# How to Audit an AI Answer’s Citation Support

> Source: CitedIndex — https://citedindex.com/guides/audit-ai-answer-citation-support
> By: CitedIndex Editorial, Research desk
> Claims last verified: 2026-09-07
> Last updated: 2026-09-07
> Covers: Preserving answer text, product, date, prompt and response context, Separating consumer citations from other Sources-panel links, Mapping individual claims to passages and dates, Classifying supported, partially supported, unsupported and unverifiable evidence, Auditing an explicitly hypothetical multi-clause answer, Handling inaccessible, changed, stale and contradictory source content, Inspecting existing API url_citation annotations without making an API call, Keeping mentions, citations, referral traffic and UI observations separate, Maintaining one canonical method with monthly reviews

Audit AI answers claim by claim: save context, compare source passages and dates, and classify support without mistaking links for proof.

## Check the claim, not just the link

Save the answer and its context, split it into individual factual claims, open each cited source, and compare the relevant passage with the exact wording and time period of each claim. Record a support classification and a reason for every claim–source pair. Treat a working link as the beginning of the audit, not its conclusion.

This is a documentation-based walkthrough. Account-level steps were not exercised by CitedIndex, and no API call was run for this guide. The interface and export details below come from OpenAI documentation; the classification system is CitedIndex’s recommended editorial method, not a provider-defined accuracy score.

Source-check date: 7 September 2026. This is the guide’s initial documentation baseline, not an announcement of a product change.

OpenAI warns that search results and citations can be incomplete, outdated, or incorrect. Its guidance is to open the cited source, check whether it supports the answer, and review its publication or update date. [Read OpenAI’s source-review guidance](https://help.openai.com/en/articles/9237897-chatgpt-search).

## 1. Preserve the answer before investigating

Create one audit record for the particular response, rather than a record for a brand or domain alone. Save:

- The full answer text, including citation markers and their positions, plus a screenshot or other permitted capture of the interface.
- The product and surface: for example, ChatGPT desktop web, a mobile app, an API response, or a monitoring-tool export. Record the displayed model or mode if available; otherwise write “not recorded.”
- The response date and time, your capture time, the exact prompt, and relevant preceding conversation. Note supplied location, language, or other context without guessing hidden settings.
- The citation destinations and any separately listed relevant links. Preserve the original URL and record any redirect destination when you inspect it.
- For an export, the tool name, export date, record identifier, and whatever original response context the export actually contains.

Keep unknowns visible. If you only have an exported URL list, do not reconstruct the answer or assign links to claims from page titles. Request the original answer, citation mapping, and response context; until then, mark the claim–source relationship unverifiable.

## 2. Separate citations from other links

OpenAI says responses using web search may include citations, and selecting a citation opens its source. [ChatGPT citation controls](https://help.openai.com/en/articles/9237897-chatgpt-search).

In ChatGPT, the Sources panel, when available, includes cited sources and other relevant links. [What the Sources panel contains](https://help.openai.com/en/articles/9237897-chatgpt-search).

For this audit, label each item as an observed answer citation, another relevant link, an API citation annotation, or an unknown relationship. Do not promote every Sources-panel entry to an inline citation. If the citation’s scope is ambiguous, preserve its position and note the ambiguity rather than assigning it to every nearby sentence.

Use separate fields for mention, citation, and referral traffic: a mention records the name appearing in the answer; a citation records the observed source link or annotation; referral traffic requires separate visit evidence. Do not infer visits from either of the first two. Keep API response evidence separate from consumer-interface observations.

## 3. Compare one claim with one source at a time

1. Copy the exact claim. Split sentences containing several propositions into smaller units, retaining qualifications such as “all,” “currently,” “in Europe,” or “compared with last year.” Keep the original sentence alongside those units.
2. Open every source attached to that claim. If several sources appear together, make a separate row for each source before writing a combined assessment.
3. Locate and save the relevant passage verbatim, with its heading or other locator. Read enough surrounding text to identify qualifications, negation, and who is making the statement.
4. Record the source’s publication date, update date, and the period its evidence concerns, where available. Record your inspection date separately. Write “not stated” when a date is missing.
5. Compare subject, action, quantity, population, location, conditions, and time period. Ask whether the passage establishes this claim or merely discusses the same topic.
6. Assign a classification and a specific reason. Then record the editorial action: retain, narrow, correct, remove, or hold pending evidence.

Interpretation rule: neither a valid URL nor quote containment proves semantic support. If an answer says a service works for every customer but the quoted passage describes a limited pilot, finding those words on the page is insufficient. Likewise, do not treat an accurate link as endorsement of a brand or proof of every nearby clause.

### Use four explicit classifications

| Classification | Decision rule for this method | Reason to record |
| --- | --- | --- |
| Supported | The inspected passage supports the complete claim at the stated scope and time. | Identify the passage and matching qualifications. |
| Partially supported | The passage supports only part of the wording, or a narrower version. | Name the supported portion and the unsupported extension; split the claim where possible. |
| Unsupported | The inspected source does not establish the claim, or explicitly contradicts it. | Distinguish “no supporting passage found” from “source contradicts claim.” |
| Unverifiable | The available record is insufficient to judge support. | Specify the missing source text, historical version, answer text, or citation mapping. |

These labels assess the evidence relationship you inspected. Do not equate “unsupported by this citation” with “false,” or “supported by this source” with independently established truth. For consequential decisions, add an authority and corroboration review rather than expanding the meaning of the support label.

## 4. Worked example: a citation supports only one clause

Everything in this example is hypothetical: the brand, answer, source labels, passages, and dates are invented. [A] and [C] represent linked citation markers in an imagined answer; [B] represents another relevant link in its Sources panel. These labels are not live links or real evidence.

Hypothetical answer captured on 7 September 2026: “LumenDesk exports reports as CSV and is the fastest reporting tool for every team [A]. It retains audit logs for seven years [C].”

- Source A: hypothetical product documentation, updated 1 September 2026. The relevant passage says: “Reports can be exported as CSV files.” The imagined page contains no comparative performance evidence.
- Source B: hypothetical scheduling documentation, updated 2 September 2026, listed as another relevant link but not cited in the answer. Its passage says: “Reports can be scheduled weekly.”
- Source C: hypothetical retention documentation. In this scenario, the reviewer cannot access its contents and has no saved copy or publication date.

| Claim or item | Evidence relationship | Classification | Annotation and action |
| --- | --- | --- | --- |
| “LumenDesk exports reports as CSV and is the fastest reporting tool for every team.” | Citation A; passage dated 1 September supports CSV export only. | Partially supported | The citation supports one clause, not the comparison. Split the sentence before publication. |
| “LumenDesk exports reports as CSV.” | Citation A: “Reports can be exported as CSV files.” | Supported | Retain the narrow export statement within this hypothetical source record. |
| “LumenDesk is the fastest reporting tool for every team.” | Citation A; no comparative performance evidence in the imagined page. | Unsupported | CSV export does not establish universal superiority. Remove or seek appropriate comparative evidence. |
| “It retains audit logs for seven years.” | Citation C; contents inaccessible and date unknown. | Unverifiable | Hold the retention statement. Do not invent a passage or infer retention from the page title. |
| Source B in the Sources panel | Relevant uncited link about weekly scheduling. | Not a cited claim–source pair | Keep it in the link inventory. It supplies no support for the speed or retention claims in this example. |

Example conclusion: record the brand mention separately from the citation relationships. Do not turn the sentence-level partial result and its two component rows into three independent findings; the first row is a summary. Leave referral traffic unmeasured because this hypothetical record contains no visit evidence.

## 5. Handle missing, changed, stale, or conflicting evidence

- Inaccessible source: record the access problem and inspection time. If no sufficiently complete saved copy is available, use unverifiable—not unsupported merely because access failed.
- Changed source: preserve the current passage separately from any dated historical copy. Without an answer-time version, state what the source supports now and leave historical support unresolved.
- Stale evidence: compare the evidence period with the claim’s time frame. A dated passage may support a historical statement without establishing a “currently” claim. Narrow the wording or hold the current claim when the time match cannot be established.
- Contradictory content: quote the conflicting passage and mark the pair unsupported with a contradiction reason. If two sources disagree, retain separate rows and record the conflict in the combined assessment; do not settle it by counting links.
- Missing date or unclear authority: record that limitation explicitly. Seek an authoritative source when the decision warrants it, but identify it as reviewer-added evidence rather than an original answer citation.

## Optional: audit an existing API export

No paid API call is required for this walkthrough. Use this branch only if you already have an API response or an export containing its original fields.

The OpenAI API guide describes answer text at message.content[0].text and cited-URL annotations at message.content[0].annotations. [API output and citations](https://developers.openai.com/api/docs/guides/tools-web-search.md).

The API guide’s example includes type, start_index, end_index, url and title in a url_citation annotation. [Inspect the documented annotation example](https://developers.openai.com/api/docs/guides/tools-web-search.md).

- Preserve the raw response and the original text alongside the annotations. Retain the response identifier, timestamp, prompt, model, relevant request settings, and tool-call records when available.
- Retain the annotation’s URL, title, and start/end indices together with the exact text they accompany. Use the positions to locate the citation in that response, then perform the same claim-by-claim reading. Do not treat those positions as a verdict on semantic support.
- If a monitoring export renamed or dropped fields, document that transformation. Do not invent missing indices or assume a URL-only row retains the original citation placement.

The API guide describes sources as the complete list of URLs the model consulted when forming its response, unlike inline citations, which show only the most relevant references. [API sources documentation](https://developers.openai.com/api/docs/guides/tools-web-search.md).

Keep that API source list separate from citation annotations, just as you separate other relevant links from consumer citations. Interpretation boundary: an API annotation is not proof that a citation was visible in the consumer ChatGPT interface. Record consumer visibility only from a preserved observation of that surface; do not substitute an API export for it.

## Finish with a reproducible editorial decision

Deliver the saved answer, link inventory, claim–source table, supporting passages, dates, limitations, and proposed corrections together. A useful row should let another reviewer understand exactly why you retained, narrowed, removed, or held a statement without rerunning the prompt.

Review this method monthly; the next scheduled review is 7 October 2026. A documentation change should trigger review and, where substantive, an update to this canonical guide—not a new daily article. Record what changed and why; a new retrieval date alone is not a substantive correction.

## Sources

- [OpenAI Help Center: Searching the web with ChatGPT](https://help.openai.com/en/articles/9237897-chatgpt-search) — consumer citation controls, Sources-panel scope, and source-checking cautions. Source packet retrieved 7 September 2026.
- [OpenAI API documentation: Web search](https://developers.openai.com/api/docs/guides/tools-web-search.md) — API output, citation annotations, and the separate sources list. Source packet retrieved 7 September 2026.

These sources document provider behavior; they are not evidence that CitedIndex exercised an account or evaluated a sample of real AI answers.

## FAQ

### What should I request if a monitoring export contains only URLs?

Request the original answer text, citation markers or annotation mapping, product or API surface, response timestamp, and prompt context. Under this method, leave the claim–source relationship unverifiable until that material is available; do not assign support from a URL or title alone.

### Can I add a better source while correcting the answer?

Yes, as an editorial correction. Label it reviewer-added evidence and preserve the original citation finding. Do not retroactively credit the AI answer with a citation it did not contain.