Product design question

If you were to create an image search engine, how would you do it?

Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.

What this question tests

Tests technical-adjacent product design: understanding the core system components of an image search engine and how quality would be measured.

Clarify scope: search by text query returning relevant images, versus reverse image search (search by an image), since the systems differ.
Cover the crawling and indexing layer: images need to be discovered across the web and indexed with associated metadata (alt text, surrounding page text, filename).
Cover the understanding layer: use image recognition models to extract visual features and tags (objects, scenes, colors) beyond just surrounding text, improving relevance for queries with no good text signal.
Cover ranking: combine text relevance, image quality signals (resolution, popularity, click-through), and safety filtering (removing explicit or harmful content).
Cover the user experience: fast-loading thumbnails, filters (size, color, type), and safe search as a guardrail.
Define success: click-through rate on results and user satisfaction (via survey or downstream engagement), plus a low rate of irrelevant or unsafe results as a guardrail.

Separates the pipeline clearly into crawling and indexing, visual understanding, and ranking, showing the system is more than just text-matching applied to images.
Names visual feature extraction specifically as the layer that differentiates image search from plain web search, since surrounding text alone often is not enough.
Treats safe search and content filtering as a non-negotiable guardrail, not an optional feature, given the sensitivity of image content.
Proposes concrete success metrics, click-through rate and irrelevant or unsafe result rate, tying the design back to measurable quality.

Describing image search as just text search applied to image metadata, ignoring visual understanding entirely.
Omitting safe search and content moderation as a core requirement, not an afterthought.

How would you handle a query with no good text signal at all, such as an obscure visual style?
How would you measure whether the visual-understanding model is actually improving relevance over text-only signals?