Sitemap
AI Advances

Ideas in depth. Where exceptional human thinking takes shape.

Deep Research vs. Broad Research: How AI turned Information Retrieval into Personalized Learning

7 min readFeb 19, 2026

--

Press enter or click to view image in full size
A surreal 3D rendered image depicts a metallic funnel twisting downward from blue at the wide top to glowing golden at the narrow base. Countless blue question mark particles cascade into the funnel’s opening, while white chat speech bubbles spiral along its shiny, reflective surface against a gradient gray floor.
(This image was generated using AI technology.)

On iteration, depth, and learning at your own pace

“Just Google it.”

When did you last say this? When did you last hear someone else say it? I had to stop and check my own reflexes. The phrase used to work as a verbal shrug for any question that didn’t require personal judgment. Now it sounds dated, closer to telling someone to “look it up in the yellow pages” than to offering a practical next step.

My default changed without a conscious decision. These days, I ask an AI, even for quick, keyword-style questions. Most of the time, the answer is enough. No clicking through. No scanning ten results to triangulate what’s correct. Just an answer that satisfies the immediate intent.

This shift happened faster than I expected, which is the part worth paying attention to. Habit change at that speed usually signals a structural improvement in the workflow, not just novelty.

About eight months ago, I wrote about the move from Google to Perplexity. At the time, it felt like a significant change in how I consumed information. The model was clear: the search engine as librarian, pointing you to shelves; the answer engine as expert, giving you the summary directly. That shift is now permanent for me. I don’t see myself going back. Looking at it now, I realize that was only step one.

The Problem with Independent Queries

Traditional search treats each query as a fresh start. You search, scan results, click around, then notice a gap and refine the question. The next query ignores the path that led there; the system doesn’t retain what you read, what confused you, or the specific understanding you’re trying to reach.

Google does “personalize” results, but mostly using history and location. That signal doesn’t reflect what you’re working through right now. There’s no thread connecting queries, no accumulation of context, so each search behaves like an island.

This is broad research. You gather fragments from different sources, hold the evolving context in your head, and do the synthesis yourself. It works, but it’s work. The burden of connecting the dots sits entirely with you.

Suggestions help a little. “People also ask” and related searches can point somewhere useful, yet they remain generic prompts rather than responses to your specific confusion. They don’t know what you didn’t understand. They can’t ask clarifying questions back.

What Deep Research Actually Looks Like

The shift runs deeper than links to answers. It’s monologue turning into dialogue.

AI research no longer expects a perfect first response.

That initial output serves as a starting point. Something to react to. Something to challenge.

“What do you mean by that?” “Can you explain it more simply?” “How does this connect to X?” “Assume I already know Y — what’s the next layer?”

Real understanding builds here. Not in the first answer, but the third. The fifth. The moment you say “wait, I thought it worked like this,” and the AI clarifies: “not quite; here’s the distinction.”

A static article can’t match this. It targets an imagined reader at an assumed level. If you’re above it, tedious. If below, confusing. You can’t interrupt a blog post. You can’t say “hold on.”

Deep research adapts information to you. Your gaps. Your pace. Your specific confusions. Not a generic audience.

The Space-Time Example

I’ve known for years that time is relative. That astronauts age slightly slower than people on Earth. That movies like 2001 and Interstellar play with this concept. I understood it as a fact. I didn’t understand why.

At some point, I wanted to know more. Just to actually grasp what I’d been nodding along to for decades.

“Why does time pass more slowly near a massive object?”

The answer involved gravity and the curvature of space-time. I didn’t understand “curvature of space-time.” I said so.

“What does it mean for space-time to be curved?”

New explanation. The fabric analogy, mass creating dips. I’d seen that visualization before, the bowling ball on a rubber sheet. But it never clicked. A flat sheet is a 2D representation. We live in 3D. What’s actually being curved?

“The rubber sheet analogy is 2D. How does curvature work in our actual 3D space?”

This is where it got interesting. The AI introduced Minkowski diagrams, a way to visualize space and time as a single coordinate system. Then gravity. I’d always thought of gravity as a force. Earth pulls the Moon, the Moon pulls the oceans, tides happen. But apparently that’s the Newtonian picture. Einstein’s version is different. Objects in free fall aren’t being “pulled.” They’re following the straightest possible path through curved space-time. A geodesic.

“So gravity isn’t a force?”

Not in the way I’d imagined. More like: mass curves space-time, and objects move along that curvature. The Moon isn’t being pulled by Earth. It’s following the geometry that Earth’s mass creates.

Get André Bergholz’s stories in your inbox

Join Medium for free to get updates from this writer.

This led to stranger questions. If time and space are unified, what does that mean for time travel? Is a “time machine” actually a space-time machine? Are “past” and “future” even meaningful as standalone concepts, or do they only make sense relative to a position in space-time?

I asked all of these. Some answers were clearer than others. Some opened more questions.

I still don’t understand most of it. Tensor math is beyond me. I couldn’t derive anything. I’m at the surface, and I know it. But it’s a surface I couldn’t reach before. Twenty minutes of conversation gave me more genuine understanding than years of vaguely meaning to “look into relativity someday.”

And I can return anytime. Pick up where I left off. Ask the next question.

The Habits That Changed

When I look at my own behavior over the past year, a few patterns stand out.

Questions, not keywords

I used to write search queries like commands. Short. Optimized. Drop articles, skip verbs, keep only essential terms. That’s how you talked to search engines. Now I write full questions, sometimes full paragraphs of context. The input field isn’t a command line anymore; it’s the start of a conversation, and context improves the outcome.

Expecting dialogue

My mental model moved from “retrieve information” to “discuss a topic.” The first answer is just a starting point. If something is unclear, I ask for clarification. If I need more depth, I request it. If the framing doesn’t work for me, I ask for a different one. The exchange continues until it clicks.

Calibrating depth

This might be the biggest change. I can request explanations at the level I actually need: “Assume I understand basic calculus but not differential geometry.” “Skip the history, just explain how it works.” “Explain like I’m new to this.” No article can do that. Every article picks one level and commits to it. Conversations adjust.

Fewer open tabs

I used to accumulate browser tabs like debt. Articles I’d “definitely read later.” Videos I’d “get back to.” It rarely happened. Now more loops close inside the conversation itself. Questions get answered. Curiosity gets satisfied. The tab count stays low.

Voice as input

I speak faster than I type. Much faster. And I’ve stopped worrying about precision. Voice mode changed how I mostly interact with AI. I ramble, think out loud, let half-formed questions spill out. The transcription doesn’t have to be perfect. It’s good enough. The AI figures out what I meant. This lowers the barrier further.

Saved calibrations

I now store prompts for recurring modes. For example, /eli5 for radical simplicity, /research for depth plus sources, /history for context and origins. They pre-set the level before the conversation begins. What once required a human tutor who knew my background now reduces to a saved prompt.

Sources as Trust Anchors

Perplexity gets one thing right: answers include receipts. You see the sources. Click through to verify, go deeper, or trace origins if needed.

I don’t always click. Most times, I don’t need to. Knowing I can builds trust. Traceability anchors reliability. Source-less answers feel like hearsay. Sourced ones feel like a briefing.

In fact, this matters more as AI answers improve. Better outputs tempt blind acceptance. Visible sources counter that. They remind us: information has origins worth checking.

The UI Catches Up

The interface now reflects these shifts. The recent Neo Browser calls their input field the “magic bar.” I don’t fully understand it yet. When does it search the web? Open a URL directly? Invoke AI? Boundaries blur. Probably by design.

Everything heads this way. Type what you want. The system figures out how to deliver it. Search, AI, direct navigation, they all collapse into one entry point. Less thinking about how to find information. More focus on what you want to know.

It’s still opaque. Sometimes I’m unsure what’s happening under the hood. The trajectory looks clear: automatic intent detection, seamless mode switching. The “search bar” evolves into a general-purpose interface for information interaction (as regular readers of mine may know I am very interested in the omnibar idea).

Broad Still Has Its Place

I’m not saying traditional search is dead. For direct navigation — opening a specific homepage, finding a restaurant’s menu, checking opening hours — it still works. In those cases, you already know what you’re looking for and where it lives. You just need the link.

And broad research still has value. You scan multiple sources to map an unfamiliar field. You compare how different outlets frame the same event. You leave room for serendipity, stumbling onto something you didn’t know you were looking for. Deep research is focused. Sometimes you need breadth first.

But for understanding? For actually learning something you didn’t know before? Iteration wins.

And now, for the first time, that’s available to anyone with a question and a few minutes to spare.

Thanks for reading! If you enjoy these texts, please consider subscribing. You can also explore my complete collection of AI writings here.

--

--

André Bergholz
André Bergholz

Written by André Bergholz

Writing about AI for leaders who need to understand it, not just deploy it. CPO & Co-Founder @Implisense.

AI Advances
AI Advances

Published in AI Advances

Ideas in depth. Where exceptional human thinking takes shape.