Search Finds Pages. Chat Finds Answers.

I typed design systems at scale into the newsroom archive on my own site and expected a paragraph back. Titles. Links. Types. No synthesis. No footnotes. For a second I thought the product had failed me. Then I remembered I had built it that way on purpose.

The archive on gaganmalik.io runs keyword search over titles and bodies: Postgres full-text search, ranked matches, twenty results at most. It is built to find pages. On the same site, Ask runs retrieval over embeddings, pulls chunks from writings and case studies, and returns an answer with numbered citations you can inspect. It is built to find answers. Josh Brown at Nielsen Norman Group reported the same split in a February 2026 study: participants reached for generative AI to explore and synthesise, but returned to traditional search when accuracy and trust mattered. nngroup-infoseeking-2026 Those are not two labels for the same feature. They are two different instruments for two different moments in how someone decides whether to trust you.

Search Finds Pages

When you know roughly what you are looking for, or you want to see everything the corpus contains on a term, search should return documents, not a synthesis. A hiring manager typing Apple or compliance or RAG in the archive is not asking the site to think on their behalf. They are asking which artefacts exist so they can choose what to open. Keyword search honours that contract. It ranks what matched. It does not pretend it read the room.

That contract breaks when search is dressed up as intelligence. You type a phrase. You get one confident paragraph with no visible sources, or worse, sources that were never retrieved. Leonid Rozenblit and Frank Keil documented in Cognitive Science in 2002 that people confuse sounding fluent with understanding until asked to explain the mechanism. rozenblit-keil-2002 You have traded browse for assertion: the same comfort I named when a clean UI feels like work but skips the mechanism. gaganmalik-in-defence-of-thinking I wrote about that failure mode when I built a personal model on my own corpus: retrieval that returns you to what you already argued is a different habit from an answer tab that lets you ship conclusions you have not earned. gaganmalik-personal-llm Margaret Mitchell, formerly of Google's AI ethics team, warned in MIT Technology Review in 2023 that citations can give wrong answers an air of correctness users never double-check. mit-tr-trust-ai-search Search keeps the user in charge of which page to trust. The page keeps the burden of proof where it belongs.

Chat Finds Answers

Ask exists for a different job. Someone arrives with a sentence, not a keyword: What did you do at Apple? or How do you think about design systems? or Can we work together on pricing? They want a direct response they can test in minutes, with footnotes. Grounded chat, retrieval-augmented generation with citations, is the right shape for that moment. The model composes; the corpus constrains; the user can click through to the lemma.

That only works if you treat chat as product, not as a search bar with adjectives. Rate limits. Query caps. Modes that route contact questions differently from opinion essays. A step ceiling on tool calls so the agent does not loop forever. Citation markers stripped when they do not match retrieved chunks, because researchers presenting at ACM SIGIR in 2025 found that up to fifty-seven per cent of RAG citations can be post-rationalised rather than genuinely derived from retrieved context. sigir-rag-faithfulness-2025 Brown's Nielsen Norman Group study found the same gap from the user side: even when chatbots supplied citations, participants could not tell which claims mapped to which sources. nngroup-infoseeking-2026 I have been sharpening that stack for months: query expansion, reranking, golden-set evals you can regression-test like any other feature. The interview question is not "Do you have AI?" It is "Can you show me where the answer came from?"

Browse Builds Trust

There is a third mode people forget when they argue for one box to rule them all: browse. Reading a full case study. Listening to an essay on a commute. Following an argument across sections until you feel the warrant, not just the conclusion. Euclid's permission stack only works if someone stays in the demonstration long enough to see the postulates. gaganmalik-demonstrate Search on design systems at scale gets you the page list from that first query: titles, links, types. Ask gets you the paragraph you can test in eight minutes. The full essay is whether you still trust the work after forty minutes of reading.

When Google rolled out AI Overviews in 2024, publishers worried that paragraph-length summaries at the top of results would choke off the traffic that funds original reporting: users would read the synthesis and never open the page. nyt-google-ai-publishers Design portfolios compress browse into thumbnails and hope evaluators infer depth. That is why I ship long-form writings, listen mode, and case studies with problem, approach, and outcome spelled out, not only a chat widget. A recruiter with eight minutes should be able to ask a question and get cited sources. A peer with forty minutes should be able to read the essay that names the tradeoff. Collapsing those paths into one generative surface optimises for the eight-minute win and taxes the forty-minute proof.

The Wrong Instrument

The counterargument is fair and I have made it myself: users do not want to learn your information architecture. They want one place to type. I tried merging archive search and Ask in a wireframe: one input field, empty archive silently calling the model. I rejected it. The merged flow looked cleaner in a screenshot than two honest surfaces do; it also hid which instrument the user had asked for.

The industry bet is still real. MIT Technology Review named generative AI search one of its ten Breakthrough Technologies for 2025 because it returns concise answers instead of link lists. mit-tr-genai-search-2025 Sundar Pichai told The New York Times in May 2025 that Google's AI Mode was a total reimagining of search. nyt-google-ai-mode Perplexity bills itself as an answer engine with inline citations; a PCMag reviewer who switched from Google for a week said it excelled at transparently sourced research queries. pcmag-perplexity

Assume one box is always better: then archive search and Ask should merge tomorrow, and every empty archive state should silently call the model. Nielsen Norman Group usability research found that site chatbots which recreate search in a conversational interface are often slower than established search, filters, and navigation, adding friction to problems already solved. nngroup-site-chatbot Empty keyword search often means the user phrased a question where the system expected a term. The fix is not to hide search behind chat. The fix is a handoff: you searched; here are pages; if you meant a question, continue in Ask with this query pre-filled. Different instrument, explicit bridge. That is product design, not feature sprawl. Maybe your corpus is tiny and browse does not matter yet. At roughly five hundred sessions a month, I still split the systems because the discipline scales before the traffic does. When the corpus grows, you will already know which index serves which job. When evaluators ask how you think about AI search, you can point at live surfaces instead of a slide.

What I Will Say in the Room

I ship search, chat, and browse on one domain because hiring managers, peers, and clients do not share one job to be done. The craft is naming the contract each surface keeps and encoding it in guardrails, not picking the smartest model. Last week I almost merged them anyway, because a single AI box looks cleaner in a screenshot than two honest ones.