Start With a Map, Not a Search Box
The old way to start a literature search was to type three keywords into a database and pray. The problem: keyword search only finds papers that use your words. Half the field calls the same idea something else, and you never see it.
Semantic search fixes this. Tools like Semantic Scholar, Elicit, Consensus, and Connected Papers match on meaning, not exact strings. You describe what you're after in plain English and they surface work that's conceptually close, even when the vocabulary doesn't overlap.
Before you search anything, get the lay of the land. Ask a general-purpose model to sketch the field so you know what you're walking into:
I'm researching [your topic] for a [course paper / thesis].
Give me:
1. The 4-6 major subtopics or debates in this area
2. The standard terms and any competing terminology researchers use
3. 3-5 names of foundational or highly-cited authors/papers, if you know them
Flag anything you're unsure about — do not invent citations.
Treat the output as a map, not gospel. The model might hallucinate a paper title or an author. That's fine here — you're using it to learn the language of the field and where the fault lines are, not to cite. The terminology it gives you becomes fuel for the actual search.
Turn a Vague Topic Into Search-Ready Questions
"AI in education" is not a research topic. It's a Wikipedia category. You can't find focused literature until you've focused the question.
Use the terms from your map to generate sharper queries. A model is good at this because it can spin one fuzzy idea into a dozen angles in seconds:
My broad topic: [topic].
Generate 8 specific, researchable questions I could investigate,
ranging from narrow to broad. For each, list 3-4 search keyword
combinations (including synonyms) I'd use in an academic database.
Now take those questions into Elicit or Consensus. These tools read across thousands of papers and return findings with the source attached, often in a table: claim, study, sample size, methodology. Consensus even shows you how many papers support or contradict a question. That's a fast way to see whether your question is settled, contested, or wide open — which tells you whether it's worth writing about.
If you're new to driving these tools well, the AI for Academic Research & Papers course walks through the search-and-screen workflow step by step.
Follow the Citation Trail Without Drowning
Once you've found two or three solid papers, stop searching and start tracing. Citation chasing finds better sources faster than any keyword query, because the authors already did the filtering for you.
There are two directions, and you want both:
Backward: who they built on
Look at what your good paper cites. The same handful of references showing up across several papers in your pile? That's almost certainly the foundational work everyone in the field stands on. Read those next.
Forward: who built on them
Look at who cites your paper. This is how you escape the trap of reading only old work. Google Scholar's "Cited by" link and Semantic Scholar's citation graph both do this. A 2015 paper with 900 citations tells you it mattered; the recent papers citing it tell you where the field went next.
Connected Papers visualizes this whole web as a graph — your seed paper in the middle, related work clustered around it by similarity. Drop in one good paper and you get a visual map of its neighborhood in seconds. It's the fastest way to spot the cluster you've been missing.
One rule that saves hours: find the review article first. A recent literature review or meta-analysis is a pre-built reading list written by an expert. Search your topic plus "systematic review" or "survey," read that one paper, and mine its reference list. You'll have a curated bibliography before lunch.
Build the Reading List in an Afternoon
Here's the workflow that takes you from blank to focused. Budget about three hours.
- Map (20 min). Run the field-overview prompt. Learn the terminology and the big debates.
- Focus (20 min). Generate sharp questions, pick one, lock your keywords.
- Seed (40 min). Use Elicit or Consensus to find 3-5 strong, relevant papers. Prioritize a recent review.
- Trace (60 min). Run those seeds through Connected Papers and Scholar's "Cited by." Pull every paper that keeps reappearing.
- Triage (40 min). You now have 30-ish candidates. Cut to the 8-12 you'll actually read.
For that last step, don't read abstracts one by one. Speed it up:
Here are 15 paper titles and abstracts: [paste].
My research question is: [question].
Rank them by relevance to my question. For each, give a one-line
reason and a tag: [core / supporting / tangential / skip].
You read the ranked list, override the model where it's wrong (it will be, sometimes), and you've got your shortlist.
Verify before you trust
This is non-negotiable. AI-powered search tools occasionally surface papers that are subtly mismatched, and general chatbots will flat-out invent plausible-looking citations — fake authors, fake DOIs, real-sounding journals. Every paper that makes your list gets confirmed to exist in a real database: Google Scholar, your library catalog, or the publisher's site. If you can't find it, it doesn't go in your paper. Building this habit now keeps you out of the integrity mess covered later in the book.
A focused reading list isn't the one with the most papers. It's the smallest one that covers the debate. Ten papers you've actually traced and verified beat fifty you skimmed and half-trust — and you can have those ten by dinner.