Make AI Cite the Source Line for Every Claim
"Cite sources" gets you one citation for a whole answer. Here's how to make AI attach a source quote and location to every factual claim, label quote vs paraphrase vs interpretation, and flag anything unsupported — so checking a claim is a glance, not a hunt.
Package a Source for CitationThe answer that sounds right and cites nothing
You paste a policy into the AI and ask for the key requirements. Back comes a clean, confident list: expense reports must be submitted within 30 days, receipts are required for all purchases, managers must approve expenses before reimbursement. It reads like a faithful summary. Then you try to check one line. Where does "30 days" come from — which sentence, which section? Is it really "all purchases," or did the source say "over $25"? Did the policy actually say "before reimbursement," or is that the model connecting two ideas on its own? Every answer sends you back into the document to find the sentence it should have shown you. The summary looks sourced. It isn't — it's a set of claims with the sources simply left off.
Telling the model to "cite your sources" doesn't fix this; it gets you a source named once at the end, not each claim tied to the line it rests on. The move that makes an answer checkable is finer: make the AI cite the source line for every claim, so each factual statement carries the exact quote and location behind it, and anything the source doesn't support is flagged instead of asserted. This guide is how to build that into the prompt — per-claim citation, a clean split between what the source says and what the model inferred, and no uncited claims slipping through as fact. NewPrompt helps you set the foundation: the Long Input Formatter packages your source with delimiters, citable section labels, and grounding rules — including the rule that a claim the model can't cite is a claim it can't make. But it's honest about the edges: it doesn't read, parse, or OCR your document, detect line numbers, check that a quote is accurate, or verify that a claim is true. You run the grounded prompt on your own source in your own AI tool, and what comes back is a candidate answer whose citations make your review fast — not a verified one.
Why "cite sources" isn't enough
The gap between "cite sources" and a checkable answer is the difference between a bibliography and a map. "Cite sources" gets you the equivalent of a works-cited list — the document exists, it's named, and you still have to read the whole thing to find where any single claim came from. A map does the opposite — it points at the exact spot. That's what claim-level citation adds: each factual statement pointing at the specific line, quote, or section it rests on, so verifying it is a glance, not a search. Without that, a few things go wrong at once:
- The source is named but the claims aren't mapped. You know which document; you don't know which sentence, so every claim is a small research task to confirm.
- "The source says" and "the AI inferred" blur together. A quoted requirement and the model's paraphrase-plus-a-little-extra look identical on the page, and the extra is where the error hides.
- Unsupported claims wear the same confidence as sourced ones. A detail the source never stated is written in the same tone as one it did, and nothing marks the difference.
- Numbers, thresholds, and scope drift quietly. "Over $25" becomes "all purchases," "within 30 days" loses its condition — small shifts you'd catch instantly against the quote and miss entirely without it.
- Checking means re-reading. The whole point of asking the AI was to not re-read the document; an uncited answer hands that job right back to you.
- Missing locations get filled with invented ones. Asked for a citation it doesn't have, a model will sometimes produce a plausible-looking section or line number — a fabricated reference that's worse than none.
Step 1: Bind every factual claim to a quote and a location
The core instruction is simple to state and does most of the work: every factual claim in the answer must carry the source quote it rests on and where that quote is. Not the answer plus a citation at the end — each claim, with its evidence attached. To ask for it cleanly, it helps to say what counts as a claim, because not every sentence needs a citation. A claim is a factual statement the source is responsible for: a requirement, a number, a date, a name, a status, a stated conclusion. Transitions, structure ("here are the requirements"), and obvious framing don't need a source; the facts that carry weight do. So the instruction becomes: answer from the source, and for every factual claim, give the exact quote and its location — and treat a claim you can't attach a quote to as a claim you don't get to state as fact.
That last rule — no citation, no claim — is the spine of grounded answering, and it's exactly what the Long Input Formatter bakes in when it packages your source: it wraps the text in delimiters, labels each section with a citable marker, and adds the grounding rules, including "a claim you cannot cite is a claim you cannot make" and a fixed answer for what the source doesn't cover. The Reduce AI Hallucinations with Grounding resource is that contract in full — source-only, gaps answered with "The source does not say," and every claim gated behind its citation. Both give you a prompt you run on your own source; the tool marks the sections you cite by, and the model does the citing — neither reads your document for you.
Step 2: Separate what the source says from what the model inferred
A citation is only useful if you can tell what kind of claim it's backing, because "supported by the source" means three different things. A quote is the source's exact words. A paraphrase restates what the source says in the model's words — usually fine, but the wording can quietly shift meaning ("should" becomes "must," "over $25" becomes "for purchases"). An interpretation is the model reasoning over the source to reach something the source doesn't state outright — "the policy implies X." All three can be legitimate; conflating them is how a soft inference gets read as a hard requirement. So have the answer label each claim: quote, paraphrase, or interpretation. The label tells you where to look hardest — a quote you spot-check for accuracy, a paraphrase you check for drift, an interpretation you check for whether you actually accept the leap.
This is the split the summary in the intro erased. "Managers must approve expenses before reimbursement" reads as a single sourced fact, but it may be a quote, or it may be the model welding "managers approve expenses" and "reimbursement follows approval" into a causal "before" the policy never stated. Labeled, the seam shows: if it's an interpretation, you get to decide whether the policy really says that or just permits the reading. Ask for the label explicitly, because a model left to itself presents its inferences in the same flat, confident voice as its quotes — and that uniform confidence is exactly what hides the claims most worth a second look.
Step 3: Forbid the uncited claim — and the invented citation
The claims that cause the most trouble are the ones with no source behind them, and there are two ways to handle them badly: state them anyway, or invent a citation to cover them. Block both. Tell the model that a claim it can't tie to the source doesn't go in the answer as fact — it goes in a separate "unsupported / needs review" section, named so you can decide whether it's a real gap, something you'll supply, or a claim to drop. And tell it, in the same breath, not to manufacture a location it doesn't have: if the source has no line numbers, cite the section or paragraph; if it can't find a supporting passage, say so — never produce a section or line number that looks precise and isn't. A guessed "section 4.2" is worse than "no citation found," because it survives a quick skim.
This is where the guide's promise and its honesty meet. Forcing every claim to carry a real citation, and quarantining the ones that can't, is what stops the answer from smuggling in the model's own additions as if the source vouched for them. But it depends on the model attributing honestly, which it can't be trusted to do perfectly — it can still attach a quote that doesn't quite say what the claim says, or cite the wrong section. The rule reduces invented facts; it doesn't eliminate misattribution. That's why "no citation, no claim" is the floor, not the ceiling: it gets the unsupported material out of the body and into a place you can see it, so what remains is at least claiming to be sourced — which is the part you then check.
Step 4: Give the citations a shape you can actually scan
How the citations are laid out decides whether you'll actually use them. Two shapes work. Inline: each claim followed by its quote and location in brackets, good for short answers and prose you want to read straight through. Or a claim table — one row per claim, with columns for the claim, the source quote, the location, the claim type (quote / paraphrase / interpretation), and a review note — which is better the moment there's more than a handful of claims, because it lines every claim up against its evidence where a gap or a weak citation jumps out. If the answer draws on more than one source, group the rows by source so it's obvious which document each claim rests on, and so a claim that silently pulls from two sources at once has nowhere to hide.
Pick the shape before the model answers, because retrofitting citations onto a finished prose answer is how you get the vague end-note version you were trying to avoid. For repeated questions against the same document — a policy you'll query all week — it's worth grounding the whole session once instead of restating the rules each time; the Create Grounded AI Workflows resource sets up exactly that, a packaged source you open the session with so every later answer inherits the same citation discipline. Whatever shape you choose, the citation belongs next to the claim, not in a pile at the end — the point is to make the evidence for any one line reachable without leaving the answer.
Step 5: Ground the source first — then verify, because a citation isn't a proof
Per-claim citation works best on a source that's been packaged for it: the text clearly delimited so the model knows exactly what it's allowed to draw from, the sections labeled so there's something stable to cite, and the grounding rules stated up front — source only, a fixed answer for gaps, no uncited claims. That packaging is the difference between a model that cites because you asked nicely and one that cites because the structure won't let it do otherwise. It's also where you set the real scope: the model cites the section or line marker in front of it — it isn't detecting true line numbers in your original file or looking anything up, so the citation points into the packaged source you gave it, and no further.
Then read the cited answer for what it is: a candidate that's fast to check, not a checked one. A citation next to a claim is an invitation — "here's where I got this, look" — and the looking is still yours. Spot-check the quotes against the source (a model can misquote or attach the right-looking quote to the wrong claim), check the paraphrases for drift, and decide whether each interpretation is a leap you accept. Read the "unsupported" section, because that's where the real gaps are collected. And for anything that carries real weight — a legal, policy, medical, financial, security, or compliance source, where the wording is the whole game — the person who signs off on the reading is a domain owner or professional, not the model. NewPrompt gives you the structure to demand citations and the packaging to enforce them; confirming the quote is accurate, the interpretation is fair, and the answer is safe to act on happens on your side, against the real source.
Common mistakes
The habits that produce a confident answer you still can't check:
- Asking to "cite sources" and getting one citation for the whole answer. Name the requirement precisely: every factual claim carries its own quote and location, not the document once at the end.
- Letting quotes, paraphrases, and inferences wear the same face. Have the model label each claim's type, so a soft interpretation can't pass as a hard quote.
- Stating unsupported claims as fact. A claim with no source belongs in an "unsupported / needs review" bucket, not in the body dressed like the sourced ones.
- Accepting invented locations. A precise-looking "section 7.3" that doesn't exist is worse than "no citation found"; forbid manufactured line numbers and section references outright.
- Trusting the citation as proof the claim is right. A quote can be misattributed or subtly not say what the claim says; citation makes review fast, it doesn't make it unnecessary — and NewPrompt doesn't verify any of it for you.
- Retrofitting citations onto a finished answer. Ask for per-claim citation up front; bolting it on afterward gets you the vague end-note you were trying to avoid.
A worked example: policy requirements, each tied to its line
Watch an uncited requirements list turn every claim into a trip back to the document, then a per-claim-citation prompt attach the quote and location to each one — and quarantine the claim the source never actually made.
An uncited requirements list flattens "should" into "must" and "$25" into "all purchases" with no way to check; a per-claim-citation prompt attaches the exact quote and location to each claim, labels paraphrase vs interpretation, and quarantines the approval claim the source never stated — a candidate you verify against the sourceTHE SOURCE (an expense-policy excerpt, given to the AI):
Section 2, para 1: "Reports should be submitted no later than 30 days
after the transaction date."
Section 2, para 3: "Receipts are required for individual expenses
greater than $25."
Section 3, para 2: "Expenses are reimbursed after manager review."
THE WEAK ASK, AND WHAT IT GIVES BACK:
ask: "Summarize the key requirements from this policy."
answer: "Expense reports must be submitted within 30 days, receipts are
required for all purchases, and managers must approve expenses
before reimbursement."
why you can't check it:
- "within 30 days" -- which line? and the source said "should," not "must"
- "all purchases" -- the source said "greater than $25", not all
- "approve ... before reimbursement" -- source says "reimbursed after
manager review"; "approve" and "before" are the model's wording
- no quotes, no locations -- every claim is a trip back to the document
A PER-CLAIM-CITATION PROMPT:
Answer only from the source text. For every factual claim, give:
claim | source_quote | source_location |
claim_type (quote|paraphrase|interpretation) | review_note
Rules:
- A claim you can't tie to a source quote does not go in the answer;
put it under "unsupported / needs review".
- Do not invent line/section numbers. If lines aren't given, cite the
section or paragraph. If you can't find support, say so.
- Keep the source's wording distinct from your own; label paraphrase
and interpretation as such.
WHAT COMES BACK (a candidate you check):
Claim: Reports are due within 30 days of the transaction date.
source_quote: "Reports should be submitted no later than 30 days
after the transaction date."
source_location: Section 2, paragraph 1
claim_type: paraphrase
review_note: source says "should", not "must" -- guidance vs rule?
Claim: Receipts are required for expenses over $25.
source_quote: "Receipts are required for individual expenses greater
than $25."
source_location: Section 2, paragraph 3
claim_type: paraphrase
Unsupported / needs review:
"Managers must approve expenses BEFORE reimbursement" -- source says
"reimbursed after manager review"; "must approve" and "before" are
not stated. Flagged, not asserted as a rule.
NEXT: you read each quote against the source, decide whether "should" is a
requirement, and rule on the flagged approval claim. The citations made
the check a glance -- they didn't confirm the reading, and NewPrompt
didn't read the policy or verify a line of it. You do, against the source.Where this fits in NewPrompt
Per-claim citation is a construction discipline — something you build into the answer as it's produced — and NewPrompt gives you the structure for it, not the source and not the verification. The Long Input Formatter packages your source into the grounded, citable shape the discipline needs: delimited text, section markers to cite by, and the rules that make "a claim you cannot cite" a claim the model can't make. The Reduce AI Hallucinations with Grounding resource is that strict contract in full, and the Create Grounded AI Workflows resource extends it from one question to a whole session against the same document. Each builds a prompt you run on your own source; none reads it or verifies a citation for you.
This guide is the construction move, and it sits just before its neighbors. Reviewing an AI answer against the source is the check you run after an answer exists — a separate, later pass; this guide puts the citations into the answer as it's written, so that later check is a spot-check instead of a re-read. Keeping the model from blending facts across sources is about not mixing; per-claim citation enforces it as a side effect, because a claim that has to name its one source can't draw from two without saying so. And a faithful summary is the goal a level up — this is the mechanism that makes "faithful" checkable, one claim at a time. The through-line: everything here happens while the answer is being built, not after it's done.
A cited answer is a museum guide who, for every claim, walks you to the case and points: this, right here. The uncited one stands in the middle of the room and describes the collection beautifully — and you have no idea which case, or whether the piece is even on display. Per-claim citation is that walk to the case: it puts you in front of the exact spot fast, so checking a claim is a glance instead of a hunt. What it doesn't do is read the label for you. Whether the quote says what the answer claims, and whether it was pulled from the right case at all, is still something you confirm standing there — the citation gets you to the glass, not past it.