Featured case study

Building Shutterstock’s first conversational search and learning where AI fit

Executive summary

Shutterstock customers often began with creative briefs, moods, and conceptual ideas that did not map cleanly to keywords. I inherited an early AI Search concept and independently led a six-month rebuild across the marketplace.

Early experiments showed a 2.3% conversion lift, but post-launch behavior revealed a more nuanced outcome: AI Search helped some customers express intent, yet most licensing still happened through traditional search. The project became as much about designing the handoff between search models as designing the AI experience itself.

Timeline

6 months

My role

Sole senior product designer

Core team

Product manager · Frontend engineer · 2 backend engineers · Engineering manager

Initial signal

+2.3% conversion lift

Promising result from early conversational search experiments.

Post-launch reality

0.5% daily usage

AI Chat Search remained a niche path within marketplace search.

Product learning

15.8% returned to standard search

Customers frequently used AI for exploration, then licensed through the familiar search flow.

01

The Product
Bet

Let customers describe intent instead of translating it into keywords

Shutterstock search asked customers to compress moods, concepts, and project briefs into exact keyword queries. Weak initial queries produced irrelevant results, repeated searches, and a slower path to licensing.

Leadership saw conversational AI as a way to reduce cognitive load, improve relevance, and differentiate the marketplace. It also addressed search friction affecting time to value, customer retention, and revenue.

Customer problem

Strict keyword matching created unguided search loops, especially when customers started with conceptual or ambiguous creative needs.

Business problem

Low-quality searches increased acquisition friction and slowed customers from finding content they were confident licensing.

Ecosystem problem

AI Search had to work across the homepage, search results, and product pages without making customers guess which search model to use.

02

Ownership &
Leadership

Taking the project from early validation to launch

I inherited an initial workshop and concept from another designer. After that handoff, I owned the product design process independently: interaction design, research, prototyping, facilitation, stakeholder alignment, and implementation support.

I partnered with the product manager and engineering team while aligning direction with the VP of Design, Senior Director of Engineering, and Director of Marketplace.

  • Defined the overall AI Chat interaction model
  • Led design explorations with Engineering
  • Used research and prototypes to navigate product and technical ambiguity
  • Facilitated tradeoff decisions and scope reduction
  • Connected AI Search to the broader marketplace funnel
03

Product
Constraints

Designing an ambitious interaction model within launch realities

The experience needed to feel modern and useful even though the first release had meaningful technical and scope limitations.

Response latency

Search results took too long to load. Engineering improved performance before launch. The interface also used additional progress animation to make the remaining wait easier to understand.

Limited asset support

The launch supported images and video only. Multimodal prompting with reference images or documents remained out of scope.

Five-response memory

The system could not preserve conversational memory beyond five chat responses, limiting the depth of iterative exploration.

Ambiguous prompts

When customers did not specify image or video, the system defaulted to image results so the experience could continue without another blocking question.

04

Research &
Prototyping

Clarifying how AI should fit into familiar search behavior

I used unmoderated usability studies, sentiment research, and interactive Figma Make prototypes to test how customers understood the chat experience, moved between AI and standard search, and interpreted transitional states and animation.

What changed the direction

  • Customers treated AI Search as another way to search, not a replacement
  • Seeing results quickly mattered more than maintaining a long conversation
  • Losing chat context when viewing an asset broke trust and continuity
  • Customers expected familiar controls such as editing, copying, stopping, and rating responses
05

Strategic Product
Decisions

Design the relationship between search models, not a standalone chatbot

Four decisions connected conversational search to the marketplace journey and protected the workflows customers already trusted.

Position AI Search as an alternate path instead of the new default

AI and standard search remained available side by side. Tabs supported direct switching and preserved the current search context, so customers did not need to start over when changing modes.

Make the relationship between search modes explicit

Put visual results ahead of the conversation

I moved away from a ChatGPT-style thread and designed a split workspace with chat on the left and the asset grid on the right. Customers could refine language while continuously evaluating content.

Keep creative results at the center of the task

Protect context across the marketplace funnel

Viewing an asset originally created a new chat and erased the prior search. I changed the flow so customers could move between AI results and product pages without losing conversation state.

Return customers to the same exploration state

Meet baseline AI expectations without overbuilding

The first release supported editing and copying prompts, force-stopping responses, feedback ratings, chat history, and reuse of earlier searches. This provided enough control to feel familiar without expanding every history-management feature.

Provide familiar control over iterative prompting
06

Cross-functional
Direction

Turning exploration into a buildable product direction

I facilitated explorations with Engineering to understand what the system could support. I then used research and interactive prototypes to make behavior, animation, and tradeoffs tangible for product and leadership stakeholders.

This shifted conversations away from abstract opinions about AI and toward observable customer behavior, technical feasibility, and the marketplace journeys the team needed to protect.

Figma Make prototype used to evaluate interaction and motion
07

Final
Experience

The launched experience added conversational discovery to Shutterstock’s universal search while keeping it connected to the marketplace. Customers could describe intent naturally, evaluate assets beside the conversation, refine prompts with familiar controls, revisit prior searches, and inspect product details without losing context.

Final AI Search experience
08

What Launch
Data Revealed

A promising experiment did not become a primary licensing path

Early tests showed conversion potential, but post-launch behavior was more mixed. AI Chat remained a niche exploration tool, and most commercial behavior still happened through traditional search.

0.5%of daily marketplace searchers used AI Chat Search
3.3%of AI Chat users licensed through AI Chat, versus 7.5% for traditional search
15.8%left AI Chat and later licensed through traditional search
16%left AI Chat before seeing results, with latency and waiting-state confusion as likely contributors

AI Chat users generated a 0.15% order rate versus 0.1% for traditional search, but only 136 orders were observed. This was too small a quantity to treat as a reliable success signal. Usage also skewed toward returning visitors. First-time visitors represented 36% of AI Chat users versus 58% of traditional search users.

09

Retrospective

The strongest learning was that customers did not need AI Search to replace traditional search. They needed it to help express intent, then hand off cleanly to the marketplace workflow that already supported confident licensing.

What the project proved

  • Conversational input could reduce the burden of translating creative intent into keywords
  • A results-first workspace fit Shutterstock better than a chat-first thread
  • Context preservation was essential across search and product pages

What I would prioritize next

  • Reduce response latency before expanding feature scope
  • Make the best-fit AI use cases clearer at entry
  • Treat movement into standard search as a designed handoff instead of a failed AI session
  • Evaluate adoption and licensing behavior alongside early conversion signals

Explore more work

Shutterstock Studios homepageRedefining Shutterstock Studios’ digital presenceAligning an expanded creative offer with a scalable digital foundationView case study Shutterstock homepage redesignRepositioning Shutterstock for the AI eraUnifying Search, AI Search, and Generate in one starting pointView case study Shutterstock contributor workflowIncreasing submission velocity by 23%Using AI-assisted workflows to drive 27% library growth in the measured timeframeView case study