Save the Answer With Its Sources — A Bibliography for What the AI Just Told You

An AI answer with citations is a research object, and research objects deserve a file. Copy the answer, the URLs it cited, and a Wayback snap of each into one plain-Markdown note whose name is the question you asked. You now have an audit trail — a record of what the answer was based on and what has rotted since. It is not a fact-check that the answer is correct.

The chat session was never going to hold this. You clear history. You switch tools. The URL the model cited yesterday returns a 404 today. Pew Research found that "38% of webpages that existed in 2013 are not available today" and that "A quarter of all webpages that existed at one point between 2013 and 2023 are no longer accessible"¹. The AI cited the web the web is losing. Save the answer with its sources or accept that neither will be there next month.

The belief: "the AI cited its sources, I'm covered"

The reassuring thought is that citation is closure. Perplexity numbers its footnotes. ChatGPT with search links its sources. Google's AI Overviews promise a "snapshot of key information to consider, with links to dig deeper"². Anthropic frames Claude's Citations API in the same shape: "Ground Claude's responses in your source documents. Citations return the exact passages that support each claim, so you can verify answers and surface sources to your users"³. The user reads a footnote and moves on.

Simon Willison put the promise most precisely when he reviewed that Anthropic launch. Grounded citations, he wrote, "even act as a form of fact-checking: the user can confirm that the quoted text did indeed come from those documents, helping provide relatively robust protection against hallucinated details resulting in incorrect answers"⁴. Read that carefully. The safeguard is the user can confirm — the citation makes the fact-check possible, not automatic. The reassurance is doing work the reader has not done yet.

The pivot: citations rot, and some of them never existed

The belief breaks in two directions. First, cited URLs disappear. A vendor GEO playbook estimates that "Between 40–60% of cited sources change from month to month as AI models update, citation patterns shift, and competitors adapt"⁵ — one industry observation on top of Pew's harder finding that a quarter of the last decade of the web is already gone. The URL the model cited last week may be a 404 today or an unrelated page tomorrow. The citation was a promise the address could not keep.

Second, and worse, the citation itself can be fiction. Walters and Wilder's 2023 study in Scientific Reports audited 636 citations across 84 AI-generated papers and found that "55% of the GPT-3.5 citations but just 18% of the GPT-4 citations are fabricated"⁶. Even the citations that pointed at real works were sloppy: "43% of the real (non-fabricated) GPT-3.5 citations but just 24% of the real GPT-4 citations include substantive citation errors" — wrong authors, wrong titles, wrong volumes, wrong dates⁷.

That was 2023. Models have improved. Fabrication has not vanished. Damien Charlotin's public database catalogues legal filings in which hallucinated citations reached actual courts, and Willison flagged it in May 2025 as the practice's growing footprint⁸. Assume the class of failure remains and design for it.

Neither risk is exotic. Both are why "the AI cited it" is a starting point for verification, not the end of one. The saved bibliography is what lets you go check.

The argument: save the received answer as a bibliography note

The correction is not a smarter model. It is a file. Treat every substantive AI answer as a small research object with three parts: the question you asked, the answer you were given, and the sources the answer cited. Save all three, together, in one plain-Markdown note. Name the file after the question so you can find it again by what you wanted to know, not by which tool answered.

The note itself stays on your device; the AI features you opt into send the text you chose to your chosen provider. That distinction is worth keeping straight — the note is yours; the ask was not private. Everything downstream of it is easier when the note is a plain .md file you can read without any app. Steph Ango's line still holds: "Apps are ephemeral, but your files have a chance to last"⁹. The chat window is the app. The dated bibliography note is the file.

Freezing the cited URL is the other move that matters. The Internet Archive's Save Page Now tool takes any URL and returns "a permanent URL for your page"¹⁰. Keep their qualifier rather than translating it into an absolute. The archive is durable-ish infrastructure at civilization scale; the Internet Archive announced in late 2025 that "In October, it surpassed the threshold of preserving one trillion web pages"¹¹.

Paste the cited URL into web.archive.org/save/<url> and paste the frozen link back into your note. If the snap fails — some pages sit behind login walls, some are single-page apps the crawler cannot render — record "capture failed" and move on. Honesty about the gap is better than pretending the snap exists.

The practice: the five-line note you can write today

You already have every tool you need. The whole method fits on five lines of Markdown, kept as one file per question, saved on your own device. Do this once for a real answer you cared about; then do it whenever the answer will outlive the chat window.

# Does the EU AI Act require training-data disclosure for general-purpose models?
Answer (Perplexity, 2026-07-07): Yes. Article 53 of the EU AI Act requires providers of
general-purpose AI models to publish a "sufficiently detailed summary" of training data.

## References
- [EU AI Act — Article 53](https://artificialintelligenceact.eu/article/53/) — Wayback: https://web.archive.org/web/2026*/artificialintelligenceact.eu/article/53 — accessed 2026-07-07
- [Commission Q&A on GPAI transparency](https://digital-strategy.ec.europa.eu/en/faqs/gpai-transparency) — Wayback: capture failed 2026-07-07 (JS shell) — accessed 2026-07-07

Five lines, one file, and the audit trail is done. The H1 is the question — the reader (later, you) can search a folder of these files and find answers by what they were looking for. The body carries the answer verbatim, tagged with the tool that produced it and the date, because the tool matters and the date matters more. The ## References section is a bulleted list of the URLs the model cited, each with a Wayback snap and an accessed date so you know what the answer was based on and what has rotted since.

Three diagnostic questions turn the discipline into a habit. Did the AI cite anything? Some answers arrive naked; flag those and treat them as opinion, not research. Did it cite something that isn't there? The Walters and Wilder rates are why this is worth asking every time — click the link before you save it. Did it cite a login-walled source? If you cannot verify it, your bibliography cannot either; either drop it or note the paywall so a later reader knows what to expect.

The subject here is the discipline, not the tool. Perplexity ships numbered footnote citations by default; ChatGPT with search ships inline source links; Google's AI Overviews render a corroborate-view; Claude's Citations API returns "the exact passages that support each claim"³. Each is a starting point that the discipline turns into a research object.

The bibliography is text you copy. It is not an auto-fetcher — a URL the model cited can host prompt-injection payloads, and we make that case at length in why your note app's plugins are an attack surface. Copy the URL. Read the page yourself. Then archive it.

The corpus around this note grows in the same shape. Save the page, not the link is the human-bookmark version of the same instinct; this piece is its diptych — the URLs the AI cited to you. Footnotes and citations in plain Markdown is the bibliography for your own writing; this is the bibliography for what the AI told you. Keep a log of what you asked the AI is the human-decision log; this is the answer plus its cited sources. Own your AI chat history exports the whole conversation; this pulls one answer out cleanly enough to cite downstream. Together they compose a plain rule: own the ask, own the answer, own the sources.

Frequently asked questions

How do I save AI answers with sources?

Open a plain-Markdown file. Make the H1 your question. Paste the answer, tagged with the tool that produced it and the date. Add a ## References section with a bulleted list of the URLs the model cited, and archive each one via web.archive.org/save/<url> so the citation still resolves if the original address rots.

How do I save a ChatGPT answer with its citations?

Copy the answer text and each inline source link into one Markdown note named after the question you asked. In a ## References section, paste every cited URL, run each through the Wayback Machine's Save Page Now, and record the returned archive link with the accessed date beside it. The saved file works with or without the chat still being open.

What is AI citation link rot?

Cited URLs decay the same way ordinary webpages do. Pew Research found that a quarter of the webpages that existed between 2013 and 2023 are no longer accessible¹. AI answers cite the same web. Whichever percentage you accept for month-to-month churn, the shape is the same: the URL you saw yesterday may be a 404 today, so archive the cited page in your note the day you receive the answer.

How do I keep AI research notes?

Use one file per question. The date and the question become the retrieval keys — if you name files after what you wanted to know, a folder of them is searchable by intent rather than by tool. Keep the note plain text, keep the references in a ## References heading at the bottom, and keep an accessed date on every link.

How do I write a bibliography for AI answers?

The bibliography is the ## References block at the foot of the note: each cited URL as a bulleted [title](url) line, followed by the Wayback archive link and the date you fetched both. Before you cite anything from that block downstream, click the original link and confirm it says what the AI paraphrased. The block records provenance, not accuracy.

Are ChatGPT or Perplexity citations always real?

No. Walters and Wilder audited 636 AI-generated citations and reported that "55% of the GPT-3.5 citations but just 18% of the GPT-4 citations are fabricated"⁶, with substantive errors even in the citations that pointed at real works⁷. Modern models are better; the class of failure remains. Save the URL so you can verify the source before you trust it.

How do I archive a webpage the AI cited?

Paste the URL into https://web.archive.org/save/<url>. The Internet Archive states plainly that "You will instantly have a permanent URL for your page"¹⁰. Copy that archive URL back into your bibliography note beside the original link. Some pages fail to capture — login walls, single-page apps, robots-blocked sites — so record "capture failed" and note the date rather than pretend a snap exists.

An AI answer with a live bibliography is the smallest possible piece of research you own. The chat window will close. The URL may rot. What survives is a plain file, on your own device, whose references section still tells you what the answer was based on and what has rotted since.

The tools we mention — Perplexity, ChatGPT with search, Google AI Overviews, Claude, the Internet Archive — did their share by showing sources at all. The rest is a file discipline that outlives any one of them. To keep each answer and its bibliography as plain Markdown you own, mnmnote.com opens a new file in a tab.