8 Research Builds on Research
This chapter is part of a book in active development and has not yet been through the author’s review. Content may change as the review advances.
The research decision. Which conversation your question joins, and which of the candidate sources a search or an AI hands you are worth chasing. You decide what goes on the list and what gets dropped. Nothing on that list counts as evidence yet.
8.1 Why this decision matters
The decision on the table: which conversation your question joins, and which leads are worth your time.
Picture a thesis advisor reading your first draft with one line circled: “You cite six papers. Where did they come from, and what argument are they having?” Research never starts from zero. Someone has already asked something close to your question, and your job is to find them before you spend months rediscovering what they already know. Get the search wrong in one direction and you miss the study that answers you. Get it wrong in the other and you build on a citation nobody ever wrote. Get it right and you inherit years of other people’s work for the price of an afternoon.
8.2 The concept
A field is a cumulative conversation: a running argument where each study answers, extends, or challenges the ones before it (Booth et al. 2024). One team measures how a policy moved turnout in one state, a second repeats it elsewhere, a third pools both. Your job is not to invent an island. It is to find the conversation your question belongs to and add one honest sentence.
Everything you gather starts as a lead. A lead is a candidate source you have not checked yet, such as a citation an AI hands you. It stays a lead until you retrieve and read it. That standard is the retrieval-verification loop, and it has four steps: ask a tool for leads, retrieve the source yourself, verify it exists and says what was claimed, and document where you found it. This chapter is about the first step, gathering well. The next chapter is where the leads earn promotion.
From one real source you grow the rest by snowballing: walking its links two ways (Greenhalgh and Peacock 2005). Backward citation search reads the source’s own reference list for the older work it built on. Forward citation search finds the newer papers that cite it. Two solid sources, snowballed both ways, quickly become ten, and the ten arrive with their relationships already visible.
Snowballing is also how you tell a real gap from a dead end. A genuine gap is a question your verified sources actually leave unanswered, confirmed by looking both ways (Snyder 2019). An unfinished search is one you simply stopped chasing. They feel identical from your desk, and only one is a contribution.
You then arrange what you found as an evidence map: each source becomes one claim node (what it asserts, in one sentence), lines connect claims that agree or contradict, and one marked spot shows where no source reaches. That spot is your candidate gap. It becomes honest only when every node around it is a source you opened yourself.
8.3 A worked example
Say your curiosity is voting. You have read that how far people live from their polling place seems to matter for whether they show up. Your question: does moving a polling place farther away lower turnout, and does the drop land harder in some neighborhoods than others?
You ask an AI for the key studies. It hands you a confident citation: two plausible author names, a 2019 date, and a journal whose title reads like a longer, more official version of the real Electoral Studies. You go to read it. The exact title returns nothing in your library portal, nothing on Google Scholar, nothing on the publisher’s site. That journal does not exist. The citation was manufactured, and it took four minutes to find out.
So you run the loop properly. You retrieve one real review of polling-place location and turnout, read it, and snowball. Its reference list points backward to the older work on the costs of voting that it builds on. Its “Cited by” trail points forward to recent studies using precinct-level administrative records. Within an hour your list holds a half-dozen real sources and one live disagreement: some studies find that relocating a polling place shaves a point or two off turnout, others find that voters switch to mail ballots and total turnout barely moves. That disagreement is the live edge of the field, and it tells you the question is unsettled.
Now the decision that is yours. The broad question “does distance affect turnout” is no gap; the literature is enormous. But your narrower question, whether the drop concentrates in neighborhoods without a car, may be genuine. You will be able to defend it only because you gathered the sources around it, and because you are about to open every one.
Snowballing backward and forward from one retrieved source is a documented search method, and for some questions it finds what database queries miss (Greenhalgh and Peacock 2005).
The block below stands in for the reading you do by hand: six retrieved records, what each one found, and the one variable that explains why they disagree. It plants no citations, because a fabricated source is the thing this chapter is teaching you to catch.
import numpy as np, pandas as pd
SEED = 464
rng = np.random.default_rng(SEED)
# Six retrieved records on polling-place distance and turnout. NO fabricated
# citations appear here: these are anonymous stand-ins for records you open
# yourself. The one thing that differs is what each record measured.
measured_switching = np.array([False, False, False, True, True, True])
effects = np.round(np.where(measured_switching, -0.2, -1.5)
+ rng.normal(0, 0.35, size=6), 2)
print(pd.DataFrame({"record": [f"record {i}" for i in range(1, 7)],
"turnout change (pp)": effects,
"measured mail switching": measured_switching}).to_string(index=False))
print(f"\nmean where switching was ignored : "
f"{effects[~measured_switching].mean():+.2f} pp")
print(f"mean where switching was measured: "
f"{effects[measured_switching].mean():+.2f} pp")
print("\nthe live disagreement is not noise. it tracks what each record measured")8.4 An AI failure case
The most common failure here is citation hallucination, the taxonomy’s confident fabrication: the model invents a source and presents it with full confidence. In the worked example it produced a plausible author pair, a 2019 date, and a journal that does not exist. You caught it in a minute by searching the exact title in quotes across your library portal and Google Scholar and finding nothing. Fluent confidence is not evidence. If you cannot locate a source where it should live, treat it as fabricated until you prove otherwise.
8.5 It is your turn
You are working inside Studio 3: Ground it in verified evidence. Keep what you write here; the studio’s milestone chapter is where it joins the other lessons’ pieces into one artifact you can defend.
You arrived with a lead question from the previous chapter. This step surrounds it with the conversation it belongs to: a pool of candidate sources, gathered fast, with AI doing the legwork.
The hands-on half of this section lives in the chapter’s companion notebook: open it in Colab with the badge at the top, and work the steps there.
Commit your own answer first, then delegate. And do not treat these as one-shot requests. A search is a loop: you prompt, you read what came back, you tell the tool what was off-target and why, and you ask again. Two or three passes beat one long prompt every time. Some tools now run that loop themselves, browsing and re-searching without asking you between steps. That changes who types the queries; it does not change who opens the papers. Each pass earns you a better list. None of them earns your trust.
Two decisions never leave your hands. Whether a source counts as verified: only reading it yourself settles that. And whether your gap is real: only retrieval that comes up empty earns the word. What your project may claim follows from those two. An AI can point; it cannot certify.
Write your lead question at the top of a fresh page. Underneath it, write the one sentence you already believe the literature says. That belief is what your search will either confirm or embarrass.
Search once yourself before you delegate. Use your library database or Google Scholar and your question’s own words, and keep the single best thing you find. That is your seed.
Ask an AI for five more candidates on the same question, each with authors, year, exact title, and venue. Then run the loop: read what came back, tell the tool which two are off-target and why, and ask again. Stop when a pass stops adding anything new.
Locate. Ask for leads to chase, not a finished bibliography.
Act as a literature scout in [your field]. My question is: [paste it]. List the five published studies most relevant to it, each with authors, year, exact title, and venue, so I can retrieve and verify each myself. Only include work you are confident exists; mark anything uncertain.After running, verify: retrieve every citation by searching its exact title in quotes before you believe it exists. Counters confident fabrication (a real-sounding citation for a paper that was never written).
Snowball your seed by hand. Read its reference list backward for the older work it builds on, and follow its “Cited by” trail forward for the newer work. Add everything that touches your question.
List-to-verify (snowball). Start from one source you have already read.
Here is one source I have verified: [authors, year, exact title, venue]. Name three older works from its reference list it likely built on, and three newer works that likely cite it, each with authors, year, title, and venue, plus one line on why it belongs on the chain.After running, verify: cross-check the “backward” names against the seed’s real reference list and retrieve the rest. A tidy chain can quietly omit the one contradicting study. Counters illusion of completeness.
Put all of them in one candidate list with a status column, and write
leadin every row. Nothing here is a source yet. The next chapter is where they earn the promotion.Log the delegation in your AI Research Ledger, and verify at least one candidate now with a named method from the Verification Guide; primary-source reading is the one that fits. An AI reviewer may run the check with you; the decision to accept or reject stays yours.