Picked from the software, topic, and level of “Agentic AI Without Coding Class #16: Web Search and Deep Research: When AI Searches the Internet Itself”.
At most one email a week. Unfollow anytime with one click.
Transcript
The full narration of this video. The current sentence is highlighted while the video plays; click any sentence to jump there.
Show transcript
Introduction
Welcome to Session 4 of Agentic AI Without Coding. In this session, AI no longer just answers from memory; it starts searching, reading, and acting on the internet. In episode sixteen, we take apart two research features: web search and deep research. You'll learn to read their source citations, check them, and choose the right feature for each question.
Step 1: From a shop assistant to AI that looks outward
In Session 3, we built Sunset Coffee's customer service assistant. That assistant knows the shop's catalog, prices, and SOP by heart, because we gave it knowledge files. But try asking about something outside those files, such as this month's coffee drink trends or the next small business bazaar nearby. The assistant can only guess from the model's memory, which has a cutoff date. This is where AI needs to learn to look outward. When does AI need to search the internet? First, when you need the latest information: news, trends, or rules that just changed. Second, market prices and products. For example, what's the price range for one-liter iced coffee in Bandung this week, and which promos are popular. Third, local data: events, suppliers, or other businesses in your city. Information like this changes quickly and is rarely stored in the model's memory. Fourth, facts that need evidence. If you'll quote a number in a proposal or in content, you need a source you can show. On the other hand, if the answer is already in your shop files, the Session 3 assistant is enough. Here's the Session 4 map. Episode sixteen, which you're watching now, covers web search and deep research: AI that searches and reads. Episode seventeen goes up a level: browser agents, AI that opens web pages, clicks, and fills things in on its own. Episode eighteen covers the brakes and the steering wheel: when AI may act on its own, and when it must ask for your approval first. Episode nineteen is the first practice: a competitor and market price research agent for Sunset Coffee. And episode twenty, practice with form-filling and data-collecting agents, ends with an automatic market research report. That report is the Session 4 output you'll create for your own business.
Step 2: Web search: quick answers with sources
Let's see web search in action. This feature is in ChatGPT, Claude, and Gemini, sometimes on automatically, sometimes turned on from a tools menu. Laras asks: what are this year's trends for one-liter bottled coffee among small businesses, and she asks for sources. Notice this gray line. The AI doesn't answer straight from memory; it turns Laras's question into search keywords, including the year. Then it opens and reads several search result pages. This usually takes just a few seconds. The result: three concise trends. The most important part is at the end of each sentence: numbers in square brackets. Those are citations, markers showing which page each claim came from. At the bottom is the list of sources. In the real app, these citations are clickable and open the original page. The site names here are fictional, and the menu layout may differ slightly because AI apps are updated often. Behind the scenes, web search works in four stages. First, you ask in everyday language, like a conversation. Second, the AI builds one or more search keywords, sometimes adding a year or a city name to make the results relevant. Third, the AI reads a few top pages from those search results. Just a few pages, not the whole internet. Fourth, the AI writes an answer and marks its sources. Remember one important thing: the answer is only as good as the pages it happens to find. AI summarizes; it doesn't guarantee the truth. To make web search results more trustworthy, use this prompt pattern. The task section names the topic, region, and period. For example, one-liter coffee, Bandung, published in the last six months. The format section asks for a citation number on every claim, plus a full list of sources with their publication dates. Dates matter, because an article from three years ago might show up too. And the limits section asks the AI to separate facts from guesses, and to be honest if it finds no source. The sentence never invent numbers sounds simple, but it helps a lot. For Laras, the full prompt goes something like: search for one-liter iced coffee trends in Bandung from sources in the last six months, answer in five points with citations and dates, and say if nothing was found. Save this pattern in your prompt library from Session 2.
Step 3: Checking sources before trusting them
A citation isn't a guarantee. So we check it ourselves. In the right panel, I note down the manual checking steps. First, click citation number one, and the original page opens. Second, find the sentence the AI quoted. The answer said orders for office events are increasingly common. And yes, it's there in paragraph three. Passed. Third, look at the publication date: March 2026. Still recent, so it's still relevant to describe current conditions. If the date were two or three years ago, a claim about a trend shouldn't be used for today. Fourth, check what the data is based on. Paragraph two mentions a survey of one hundred twenty buyers. That's a small survey, not national data. The number can be used, but don't exaggerate it. Fifth, who's the publisher? A business news outlet, not a promotional page owned by a one-liter coffee seller. Sources selling their own product tend to exaggerate. The conclusion: this claim can be used, with a note that it comes from a small survey. The whole check takes about two minutes, and this habit is what separates trustworthy research from simply copying an AI answer. From that exercise, we can sum up the signs of strong and weak sources. Publisher: official bodies, media, or business associations are stronger than anonymous sites or a seller's promotional blog. Date: strong sources show a clear, recent date. Be wary of pages with no date, or ones that are years old. Content: strong sources explain where the data came from and how it was collected. Round numbers with no explanation, like up fifty percent, deserve suspicion. And match: the sentence in the AI's answer must actually be on that page. Sometimes AI attaches a citation to a page that turns out not to contain the claim. That's the main reason we always open the source. Keep these five steps as a habit. One, click the citation and make sure the page actually opens. A dead link is the first red flag. Two, find the quoted sentence or number on that page. Use the browser's find feature to be quick, then read the paragraph around it so you don't lose the context. Three, check the publication date and who the publisher is. Articles with no date, or promotional pages owned by sellers, need to be read more carefully. Four, find at least one other source saying something similar. One source can be wrong; two independent sources are far more convincing. Five, don't use claims that fail in your content or decisions. Or, if you still use them, state their limitations honestly.
Step 4: Deep research: step-by-step reports from dozens of sources
Now let's go up a level: deep research. If web search is like asking a friend a quick question, deep research is like assigning the work to a research assistant. It builds a plan, reads dozens of sources over a few minutes, compares them, then writes a full report with citations. ChatGPT, Claude, and Gemini each have a feature like this, with similar names. Let's look at an example. Laras turns on deep research mode, then writes: do deep research on the opportunity for one-liter iced coffee for office orders in Bandung. She also gives her business name and the price of her one-liter bottle. Notice, the AI doesn't set off right away. It usually asks clarifying questions first: focus on private companies or government offices, whether a price comparison is needed, and what decision the report is for. These clarifying questions are valuable. Answer them specifically. Laras replies: both, yes to price ranges, and the report is for designing an office subscription package. Then the AI builds a research plan. Laras's big question is broken into four smaller sub-questions, such as what office ordering patterns look like and what the market price range is. After that, the AI explores. It searches, opens, and reads pages one by one. In this example, twenty-four sources are read. It also compares the findings, including looking for information that contradicts. This is what makes deep research deeper than regular web search. A few minutes later, the report is ready. There's an executive summary, office ordering patterns, market price ranges, package recommendations, and data limitations. This report includes twenty-four source citations. Much more complete, but also much more for you to check. We'll cover how in the next step. So deep research has five stages. First, clarify: the AI asks questions back so it doesn't head the wrong way. Second, plan: the big question is split into several sub-questions. Third, explore: dozens of sources are read over a few minutes. Sometimes the AI even adjusts its plan midway, if it finds a more promising direction. Fourth, compare: the findings are matched, and contradictions are flagged. Fifth, a report with chapters, tables, and citations. While it runs, you can do other work. If you look closely, this is the agent pattern from Session 1: goal, plan, action, then check the results. As a rough picture, web search answers in seconds. Deep research takes several minutes, sometimes up to about half an hour, depending on how broad the question is. And it can read dozens of sources. One more thing: deep research is usually limited by a quota based on your subscription plan, and the rules change often. So check the latest pricing page before relying on it for routine work.
Step 5: Reading a deep research report critically
A deep research report can be very long. So you don't drown in it, get to know its anatomy first. Part one, the executive summary: the core answer in one paragraph. Read this first. Second, findings per sub-question, each with source citations. This part answers the smaller questions set out in the research plan. Third, comparison tables. This is where the important numbers live, and it's exactly the part that most needs double-checking. Fourth, recommendations. Good recommendations are tied to your decision, such as Sunset Coffee's office subscription package. Fifth, limitations: what wasn't found or isn't certain. A report without this section may be overconfident. If it's missing, ask the AI to add it. Let's practice reading critically. Here's a slice of Laras's report: a table of one-liter price ranges, split into home-based shops and trendy coffee shops, then a recommendation for an office package of ten liters a week. Numbers like these will be used to set prices, so Laras asks back: which sources are these ranges from? Separate the official shop menus from buyer reviews. The answer reveals something important. Four numbers come from shop menu pages, two from buyer reviews. And two of the menu pages turn out to have no date, so their prices may have changed. You'd never see this if you only read the summary. One question back is enough to learn how solid a number's basis is. Watch out for these three traps. First, sentences that sound very confident but have no citation number. They may be the AI's own conclusion, not what the source says. Second, old sources used to describe current conditions, like prices from three years ago. Third, data from one city or one small survey treated as if it applies everywhere. Coffee prices in Jakarta, for example, aren't necessarily the same as in Bandung. The fix is simple: ask where the most important numbers came from, then manually check the three claims that most affect your decision using the five steps from earlier. There's no need to check all twenty-four citations.
Step 6: When to use web search, when deep research
Now let's put them side by side. Time: web search takes a few seconds, deep research a few minutes up to about half an hour. Depth: web search reads a few top pages, deep research reads dozens of sources and compares them. Result: web search gives a short answer with citations, deep research gives a report with chapters and citations. Fit: web search is for quick facts or checking one thing. Deep research is for important decisions like market research. And usage limits: web search is generally generous, while deep research has a limited quota. So the practical rule: start with web search, and move up to deep research when the answer really needs comparing from many angles. Let's practice. Laras wants to know the opening hours of this weekend's coffee expo. The answer is web search. One fact, needed fast. Still open the organizer's official site to confirm. Dewi, operations manager at Maju Bersama Co., wants to compare five paper vendors for a one-year contract. That's deep research: a big decision that needs many sources. Mr. Hendra from Cerdas Tutoring wants the latest school exam schedule announcement. Web search, then go straight to the official notice from the school or the education office. Nadia, a social media freelancer, wants to know food content trends over the last three months for her clients. Deep research, because she needs patterns from many sources, not a single fact. Before moving on, do this small assignment. First, ask one question about your business with web search, using the prompt pattern from earlier, and ask for its citations. Second, pick two citations from the answer, then run the five source-checking steps. Third, if your plan includes deep research, try one small study of your own market. If not, that's fine; web search is enough for this exercise. Fourth, note how many claims passed the check and how many didn't. That number will make you wiser about using AI research results in the episodes ahead.
Done!
Let's sum up. Web search gives quick answers with citations from a few pages. Deep research asks questions back, builds a plan, reads dozens of sources, then writes a report. Either way, you still need to check the sources. In the next episode, AI goes a step further: browser agents like Claude in Chrome and ChatGPT Agent, which can open pages, click, and fill things in on their own. See you in episode seventeen.
Practice exercises
Try these on your own after watching to lock in what you learned.
Ask one business question with web search using the task, format, and limits prompt pattern. Ask for a citation number on every claim.
Pick two citations from the answer and run the five source-checking steps. Note which ones pass and which don't.
If your plan includes deep research, run one small study of your market, then ask where its three most important numbers came from.
Ask the class material
Type a question or keyword; we find the part of this episode that explains it and jump the video there.
My notes
Notes are saved with the current video time and are only visible to you.
Questions & answers
Stuck on a step? Ask here; the instructor and other learners can answer.
No questions yet.
Your instructor
Cerdas Inovasi Digital Team
Software developers, BIM & 3D engineers, and AI practitioners
53 video classes1,281 episodes
We build software, BIM models, games and AI automations for clients every day, and turn that hands-on work into step-by-step classes: every step is shown on screen and comes with a PDF guide you can keep.
