A Shared AI Conversation Link Is Not a Private Evidence Export

By Daoyu Guan

A shared conversation link can help a colleague inspect an AI answer. It is a poor substitute for deciding who may see a client's evidence, what the shared object contains, and how that object will be retained. Before using such a link in an AI visibility report, establish the account type, sharing scope, update behavior, and deletion limits. A private conversation, a public snapshot, a restricted workspace link, and an exported file are different objects.

The distinction matters even when the answer is commercially flattering. A recommendation can sit beneath a prompt containing a confidential launch date or alongside an uploaded drawing. “Share the result” is too vague an instruction for a reporting workflow.

The link's audience depends on the product context

OpenAI's current conversation-sharing guidance distinguishes personal-account links from managed-workspace links. A personal-account link can be viewed by anyone who has it. Workspace-created conversation links are restricted to eligible members of the originating workspace. The documentation also distinguishes a personal snapshot from workspace conversations that can include later messages.

Google's Gemini Apps sharing guidance describes public links that can be read and reshared by people who receive them. Its shared page is a snapshot rather than a live view of subsequent edits. Some organizational settings can disable sharing.

These are product-specific descriptions reviewed on November 5, 2026. They are not a statement that every enterprise product, mobile interface, or account plan behaves identically. In particular, do not transfer a rule from Gemini Enterprise to the consumer Gemini app because the names sound related.

Preview the shared object, not just its title

A report author may want to show one recommendation sentence. The sharing experience may include a larger conversation or supported attachments. Review the actual preview and available options before creating the link. A familiar title does not establish what a recipient can open.

For a hypothetical factory launch, inspect the prompt for unreleased model names, production dates, buyer names, and drawings. Look at earlier turns as well as the final answer. A harmless last paragraph does not make a preceding customer specification suitable for circulation.

Removing the account name is insufficient. A conversation can identify a business through the text, an attachment, a rare product description, or a combination of details. Anonymous-by-default presentation and confidentiality are separate questions.

Use a fabricated conversation to document the sharing procedure. That allows the team to test its workflow without making a real customer conversation public. Keep the demonstration separate from performance evidence.

Record which object the reviewer saw

An evidence record can contain more than one of these. It should identify each object's role rather than replacing them all with a field called “proof URL.”

For a reported recommendation, retain the exact question, answer, visible sources, timestamp, platform, mode, and entity review separately from any presentation link. The link is an access mechanism, not the complete research method.

Sharing can create another version

Suppose a fictional agency captures an answer on Monday, creates a personal snapshot on Tuesday, and updates the shared snapshot on Wednesday after correcting the conversation. A reviewer opening the same URL on Thursday may inspect a different object from the one first referenced in Monday's report.

Keep the capture time and the shared-object time distinct. Where updates are supported, record them. Preserve the report's actual evidence version in an authorized store, together with the hash of the file used for review. A matching hash establishes that two copies contain the same bytes; it does not independently authenticate the platform or the account.

The same issue runs in the other direction. A snapshot may remain unchanged while the original conversation develops. Do not assume that a new answer is included merely because the old share link still opens. Review the content actually exposed.

If a recipient continues a shared conversation, that continuation is a separate interaction. Its recommendations cannot be backdated to the original capture or counted as another response to an unchanged monitoring run.

Keep a private evidence index that identifies the approved file version used in each report. If a presentation link is replaced, update that index without overwriting the original observation. A colleague should be able to explain which answer supported a published statement even after the report's sharing arrangement changes.

Revocation is narrower than recall

Deleting a shared link can stop future access through that link. It does not pull every saved copy back from every recipient. Both OpenAI and Google describe limits involving conversations copied or continued in recipients' accounts. Google also distinguishes deleting a link from removing third-party posts containing it.

An agency should therefore treat revocation as one control with a defined scope. Keep a record of where an approved link was distributed, who owns it, and how it will be disabled. If a client requests removal, identify the controlled objects and the copies outside the agency's direct control rather than promising universal erasure.

Do not create a real public link merely to see whether an AI search engine discovers it. That mixes evidence handling with a distribution experiment and may expose the underlying conversation. A collection workflow should not require public publication of private client material.

Separate client review from public case-study approval

A client may authorize an analyst to inspect an account without authorizing an agency to publish the answers. It may permit a private report while withholding permission to use its name, logo, prompts, or screenshots in a public case study.

Make those decisions explicit. Specify the approved audience, passages, attachments, dates, attribution, and publication destination. Redact information by checking the resulting object, not merely by placing a black rectangle over one screenshot while leaving the text in a downloadable file.

For a public results article, disclose the question set and denominator, including failures and negative answers, to the extent permission allows. One shared answer naming a business cannot establish an overall recommendation rate. A source link in an answer also cannot establish that the source's publisher was recommended as a service provider.

These are evidence and editorial practices, not a legal conclusion about the ownership of any particular conversation. Contractual, privacy, and copyright questions depend on the actual material and applicable terms.

A practical handover test

Use one fabricated conversation and one authorized, non-public export example. First, write down who should be able to open each object. Then have a permitted reviewer check the contents through the intended access route. Record any restrictions without bypassing them.

Next, compare the portable file with the approved screen record: full answer, exact question, visible citation URLs, and relevant metadata. Check whether the file can be opened after the presentation link is removed from the report. A link working today is not a retention test.

Finally, test revocation on the fabricated share object only. Record what stops working and what remains in a recipient's separate copy. This exercise establishes the procedure's scope; it does not certify the provider's entire access-control system.

Semvi by Xindar should be evaluated through the same distinction. The product team has confirmed complete-answer and citation-URL export. That confirmation is separate from a successful account export, verified client permissions, or permission to publish a customer's results. Buyers can use the product scope reference to identify what still needs an account-level acceptance test.

FAQ

Is a link private if I send it to only one person?

Distribution to one person does not change the link's access model. A personal public-share link can be forwarded. A restricted workspace link depends on the workspace's eligibility and settings. Check the actual type before sending it.

Will locking a ChatGPT conversation revoke an existing link?

OpenAI's current guidance says locking a conversation does not delete an existing shared link. Link deletion is a separate action. Check the current sharing notice rather than assuming that an in-account lock recalls previously shared content.

Does a shared link prove the answer was not edited?

It shows the content exposed by that link when inspected. It does not, by itself, provide an immutable capture history. Keep the original record, version information, and a reproducible comparison.

Should every report include a public conversation link?

No. Choose an access method that fits the client's audience, permission, and retention needs. A public link is optional presentation infrastructure, not a prerequisite for rigorous AI visibility monitoring.

Sources and scope

The product distinctions above use current OpenAI sharing documentation and Gemini Apps sharing documentation. No customer link was created, revoked, or publicly distributed to prepare this article. The launch scenario and handover procedure are editorial examples, not a claim of verified Semvi account behavior.