Rexz Bot logo Rexz BotBots and automation, explained plainly
AI Assistants

How to Get an AI Assistant to Cite Sources You Can Check

I asked an assistant for the Swedish rules on invoice retention periods and it gave me a clean answer with a reference: a named section of the Bokforingslagen, a year, and a link to a government page. The answer was roughly right.

A laptop showing a chat answer with footnotes beside a browser window open on a government web page

I asked an assistant for the Swedish rules on invoice retention periods and it gave me a clean answer with a reference: a named section of the Bokforingslagen, a year, and a link to a government page. The answer was roughly right. The section number was wrong, the year was wrong, and the link went to a page that had never existed. It took me eleven minutes to establish all three, which is longer than it would have taken me to look it up myself.

That is the trap. A citation that looks correct costs you more than no citation at all, because it switches off the part of your brain that was going to check. After a year of pushing on this with ChatGPT, Claude, Gemini and Copilot, I have landed on a handful of prompt habits that genuinely raise the hit rate, and one popular fix that does much less than people think.

Why an assistant's citation is a different object

A search engine finds a document and then shows you where it came from. A language model produces text that fits the pattern of an answer, and a reference is part of that pattern. Without a retrieval step, the model is generating something citation shaped: a plausible publisher, a plausible title, a plausible year, a plausible URL. All four can be individually plausible and collectively fictional.

The failure modes are specific enough to recognise. Author names get swapped between real papers in the same field. Section numbers drift by one or two. Dates land in the right decade. URLs follow the real structure of a publisher's site but point at nothing, which is the hardest kind to spot because the domain is real and the path looks like every other path on that site.

When retrieval is switched on, the model has actual documents in front of it and the picture improves. It does not become clean. The assistant can still attribute a claim to the wrong one of five retrieved pages, or summarise a sentence that qualified the claim into one that asserts it. That is a different error than invention, and it needs a different check.

Prompt habits that raise the hit rate

None of these are clever. They work because they make the model do something other than produce fluent prose.

Ask for the quote, not the reference. This is the one change with the biggest effect. Instead of asking for sources, ask for the exact sentence from the document that supports each claim, in quotation marks, with the heading or section it sits under. A fabricated quote is far easier to catch than a fabricated title, and in my experience models invent quotes less readily than they invent bibliographies. When the answer comes back with a paraphrase where a quote was requested, that is a signal in itself.

Separate the claim from the evidence. Ask for the answer first in plain prose, then ask in a second message which parts of it the assistant can point to a source for and which parts it cannot. Splitting it is what produces useful hedging. Asked in one breath, the model writes a single confident block. Asked in two, it will often tell you that two of its five claims are general knowledge rather than sourced.

Give it permission to find nothing. A line as blunt as: if you cannot find a specific source, write no source found rather than guessing. This sounds too simple to matter. It matters a lot, because the default behaviour is to complete the shape of the answer you asked for. I tested this on 30 questions where I knew a clean public source existed for about half. With the permission line, the assistant reported nothing found on 11 of the 15 unsourceable ones. Without it, on the same 15, it offered a reference every single time.

Ask for publisher, title and date instead of a link. Links are the easiest thing to fabricate and the most tedious to verify. A publisher name plus an exact title plus a year is something you can paste into a search box in five seconds, and if the title does not exist, you know immediately. I now explicitly say no URLs unless you retrieved the page in this conversation.

Name the kind of source you will accept. Primary legislation, the vendor's own documentation, a peer reviewed paper, the official statistics office. Being specific narrows the space the model is working in. Asking for sources in general invites a blog post that summarised the thing you wanted.

The counter-take: turning on web search is not the fix

The standard advice at this point is to switch on browsing or use a search grounded product, and treat the output as verified. I use those modes daily and I think that conclusion is wrong in a way that quietly causes problems.

Retrieval changes what can go wrong, not whether anything can. The assistant now has to pick which of its results answers your question, and it picks badly in predictable situations: when the top result is an aggregator repeating a figure from somewhere else, when a page has been updated and the model is working from a cached version, or when the real answer sits in a PDF that the tool could not read.

I ran a small test in September on 20 factual questions about pricing and policy pages, where I could independently confirm the correct current answer. With browsing on, 17 of 20 answers were correct and all 20 carried links. Of the three wrong ones, two cited a real page that genuinely contained an outdated figure, and one cited a correct page but reported a number that was not on it. The links made all three look stronger than the no browsing answers, including the ones that were right. That is the problem: retrieval improves accuracy and improves apparent credibility by more.

So use browsing, and keep checking. It shifts your work from does this exist to does this page say what the summary claims, which is faster but not free. The pattern of sounding certain regardless of footing is the same one we picked apart in why an assistant sounds confident when it is wrong.

A verification routine that takes five minutes

For anything that leaves my desk, I run this. It is fast enough that I actually do it.

  1. Open every link before reading the answer again. Not after. Dead links and redirects to a homepage are the cheapest catch and they take seconds.
  2. Search the exact quoted sentence in the page. Use the browser find function. If the words are not on the page verbatim, the quote was reconstructed, and the claim attached to it is now unsupported.
  3. Check the date on the page itself, not the date in the citation. Pricing, tax rules and product limits are the three categories where a correct source from eighteen months ago produces a wrong answer today.
  4. Look for the qualifier that got dropped. Open the paragraph around the quote and read the sentence before and after. Most of the errors I catch are not false statements, they are true statements with the condition removed.
  5. Verify one number independently. Pick the figure the answer depends on most and find it from a second source. If two independent places agree, the rest of the answer is usually safe to use.

If the material is a document you already have rather than something on the web, the whole problem gets easier, because you can ask the assistant to work only from the file and point at page numbers. The method in summarising a PDF accurately applies directly, and page references are far more checkable than citations to the open internet.

One habit I have kept from the invoice retention episode: when the stakes are real, I ask the question twice in separate chats and compare the references. Agreement between two fresh conversations is weak evidence, but disagreement is strong evidence that something is invented, and it costs about forty seconds.

FAQ

Why does an AI assistant invent citations at all?

Because it is producing text that fits the shape of a sourced answer, and a reference is part of that shape. Without a retrieval step there is no document underneath, so the details are generated like any other words.

Can I tell a fake citation without clicking the link?

Sometimes. Suspiciously round dates, a title that reads like a summary of your question, and author names you cannot place together are all hints. Clicking remains the only reliable test.

Are the paid plans better at sourcing?

Paid tiers usually include browsing and larger context, which helps. The difference comes from the tools attached rather than from paying, and a paid model with browsing off still fabricates.

What should I do when no source exists?

Treat the answer as a hypothesis and go to the primary holder of the fact: the vendor, the agency, the registry. For anything with legal or money consequences that is the only acceptable end point, and the AI assistants section has more on where the line sits.

HL
Hedda Lindgren

Hedda covers AI assistants and everyday automation, translating the hype into practical habits anyone can use.

More posts by Hedda

More in AI Assistants

01doorganico 084goad 3hot8 4465456 after5internprogram ahandfulofblessings ahmadpc ahomcybersecurity allodialresources annodex antiquesandperiodcollectibles aotc asianworldkashmir beedlectrics bloggertemplatefree bluefeatheragency bodyweightsolution campaign08blog carfixadvisor catstreetcollective cbstvsports cnfanspurchasing committeeof500yearsofdignityandresistance covertopsdjs cs-kaiyuangame cuestacomunicaciontotal cxoutlet dancebetgame deskgearhub destinyvillastay detectmagazine di33 diamondsteelbuildings donmo dyanadropship emarsysfinance emimattl escapechan eskitarih evautoplay everythingweddings-orlando farmzone feedviewerapp foxstudiosscratch gamevui123 gleasongrade greetingsday hacklabasu haskell iw-kaiyun jagoanproperti jilidskripsi joltbusinessfinance juegodepokergratis juegospucca kaeru-festival latestmehndidesigns lestanzediottaviano litemagazines localcommunityhero lustrsazan mariashebets menswheels newwebstechs1 nikkit ootheday oxigenvpn plurfitness prefetial probudgetcarpetcleaning psychicmrsmartin pyaten qvantor rejoignezlogis riselayouts ruchin rwxqsgd5 sakaltime shibiaowang shop-bot skrtips spotlesshaven strcoding tetapatrium topselller uneplanteartificielle vipertechcr webtech1223411 wo-kygaming yamaha-mekar yeknava yoybuysheets zenunwa yamaha-mekar jilidskripsi ankertoto ankertoto ankertoto