Keep a record of AI sources you can actually cite
The model, the date, the working link and the text itself, kept together, so a citation you write in March still checks out in November.
For students, researchers and anyone writing something that has to survive being checked.
The situation
Citation styles have caught up with generative AI. APA's guidance asks for the name of the tool or model and the URL of the chat. MLA asks you to cite the tool when you quote, paraphrase or otherwise use it. Both assume the conversation is still there.
It usually is not. Chats get deleted during a clean-up, share links get revoked, and a platform can change how sharing works between when you write and when someone checks. A citation with a dead URL is a citation that fails.
Then there is the harder problem, which is your own record. Three months into a project you have a paragraph resting on something a model told you, and you cannot say which model, on what date, or in response to what prompt. Rewriting the paragraph is easier than proving it, so people quietly rewrite the paragraph.
What Savelore does about it
- Model and date on every entry
- Each saved item records the platform, the model and when you saved it. Those are two of the three things a citation needs, kept without you doing anything.
- The source link, kept
- Entries hold a link back to the conversation. For a saved response it reopens the conversation and scrolls to that message, so you can show what you were shown.
- Your prompt as well as the answer
- Collection mode lets you highlight your own prompt alongside the response. What you asked is part of the evidence, and it is the part people forget to keep.
- Citations inside a collection
- A collection stitched from several conversations carries numbered citations, each one linking back to the exact message it came from.
- A page a reader can open
- Publish an entry as a public link and anyone can read it with no account and no install. The page is a snapshot, so it does not change under a reader who has already cited it.
- Export for your own records
- Pro exports entries as PDF, HTML, Markdown, TXT or JSON, which is what an appendix or a supervisor's copy usually needs to be.
How to keep a source you can stand behind
Do this at the moment you decide to use something. Reconstructing it later is the part that fails.
- 1
Save the response you are going to use
Save the specific message, not the whole thread, so the record and the claim line up.
- 2
Keep the prompt with it
Use collection mode to include what you asked. A response without its prompt is much weaker evidence.
- 3
Write down what you took from it
The description field is where you record what you concluded and why. Future you is the one checking.
- 4
Note the model, not just the platform
The entry records this, and citation styles ask for it. ChatGPT and GPT-5.6 are not the same level of detail.
- 5
Publish it if a reader needs to see it
A public link is one thing a marker or a reviewer can open without an account. Unpublish it after the review if you want.
- 6
Export a copy for the appendix
PDF for submission, JSON if you want the record machine-readable.
What the style guides actually ask for
APA's reference format for a generative AI chat names the tool or the model and ends with the URL of the chat. MLA's guidance is to cite the tool when you quote it, paraphrase it, or use it functionally, and to say what you asked.
Both are worth reading from the source rather than from a summary, because they have both been revised. APA publishes theirs at apastyle.apa.org, and most university libraries keep a current page on it.
A link is not the same as a record
The URL in a citation points at something the platform controls. It can be deleted by you, revoked by them, or changed by a product decision, and none of those events tell you they happened.
Keeping the text alongside the link means the record survives the link. If a reviewer cannot open the URL, you can still show what the model said and when.
Disclosure, not just citation
A lot of institutional policy is about disclosing how you used AI rather than quoting it. That is harder to do honestly six months after the fact, when you are trying to remember whether the model wrote the outline or you did.
An archive with dates, models and your own notes turns that into something you can read off rather than reconstruct. Tags help here: tag by chapter or by section and the disclosure writes itself.
What this does not settle
Keeping a good record does not make an AI-generated claim reliable, and it does not tell you whether your institution allows the use you have in mind. Both remain your call.
Savelore also does not format citations for you. It keeps the pieces a citation needs, in one place, with the dates intact.
Frequently asked questions
What does a citation for an AI conversation need?
Broadly the tool or model, the date, and a URL for the chat, though the exact format depends on the style. Check APA's or MLA's current guidance directly, because both have revised theirs.
What if the conversation gets deleted?
The link stops working, and the text stays in your archive along with the model and the date. That is usually enough to satisfy a reviewer even when the URL will not open.
Can I give someone a link they can open without an account?
Yes. Publish the entry as a public page. No Savelore account and no extension is needed to read it, and you can unpublish it later.
Does the published page change if I edit the entry?
No. It is a snapshot taken when you published, which is what you want for something being cited.
Can I keep the prompt as well as the answer?
Yes. Collection mode lets you highlight your own prompt and the response into the same entry, each with its own citation back to the message.
Can I export the record?
Pro exports entries as PDF, HTML, Markdown, TXT or JSON. PDF is the usual choice for an appendix, one entry at a time, and JSON if you want the record machine-readable.
Keep the evidence while it still exists.
The conversation you are citing is one clean-up away from being unverifiable. Savelore keeps the model, the date, the link and the text together, across all nine supported platforms, from the moment you decide the answer matters.