show question history. Keep a visible summary of what the person has asked so far, and let them edit an earlier question without starting a new chat.
- Add refine controls. Put buttons for narrowing, excluding, broadening, and checking against the original request next to every answer, so nobody has to phrase those moves as a prompt.
- Set automation per task. Decide which tasks the system only suggests on and which it can run by itself, and give people a setting to change that.
- Fix metadata first. Tag documents with type, owner, date, and status, define shared terms in a semantic layer, and show the sources behind every answer, including when nothing relevant was found.
Resources
- Information and the Information Professions, Selected Works of Marcia J. Bates, volume I (Ketchikan Press, 2016). Her essays on what information is, including “Fundamental Forms of Information” and “The Invisible Substrate of Information Science.”
- Information Searching Theory and Practice, Selected Works, volume II (Ketchikan Press, 2016). Collects “Information Search Tactics” and “What Is Browsing — Really?” along with her work on search techniques.
- Information Users and Information System Design, Selected Works, volume III (Ketchikan Press, 2016). Includes the berrypicking paper and her 1990 paper on how much a search system should do for the person.
- Understanding Information Retrieval Systems: Management, Types, and Standards (Auerbach Publications, 2012). A reference volume she edited on how retrieval systems are managed, the main types, and the standards behind them.
- Marcia J. Bates Home Page. Her UCLA page, with free copies of many of the papers cited in this piece.
Berrypicking still beats the single answer for search, even in the world of AI was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.