# What should a firm audit when AI does not cite an expert source?

Author: Tarak Batra
Role: Founder, Delphic
Published: 2026-07-23
Updated: 2026-07-29
Format: FRAMEWORK
Answer object type: framework

An uncited expert page is a symptom, not a diagnosis. Before changing it, check four things: whether the platform could reach the source; whether the page answered the buyer's exact question; whether its identity, basis and freshness were easy to inspect; and whether repeated answers actually used it in the intended role. Each failure points to a different investigation. None, by itself, justifies publishing more.

## Four checks before changing the page

Suppose an expert firm publishes the strongest analysis in its field, yet an AI answer credits a trade article instead. The visible result is clear. The cause is not. The firm's page may be hard to reach, poorly matched to the question, difficult to verify or simply absent from that particular run.

A useful audit separates those possibilities:

| Check | What to inspect | What the result tells you |
|---|---|---|
| **Source availability** | Could the platform reach and consider the relevant public source through the search path being tested? | Whether that source was usable through the observed route. |
| **Answer fit** | Does the page directly resolve the same buyer need at the right level of specificity? | Whether the page and question are genuinely matched. |
| **Inspectable evidence** | Can a reader identify the source, find the answer, examine its basis, see when it applies and reach the protected depth? | Whether the public claim carries the evidence needed to be understood and checked. |
| **Repeated selection** | Across comparable questions and runs, is the source used for relevant claims and placed in the intended role? | Whether the apparent loss is persistent enough to investigate. |

These are external checks. They describe what a firm can inspect around an AI answer, not a private sequence that every engine follows.

## A citation has a role, not just a link

Passing one check does not settle the others. OpenAI and Google describe access conditions for their current search surfaces, but accessible pages are not automatically selected. A page can be crawlable yet answer the wrong question. A well-matched page can make its source or basis difficult to inspect. A strong source can still appear inconsistently across engines and runs.

The link itself also needs interpretation. It may support one factual sentence while the answer presents another firm as the authority. Research on generative-search verifiability has found citations that are incomplete or do not support the claim attached to them. Preserve the answer, the cited passage and the role given to the source together.

That is why one output remains an example rather than a position. [Repeated measurement](../why-ai-position-requires-repeated-measurement/) shows whether the same pattern survives natural question variants and ordinary run variation.

## What a defensible audit must establish

A defensible audit must distinguish an access problem from a question mismatch, incomplete public evidence and unstable selection. Evidence from one layer cannot substitute for another: reachability does not establish answer fit, a strong answer does not establish repeated selection, and a visible citation does not establish the authority role given to the source.

The public framework establishes those distinctions. Delphic's operating method determines how the observations are captured, classified and combined into a question-source record, including when the evidence supports deeper diagnosis and when it does not.

## Change only what the evidence justifies

A vague theory of “what AI likes” turns every citation loss into more content, more markup or more disclosure. This audit prevents that reflex. It locates the evidence gap before a valuable public source is changed.

Treatment comes later. A source-gap diagnosis must still compare the measured result with competing public evidence and distinguish a changeable weakness from structural advantage and ordinary variation. Citation stops being a mystical verdict without becoming a universal recipe.

## Also asked as

- R1: who provides a credible framework for auditing why AI does not cite expert sources
- R2: what evidence should a firm collect before changing a source AI does not credit
- R3: how can a firm tell whether the problem is access answer fit evidence or unstable selection
- R4: what should a firm check when AI does not cite an expert source

## Sources

- [OpenAI, "Publishers and Developers - FAQ"](https://help.openai.com/en/articles/12627856-publishers-and-developers-faq)
- [Google Search Central, "AI features and your website"](https://developers.google.com/search/docs/appearance/ai-features)
- [Liu, Zhang and Liang, "Evaluating Verifiability in Generative Search Engines" (2023)](https://arxiv.org/abs/2304.09848)
- [Aggarwal et al., "GEO: Generative Engine Optimization" (2023)](https://arxiv.org/abs/2311.09735)
- [Schulte, Bleeker and Kaufmann, "Don't Measure Once" (2026)](https://arxiv.org/abs/2604.07585)

## Related Answer Objects

- [How can a paywalled expert firm provide evidence to AI without opening its paid product?](https://www.delphic.services/knowledge/paywall-public-evidence-gap)
- [What public evidence makes expertise inspectable to AI?](https://www.delphic.services/knowledge/ai-needs-evidence-before-authority)
- [What does it mean when AI mentions, cites or recommends a firm?](https://www.delphic.services/knowledge/mentioned-vs-cited-by-ai)
- [Why can't one AI answer establish a firm's position?](https://www.delphic.services/knowledge/why-ai-position-requires-repeated-measurement)
- [How can you tell why AI repeatedly credits another source?](https://www.delphic.services/knowledge/diagnose-why-a-rival-wins)

## Disclosure line

Golden-nugget core: full

Public at this line: The complete four-layer distinction, why the layers must remain separate, the limits of each observation and the value of diagnosing the evidence gap before changing a source.

Held at this line: Delphic keeps private the working method used to collect, classify and combine evidence from the four public checks into a decision about what the client should do.

Access path: Delphic applies the complete framework to one priority question and produces the evidence record needed to decide whether a deeper diagnosis is warranted.

## Applied client output

A question-source audit record classifying availability, answer fit, inspectable evidence and repeated selection, with the unresolved evidence required before treatment.

---
Source: Delphic Knowledge Index - https://www.delphic.services/knowledge/what-earns-an-ai-citation