Enterprise Search

I designed for an answer years before AI could give one

Three years on Meta's internal search: an ambitious overlay, a results-page rebuild, then the AI answer once Metamate needed our infrastructure. I led design across all three.

Problem:

Meta's keyword search was weak enough that people stopped typing more than a couple words. A recent redesign had changed the visuals, not the problem.

Outcome:

A results page built around how people actually search, which became the foundation for an AI answer that outperformed much of the industry doing the same thing at the same time.

Problem:

Meta's keyword search was weak enough that people stopped typing more than a couple words. A recent redesign had changed the visuals, not the problem.

Outcome:

A results page built around how people actually search, which became the foundation for an AI answer that outperformed much of the industry doing the same thing at the same time.

Role:

Sole designer

Timeline:

2021-2023

Role:

Sole designer

Timeline:

2021-2023

An overlay for everything

Connected Search was a Spotlight-style layer for every internal tool. It found real traction with people who had a specific need, like code search or employee lookup.

It never won broad adoption. It competed against a browser address-bar override with no type-ahead, but one already where people were.

…

Rebuilding the pagearound what people actually knew

Most people knew what kind of thing they wanted before they knew what to call it. I redesigned filtering around that instinct.

Filters that adapt to intent

Gen-1 : New chat start screen & menu

Selecting a document-type facet dynamically updated every sub-filter below it, instead of showing options that mostly didn't apply.

For high-confidence entities, I prototyped widget-style result cards — closer to a knowledge panel than a document list. Same instinct AI answers would later formalize.

Curated, authoritative results

Gen-1: Start AI chat from sticky CTA on articles

Then Metamate needed us and the answer became real

Metamate needed our retrieval infrastructure for RAG. That opened the door to a real AI answer, built on a year of filter thinking already done. Placing the answer at the top, not in type-ahead, is the decision that let it survive short-query habits.

Before

Roughly 80% of queries were three words or fewer — a habit that risked breaking a GenAI feature built for longer input.

After

I kept the composer and solved the real problem: a radial glow entrance built from the assistant's own palette, resolving on arrival and settling out of the way. Engagement came back level or better.

Results

−43%

Time-to-find: 12.8s to 7.25s, after the filter and results-page redesign.

80%+

Zero-click AI answers: Most AI answers resolved a search without a further click.

Jonathan LoPorto