To cite Gemini correctly, first identify what you used: a specific response, a generated artifact, Gemini as a tool, or an original source that Gemini helped you find. Cite or attribute AI-generated wording, images, code, tables, or other output that you quote, reproduce, paraphrase, analyze, or incorporate. If Gemini only helped with editing, brainstorming, translation, organization, or feedback, check the applicable policy to see whether disclosure is required.
If Gemini points to a report, dataset, article, legal text, or other publication that supports a factual claim, verify that source and cite it directly. A citation to Gemini identifies the AI interaction when the output itself matters. It does not replace the source that provides evidence for the claim. Current instructor, institutional, journal, publisher, and style-manual requirements take precedence over any example in this guide.
Use a three-way decision: cite incorporated Gemini output, disclose substantive assistance when policy requires it, and cite verified original sources for factual claims.
Choose the right record before formatting a citation
A citation should identify the thing a reader needs to evaluate. A conversation, one response within that conversation, a generated image, Gemini as a general tool, and an underlying publication are different records.
- Specific response or chat: Record the exact product, response or conversation description, generation date, model or version if displayed, and a stable share URL when one is safe and available.
- Generated artifact: Record whether the output is text, an image, code, a table, or another artifact. For an image, also note its description, later edits, and any uploaded or external source assets.
- Gemini generally: Identify the actual tool and responsible organization according to the applicable style. Current APA guidance uses Google as the author for its Gemini example. Do not automatically substitute Google DeepMind across every style or product.
- Underlying source: Record and cite the original publication, dataset, or official document when it supports a statement in your work.
Google’s Gemini models page describes Gemini generally. It does not identify a particular response, artifact, or historical model version. Use a specific share link for a specific output when possible, and never infer a model version from a date or a generic product URL.
APA: choose a specific-chat reference or a general tool reference
Current APA guidance distinguishes a reference to one AI chat from a reference to an AI tool generally. Use the specific-chat approach when the particular interaction matters and a useful retrievable link is available. Use the general tool approach when documenting the tool rather than one response, or when a specific-chat reference is not useful or appropriate.
These are information patterns, not finished references. Replace every bracketed field with facts from your interaction and apply the current APA formatting rules:
- Specific chat: Company. (Year, Month Day). Chat title [Generative AI chat]. Tool or model name. Specific chat URL.
- Tool generally: Company. (Year). Tool or model name [Large language model]. Tool URL.
Do not automatically classify every inaccessible Gemini response as personal communication. APA now provides specific-chat guidance and a general-tool option. If Gemini helped edit, translate, organize, refine writing, generate code, create a table or figure, or support a method, describe that use in the body, Method section, or author note when appropriate. That does not mean every internal editing interaction needs an individual reference entry.
MLA: describe the generated material and identify its container
MLA does not recommend treating the AI tool as the author. Its updated guidance starts with the generated material, then identifies the AI tool as the container. Include the model or version if known, the responsible company, the generation date, and a stable link to the specific conversation or artifact when one is available and safe to share.
In practical terms, describe the response or artifact first, then add the tool, model or version if shown, company, date, and specific share URL. A general Gemini URL is a fallback when a stable output link is unavailable, not an equivalent identifier for a particular response. Use the current revised MLA template rather than copying older examples that treat the general product page as the location of every output.
MLA also distinguishes citation of incorporated output from acknowledgment of functional use such as editing or translation. Its guidance on describing AI use explains why acknowledgment can extend beyond a Works Cited entry. For a generated image, follow the current MLA advice for describing the image and supplying citation or caption information rather than applying a text-response example unchanged.
Chicago: use prose attribution or a note when formal detail is needed
The Chicago FAQ says AI-generated material should be credited. For a brief, ordinary reference, naming Gemini in the text may be enough. Use a note when readers need more detail, such as the generated material, prompt, tool, responsible company, date, or URL.
A note can follow this descriptive pattern: Description of generated material, prompt if needed, tool and responsible company, generation date, URL if available. Treat this as a record of the information to include, not a universal finished citation. The FAQ does not establish an absolute rule that every AI interaction must, or must not, appear in a bibliography. Confirm the treatment required by the relevant Chicago edition, instructor, publisher, or journal.
Record the interaction and check the share link before publishing
Capture the record while the interaction is available. A useful provenance record includes the provider, exact product, visible model or version, generation date, prompt summary, output type, human edits, underlying source URLs, citation treatment, policy decision, and either a safe share URL or an authorized local archive.
Google’s Gemini Apps sharing guidance says anyone with a public share link can view the conversation. The shared page is a snapshot from when the link was created. A link can be reshared or continued, and deleting the relevant chat or activity can make it unavailable. Treat public sharing as a disclosure decision. If the chat contains personal, client, student, patient, confidential, proprietary, or regulated information, do not publish the link. Preserve an authorized local record and use the applicable style’s fallback.
Google also documents interface options for exporting Gemini responses to supported products such as Google Docs, Gmail, Sheets, Colab, and Replit, subject to account and Workspace conditions. These documented export options are not evidence of a general Gemini Apps conversation-history API or automatic citation integration.
Verify AI-supplied sources instead of citing them blindly
Gemini may display related sources, but Google says the Sources control is not guaranteed for every answer. A displayed link is a lead to investigate, not proof that a claim is supported. Open the source, confirm that it exists, compare the relevant passage with Gemini’s statement, and cite the original source when you rely on it.
Google also warns that Gemini can produce inaccurate information and says responses do not represent Google’s views. Do not present Gemini output as an official Google position, and do not treat the presence of a source panel as evidence that the answer is accurate.
The Gemini record identifies an AI interaction when the response or artifact is part of your work. The original publication, dataset, or official document supports the factual claim. Store and cite them as separate records, then decide whether the final work needs one citation, an AI-use disclosure, or both.
Separate citation from AI-use disclosure
Citation answers, “What output or source should a reader identify?” Disclosure answers, “How did AI assistance shape this work?” The answers can overlap but are not interchangeable.
- If Gemini-generated wording, an image, code, table, or other artifact appears in the final work, cite or attribute it according to the applicable style and policy.
- If Gemini only edited, translated, organized, brainstormed, or provided feedback, disclose that assistance when the applicable policy requires it. Do not automatically create a reference entry for every internal interaction.
- If any generated wording or artifact was substantially edited, describe the human editing accurately rather than presenting the final result as untouched AI output.
- If Gemini supplied a factual source, verify and cite that source directly. Add an AI citation only when the Gemini response itself is also being used or discussed.
A proposed provenance workflow for teams
The following is an editorial and research operations design, not an official Google, APA, MLA, or Chicago schema. It is intended for teams that need repeatable records across documents containing multiple prompts, regenerated answers, images, or underlying sources.
1. Define the record grain
Keep a conversation, individual generation run, output artifact, underlying source, and final document distinct when they have different evidence or decisions. A regenerated answer is a new run even when it appears in the same conversation. An AI-generated paragraph and image should have separate output records if they are separately cited or reviewed.
2. Capture bounded fields
Use ordinary validation for provider, product, generation timestamp, visible model or version, prompt summary, output type, share URL, local archive location, human use, human edits, citation style, policy status, reviewer, and review status. Record “not displayed” for an unavailable model version. Do not infer one from a date or a general Gemini URL.
3. Verify before approval
For each underlying source, store the original URL and a source-verification status. A reviewer should confirm that the source exists and supports the cited claim. A broken source link, guessed model metadata, unclear output grain, or unsafe share link should return the record for correction rather than move it to approved.
4. Prevent duplicate records safely
Use a provider-issued or application-generated interaction or run identifier when available. For a local transcript without a stable link, a proposed key can combine product, generation timestamp, and a cryptographic hash of the preserved transcript or output. This is an implementation recommendation, not a vendor-published identifier.
When concurrent writers or automations can submit the same record, enforce uniqueness in the database and use a transactional upsert. A lookup-then-create sequence alone is not race-safe and can create duplicates. The key must match the row grain: a run key should distinguish regenerated runs, while a source record should retain its own source URL and verification state.
5. File the approved decision
Place the final APA reference, MLA Works Cited entry, Chicago note, figure note, or disclosure in the document location required by the applicable policy. Keep the supporting transcript, export, or archive record in the approved system of record. If no safe share URL exists, document that limitation and use the relevant style’s fallback instead of inventing a link.
{
"record_grain": "one generated response run",
"provider": "Google",
"product": "Gemini Apps",
"run_id": "application-generated unique identifier",
"model_version": "not displayed",
"generation_timestamp": "actual timestamp",
"output_type": "text",
"share_url": "specific link only if safe and available",
"source_verification_status": "not applicable",
"policy_checked": false,
"review_status": "captured"
}This illustrative record describes one run, not an entire document. A project with several prompts, regenerations, outputs, or sources needs related records at those grains. AI may help summarize a long prompt history into a reviewer-checkable description, but it should not invent dates, model names, URLs, identifiers, or verification results.
- The record identifies one conversation, run, artifact, or source at a clear grain.
- Product, date, and model details match the actual interaction; unavailable version data is not guessed.
- Underlying sources are separately recorded and verified when the work relies on them.
- Citation and disclosure treatment match the current style guidance and applicable policy.
- A public link does not expose restricted information, and a named human reviewer has approved the record.
Check the current guidance before submission
APA, MLA, and Chicago guidance can change independently, and a publisher or institution may impose additional requirements. Before submitting, review the current official guidance, the exact assignment or publication policy, and the privacy conditions of any share link.
Teams defining ownership, review gates, and an approved system of record may benefit from workflow and systems consulting. The process described here is designed to make citation decisions auditable without pretending that consumer Gemini Apps provides a universal automatic citation API.
For reference, consult APA’s generative AI citation guidance, MLA’s revised generative AI guidance, and the Chicago AI documentation FAQ.