Research for work
~ min read
30-second summary
- Three ways to use AI: map a new territory (figure out what exists before you start searching), synthesize multiple sources you already have, do live research with engines that cite (Perplexity, ChatGPT with search, Gemini).
- Unlike Module 2: here you produce for someone else who trusts your name. A hallucination isn’t a mistake you’ll catch yourself, it’s a mistake you’ll sign.
- The specific risk has three shapes: invented citations, round numbers without a source, syntheses that attribute to a source things the source doesn’t actually say.
- Three-step pattern: ask for the source, open the source, track where each thing came from in your final document.
- For output you sign (proposals, reports, briefs), AI’s first synthesis is the starting point, never the ending point.
You have a meeting tomorrow morning with your sales director and they want five points on how the three main competitors are moving on topic X. Or: you’re preparing a proposal for a client in a sector where you don’t have deep experience, and you need market numbers, regulations, recent trends. Or: the board has asked you for a status update on regulation Y that comes into force in six months, and you didn’t know it existed.
In all three cases the pattern is the same: short on time, lots of potential sources, an output you’ll sign with your name.
AI speeds this work up considerably. It maps a territory in two minutes where by hand it would take you twenty, summarizes five articles in thirty seconds, and searches in real time with citations. It’s the work use case with the most unbalanced speed-to-risk ratio, in both directions.
The hallucinations you saw in When to trust it and when not to aren’t an abstract problem here: they become a wrong number in a client pitch, an invented citation in a board report, a trend that never existed presented as industry data. Difference from Learn something new in Module 2: there you study for yourself, and if AI is wrong reality corrects it later; here you produce for someone else, and if AI is wrong the problem reaches a client, a manager, a board.
This lesson is the playbook to use it well: three standard moves, the specific risk, and a verification pattern you can’t skip.
Three ways to use AI for research
Section titled “Three ways to use AI for research”Three distinct moves, picked based on where you are in the process: you don’t yet know what to search for, you already have something to synthesize, you need a fresh fact with an open source.
1. Map a new territory
Section titled “1. Map a new territory”When you’re asked to work on a topic where you’re not an expert, the first problem isn’t searching: it’s figuring out what exists. What are the sub-topics? Who are the main actors? What’s the specific vocabulary? What’s connected to what? Without a map, every search on a search engine starts blind.
This is where AI is an index of the problem. You ask it for a map, not the facts.
“I need to prepare a brief on ‘European AI Act regulations for SMEs’. I’m not an expert. Give me a ten-point map of the territory: the main sub-topics, the institutional actors involved, the technical vocabulary I should know, and the three or four questions an SME client would ask me.”
You get something like: four sub-topics (risk classification, transparency obligations, use in HR, regulatory sandboxes); actors (European Commission, national supervisory authorities, data protection authorities); key terms (high-risk AI, foundation model, General-Purpose AI); typical questions (do I need to notify? what fines? when does it come into force?).
That map isn’t the final answer. It’s the compass: it tells you where to look, not what to write. From there you know which terms to search on an engine, which institutional sources to open, which questions to ask. Without the map, you waste twenty minutes just figuring out what exists.
The map itself does need checking, but it’s a structure that’s easy to verify: you search for the actor’s name, you search for the technical term, and in two seconds you see whether they exist. If the actor exists and the term is in use, the map node makes sense. If you find nothing, AI probably pulled it out of thin air.
There’s a point where readers of Summarize a long document might get confused. There Technique 3 asked for the index of a specific document you had uploaded; here the map is of a topic, with no source document to start from. The first is the outline of a text, the second is a compass on a territory. The risk also changes: the index of a document is verified by scrolling the document, the map of a topic is verified by searching the names on a search engine.
2. Synthesize multiple sources you already have
Section titled “2. Synthesize multiple sources you already have”You’ve opened three articles, downloaded two reports, saved a PDF. Five documents, you want one synthesis with the points of convergence, the divergences, and the key facts that repeat.
This is where the techniques from Summarize a long document scale up: instead of one document you have five, and the output isn’t a summary but a comparative analysis.
“I’m uploading five sources on topic X (three articles, two reports). For each one, give me a five-point summary. Then give me a comparison table: where they converge, where they diverge, and which key facts (numbers, dates, names) are repeated in at least two sources. For each key fact, write in parentheses which source supports it.”
The three constraints that make the difference: per-source summary (not a single blob), comparison table (makes convergences and divergences explicit), source attribution per fact. The last one is what saves you: a fact appearing in three sources carries more weight than one appearing in only one, and as the person signing the deliverable you need to know that.
Asking AI to show attribution also changes its behavior: if it can’t say which source a fact comes from, that’s the signal that it probably added the fact from its training, not from the sources you gave it. If it writes “not directly from any source, but it’s a generally known fact”, you know that fact needs to be verified separately.
On volume: five news articles or blog posts fit into a chat without trouble. Five PDFs of a hundred pages each is a different story, because every platform has a context limit beyond which documents get truncated silently. For long documents, the same advice from Summarize a long document applies: either you pass the documents one by one, get a summary for each, and then ask for the comparative synthesis on the summaries; or you use a dedicated tool like Google NotebookLM, which handles bundles of bulky sources and keeps per-source traceability in its responses.
3. Live research with citations
Section titled “3. Live research with citations”When you need a fresh fact (last quarter’s statistics, regulation that came into force last week, yesterday’s news article) standard AI is the wrong tool: ChatGPT base, Claude base, Gemini base have a training cutoff date and don’t know what happened after. On specific facts they didn’t see in training, they can invent easily.
Here you need engines with web access and citations:
- Perplexity: built specifically for research with inline citations. The response comes with footnote numbers and clickable links.
- ChatGPT with search (selecting a model with web search enabled): searches, cites, and links.
- Gemini: integrated with Google Search, responses come with sources.
- Claude with tools (on business plans or via specific plugins): can search and cite in enterprise scenarios.
How to choose between these four? For pure research (a fact, a statistic, a name with a source) Perplexity is the most direct: that’s what it was built for and it cites by default in a very readable way. For research inside a wider workflow (you’re already writing, rewriting, iterating) ChatGPT with search and Gemini are more convenient: you stay in the same tool you’re doing the rest of the work in. If you already pay for one of the three, that one is fine to start with; if you don’t pay for any, Perplexity has a generous free tier for the research case. Newer modes like “Deep Research” on ChatGPT or agentic search on Perplexity Pro (modes where AI does multiple search rounds on its own, digs deeper, and hands you a longer, more structured output) take more time and cite more sources. They don’t change the underlying problem: the verification pattern you’ll read in the next section is the same, because invention doesn’t disappear with more automated steps, it just hides better.
All four share the same non-negotiable constraint: the link must be opened. The citation is part of the prompt, not the verification. Sometimes the engine cites an article accurately but paraphrases a sentence the article doesn’t actually contain. And yes, live engines also invent: less often than the base versions, but the problem hasn’t disappeared. You’ll see it in the ChatDemo of the A concrete example section.
The professional hallucination: the specific risk
Section titled “The professional hallucination: the specific risk”On research-for-work topics, AI can invent in three specific ways. Knowing them by name helps: they’re recognizable at a glance once you know what to look for.
Invented citations. The book doesn’t exist, the author doesn’t exist, the year is plausible but wrong. The title often “sounds” authoritative (Smart Markets in the Age of AI, Harvard Business Press, 2023). Classic case: you ask AI to cite three books on topic X and it gives you three. When you search, you find that one is real, one is by a real author but the title doesn’t exist, and one is completely made up.
Round numbers without a source. “73% of companies report that…”, “The market is worth 12 billion dollars globally”. Precise numbers, suspiciously round percentages, with no indication of who measured them or when. On LinkedIn these numbers propagate from one post to another because nobody verifies them. In a work deliverable they’re poison.
Synthesis that invents attribution. You gave AI five sources. A claim shows up in the synthesis. You go look for it in the five sources and it isn’t there. AI has synthesized by combining things it found in the sources with things it already “knew” from training, and the two layers blend together in the single output.
These three patterns aren’t a built-in flaw of AI: they’re the behavior of a system that produces plausible text from a request, even when it doesn’t have the exact information. Same mechanism as in When to trust it and when not to. What changes here is the consequence: in a professional deliverable, an invented citation is an embarrassing moment with a client, an invented number is a flawed analysis that will steer a decision, an inflated synthesis is your name next to a false claim.
Pattern: verify, cite, track
Section titled “Pattern: verify, cite, track”For every piece of information that will go into the deliverable, three steps you can’t skip.
1. Always ask for the source. If AI gives you a number, a citation, a name, a fact, ask “where does it come from? Give me the specific source with link, author, and date”. If it can’t point to one, throw the fact out. A good rule is to ask for it already in the initial prompt: “for every factual claim, indicate the source in parentheses”. Prevention beats chasing. The constraint isn’t a guarantee: the model can still invent a plausible source. But it works as a brake: most of the time it prefers to declare “no recent source” rather than fabricate one if you’ve given it an explicit safety exit. The verification at step 2 stays indispensable.
2. Open the source. If the source exists, go to the link and verify that the sentence AI cites is actually there. Often it’s there but rephrased, and the rephrasing changes the meaning (it goes from “grows by 10%” to “could grow up to 10%”, or from “in some countries” to a general claim). When the response cites many sources (easily happens with Perplexity in deep mode), you don’t need to open them all: open the ones that support the facts going into the deliverable. If the brief has five key points, that’s five sources to open, even if the response cites fifteen. For a report you sign, cast a wider net: every fact cited as a specific data point (numbers, dates, names) needs checking.
3. Track where each thing came from. In your final document, keep a record of the source for every important data point: a “source” column in the working sheet, a footnote in the report, a margin note that won’t end up in the deliverable but that you have. The minimum acceptable is one row per fact with four fields: the fact, the source, the link, the date you checked. Like:
“Diadora e-commerce relaunch 2024 / source: Diadora press release / link / verified 24/04/2026”.
It serves two purposes: it lets you go back if someone asks “where did you get that?” (and someone always asks, sooner or later), and it forces you not to leave any fact unverified, because you notice when you have to fill in the column.
The pattern isn’t academic. It’s how professionals who’ve learned that shipping a deliverable without traceability is the fastest way to end up justifying a mistake in front of a client six months later, with no memory of where the fact came from.
A concrete example
Section titled “A concrete example”You’re preparing a half-page brief for the sales director. The topic: how the three main Italian sportswear companies have moved on the e-commerce channel over the past year. You’re not an expert in the sector. You have thirty minutes.
Strategy: start with a live search engine, ask for five points, and in the prompt require sources for every fact. If a point has no source, AI must say so.
AI did well on two axes: it attributed the facts in the first three points to specific sources (with links, dates, and page numbers where relevant), and on points 4 and 5 it declared the limit rather than filling in with a plausible number. Point 5 is exactly where an AI without the “if you don’t have a source, say so” constraint would probably have invented a suspiciously round figure.
From here your work isn’t done: the three links in the first points need to be opened one by one, the sentences need to be verified word for word against the original text, and the two missing data points get gathered from elsewhere (an internal team, a specialist source, a sector agency) or you accept that the brief leaves them blank with a note.
What NOT to do
Section titled “What NOT to do”Don’t ship a deliverable without verification. Even if AI gave you five cited sources and everything seems coherent. Opening links and checking sentences isn’t optional: it’s the difference between a professional brief and an embarrassing slip. A wrong citation in a board report, surfaced six months later, is the kind of mistake that sticks.
Don’t ask “do the research for me” as a one-shot. AI doesn’t know your client, doesn’t know what angle to take, doesn’t know your internal constraints. Without orientation it gives you a generic synthesis. You’re the one who has to steer it with questions, constraints, and specific context. Same principle as Things NOT to do: the decision on what goes into the deliverable and what doesn’t stays with you.
Don’t trust round numbers without a source. “73%”, “a third of companies”, “over 10 billion”. Every number AI hands you without a traceable source is a number to throw out. Same pattern as Working with data and tables: AI can be wrong about a number with the same confidence it has when it’s right, and on market data going into a brief, confidence isn’t enough.
What’s next
Section titled “What’s next”The next lesson, AI as a thinking partner, shifts the axis: from research (finding and synthesizing information that exists) to thinking (using AI to anticipate objections, play devil’s advocate on your ideas, brainstorm in a structured way). The risk changes shape: there hallucinations matter less, but a new risk shows up, the risk of being reassured rather than challenged.