AI Prompts for Research: Every Number Needs a Source Line

Most AI prompts for research demos end the same way. You paste the question, and back come five findings, each with a source name and a URL. The prose reads like a briefing: units, round numbers, a conclusion. What the demo never shows is that none of those five links has been opened yet - and a model asked for a citation does not go quiet when it lacks one. It writes the reference in the same confident register it uses for what it knows. Here are the four slots of a source line, the three ways a citation fails, and the one check that scales: open two.

What AI Prompts for Research Should Return: Four Slots, Not One

A source line is not a name. It is four slots that have to close, and the first two are the ones people skip when a number looks useful.

The Claim, With Unit and Scope

"The average went up 8%" is not verifiable. Up 8% of what, over which period, in which population? A claim without unit and scope cannot be checked by anyone. First pass over the output: complete them before the number travels into a slide.

The Name Behind It

"A recent study" is not a source. The slot asks for a named entity: an institution, a company, a named author. If the model does not have it, the gap stays a gap - a plausible author name reads exactly like a real one, so it buys nothing.

The Date of Measurement, Not the Date of Reading

A 2019 figure with a 2026 access date is still a 2019 figure. The date that belongs in the line is the day of the measurement, not the day the document repeated it and not the day you opened it. Documents that recycle old numbers rarely advertise the vintage; treat a missing date as a warning.

The Place You Can Open

A title without an address is not a pointer. The slot closes only when there is a page you can open today that contains the claim - not a search result mentioning the paper, the page where the number lives. This is the slot that costs thirty seconds and disqualifies most of what a model hands you.

Four slots filled and the claim survives a client question. Three filled and you have a lead, not a fact.

Three Ways a Citation Fails

The Reference That Does Not Exist

There is literature on this. A 2023 audit in Scientific Reports had ChatGPT-3.5 and ChatGPT-4 write short literature reviews on 42 multidisciplinary topics, then collected the 636 citations inside the 84 papers they produced and searched databases for each one. Fabrication and formatting errors were widespread enough to be the paper's headline. What matters here is the mechanics, not the rate: the model produces the shape of a citation - author, year, journal - because the shape is what it learned. Whether the thing exists is a fact about the world, and only the world supplies it. Read the 2023 audit in Scientific Reports.

The Pointer That Does Not Resolve

The domain is real and the page is not. The price arrives with no currency and no date. The survey has no author, only a year. These failures survive a glance because nothing looks broken: the line has every mark of a reference except the ability to resolve. One-step test: paste the link and watch it answer. A citation that does not open is a sentence, not a citation.

Borrowed Authority

The aggregator cites another aggregator, and the article cites "a study" that cites nothing. Authority gets borrowed in chains that usually end at a page which does not carry the number, or carries it with a different method. The rule: the source line ends at the origin, at whoever measured the thing, not at whoever repeated it loudest.

Which is why "be more careful" fixes nothing. The paper Why Language Models Hallucinate (Kalai et al., arXiv:2509.04664) argues that models guess when uncertain because training and evaluation procedures reward guessing over acknowledging uncertainty - a model graded as a test-taker scores better by answering than by abstaining. The practical consequence: make not knowing cheap. Say in the prompt that UNCONFIRMED and "not found" are valid, expected outputs. If the only format you accept is a list of findings with sources, the model fills the list. Read the paper on arXiv.

Check Whether Your Tool Actually Searches

A grounded prompt needs the tool to consult the outside world. Search grounding, as the provider documentation describes it, connects the model to public web content and returns citations plus metadata tying spans of the answer back to the source they came from. A span with no such metadata did not come from the search; it came from memory. One is a lookup, the other a recollection, and recollections do not update.

The thirty-day test settles it. Ask something whose answer changed this month: a current price, a version published last week, a confirmed release date. If the answer arrives with a page that mentions that change and its date, the tool is searching. If its newest citation predates your question, you are reading training memory.

A Confidence Label Is Not Evidence

A confidence field such as High, Medium or Low is a self-assessment, and a self-assessment is not a source. It becomes information only when tied to the fourth slot: High if there is a primary source with a date that opens, Low if all you have is an aggregator repeating a number. Never reward a label without the source beside it.

Open Two Claims, Then Fill the Table

The verification that scales is a sample, not an audit. Take two claims from the output - one number and one statement carrying the conclusion - and open both sources. If both survive the four slots, the rest of the batch earns more credit. If one fails, run the prompt again naming the claim that failed and asking for the source that supports it or for the unconfirmed mark.

The deliverable that makes this permanent is a claim table with five columns: claim, who measured it, when they measured it, where you opened it, verdict. Four rows are enough for a one-page brief, and the table is what lets you defend the number when someone asks where it came from.

Three limits the tool does not cross. It does not open a paywalled paper or check what is behind the wall. It does not know your market, your data or your operations. And it does not certify: a claim with a legal, financial or health consequence gets signed by a qualified professional, not by a language model.

The Ten-Minute Pass

Run this before the brief leaves your desk. Pick two claims. Open their sources and count the filled slots. Delete any claim that does not close all four. Save the table next to the brief. Next time, paste the line "if you cannot verify the source, write UNCONFIRMED" into the prompt before you generate, so the model has a cheap exit from guessing.

Three Prompts That Already Write the Source Line

The catalog holds three prompts that demand this discipline from the model instead of hoping for it. Each covers a different kind of claim, and each is a starting line, not a replacement for the sampling pass.

For Claims Out There

Gemini Grounded Research Assistant answers first with a confidence level, then delivers five to eight key findings each with a source name and a URL, separates the conflicts where sources disagree, lists the three most reliable sources and closes with follow-up questions. Its written rule matters most: never invent a URL, and if a source cannot be verified, say so.

For Market Claims

Gemini Product Comparison Research returns a table where every product carries its price with currency and date, plus key specs and a one-line verdict, then a review consensus with its source and a buy-timing read. Every cell is a miniature source line, which is why a price without a currency stands out immediately.

For Your Own Numbers

Gemini Multimodal Data Analyst takes screenshots, charts and CSV files and requires every insight to declare which file it came from, flags low-confidence inferences and lists inconsistencies between files. Same four slots, applied to evidence you already own.

All three write the line. None of them can open it. That last step - two claims, two clicks, one table - turns a fluent answer into a defensible one. Start with the grounded research prompt for outside claims, and keep the claim table open while you read the output.