# How to measure AI citations and visible brand attribution

> A repeatable rubric for checking whether an AI answer consulted, linked, displayed, or named your source. Includes a blank ledger and calibration examples.

By George Kelly · Published 2026-09-22 · Updated 2026-09-22

Publication: iamgeorgekelly (https://www.iamgeorgekelly.com)

Canonical: https://www.iamgeorgekelly.com/field-guide/ai-citation-observation-protocol

## Short answer

To measure recognition in AI answers, save the answer and record separate observations: whether your source was consulted, whether it was cited, whether its link or domain was visible, and whether your brand or author was named. A citation can serve an awareness goal without a click. Keep the evidence and the observation conditions; none of these counts measures a person's recall.

## A source can be used without being visibly credited

OpenAI's Responses API distinguishes consulted sources from inline citations. The sources field can include more URLs than the answer cites; url_citation annotations identify cited URLs. A source-list entry is therefore insufficient evidence of a citation. An annotation also needs a rendered answer to establish what a person could see. These API fields describe an API run, not the consumer ChatGPT interface.

Google's generative AI Search report measures links shown in AI Overviews and AI Mode. It is useful for exposure without a click, but it does not tell you whether an answer spelled out your author's name. Its data overlaps ordinary Web performance; Discover is separate. Do not add these report totals to manually collected answer observations.

Sources: [OpenAI API: Web search](https://developers.openai.com/api/docs/guides/tools-web-search); [Search Console: Generative AI performance report (Search)](https://support.google.com/webmasters/answer/16984139)

## Record seven signals without turning them into a score

This rubric is an editorial convention, not a metric supplied by an AI platform. The signals are independent: an answer can name a brand without a link, or link to an article without displaying its author. Record unknown when the required evidence is unavailable. A Sources panel may list consulted pages as well as cited pages; a visible link there establishes source-link visibility, not necessarily an inline citation supporting a claim.

For each visible signal, record first appearance as initial_answer, expanded_sources, absent, or unknown. Initial means the complete saved answer before opening an additional source panel, not a claim that every word was in the first viewport. Expanded means it appeared only after opening the source interface. Preserve the capture state.

| Signal | Evidence needed | Keep separate from |
| --- | --- | --- |
| Source consulted | Complete available retrieval metadata containing the target URL | A citation or a displayed credit |
| Citation in API metadata | A target URL in citation annotations from the saved response | The application's actual rendering |
| Visible source link | Saved answer/source panel and a verified target destination | Visible domain text |
| Visible domain | The target domain is readable in the saved answer or source panel | A hostname present only in the link destination |
| Brand named | The target brand is explicitly named and its identity is unambiguous | An author name or an inferred brand association |
| Author named | The target person's name appears with sufficient context to identify them | A generic source title or another person with the same name |
| Attribution supported | The credited claim agrees with the retained source passage | Presence of a link alone |

[Download the rubric, blank record, and synthetic calibration cases (JSON)](https://www.iamgeorgekelly.com/downloads/ai-citation-observation-protocol.json)

## Capture the answer before inspecting its sources

Choose the question panel before collecting results. Keep discovery prompts free of the target name, domain, supplied page, and instructions to cite that target. Branded questions and tests that provide a source can be useful, but label them separately. They test response behavior after introduction, not independent discovery.

For each planned run, retain the exact prompt, surface, visible model label, capture time, locale, search setting, prior-context condition and run status. Save the initial answer before opening source panels, then capture the expanded state. If the product exposes no model label or retrieval metadata, record unknown instead of guessing.

Verify citation destinations. A title resembling your guide does not establish a match; example.com.evil.test is not example.com. Declare accepted host aliases before scoring, preserve the original URL, and record any inspected redirect. A common author name needs context; uncertain identity remains unknown.

Retain failed runs and answers with no target source link. A completed, fully inspected answer can establish absence for that observation. An incomplete capture or an inaccessible source panel cannot. Use a saved copy or evidence reference that remains available to the reviewer.

[Download the blank observation ledger (CSV, no measured results)](https://www.iamgeorgekelly.com/downloads/ai-citation-observation-ledger.csv)

## Calibrate reviewers on the ambiguous cases

The following situations are invented examples. The downloadable protocol gives their expected labels so two reviewers can compare decisions before reviewing real answers. They are not observations of this website or any assistant's actual behavior.

| Synthetic situation | Record | Do not claim |
| --- | --- | --- |
| API sources contain the target; complete citation annotations were inspected and contain no target URL; no rendered answer is retained | Consulted yes; API citation absent; visible signals unknown | The user saw the source |
| A source chip links to the target article but displays only its title | Visible link in initial answer; domain and names absent after full inspection | The author was named |
| The initial answer has no target; opening Sources reveals its domain link | Link and domain appear in expanded sources | The domain was displayed in the initial answer |
| The answer names the target author and links to the guide, but attributes an unsupported result to it | Author and link present; attribution misattributed | Accurate recognition |
| The answer names the target brand without providing a link | Brand present; source link absent; retrieval unknown | A citation or a website visit |

## Use a denominator that matches each signal

Count an answer once per target and signal, even if it has several matching links. For a given surface, prompt panel and capture procedure, the observed source-link rate is answers with a verified visible target link divided by completed answers whose link state was fully inspected. It includes links in expanded Sources lists and is not an inline citation rate. Report initial-answer and expanded-only links separately. Exclude unknowns from that signal's denominator and show their count beside it.

For a fictional panel with ten planned runs, two failed and one completed answer with an inaccessible source panel, suppose three of the seven fully inspected answers link to the target. Report 3/7 inspected answers, plus two failures and one incomplete inspection. Do not report 3/10 as measured absence, or 43% as market share. The unknown answer could change the result.

Use the same method for visible domain, brand and author, with each signal's own known denominator. Report unsupported attribution separately. These are sample observations, not unique people, platform-wide citation share or proof of brand recall. Keep automated API observations separate from consumer-interface observations.

Next step: [Design the broader question panel](https://www.iamgeorgekelly.com/field-guide/ai-visibility-measurement) — Choose stable questions and an observation cadence before collecting a baseline.

## Pair reported exposure with observed attribution

Keep Google AI link impressions as a separate exposure series. Google applies standard impression rules to AI Overviews and AI Mode: an impression can mean a link was seen or potentially seen; expandable widgets can require expansion. The report is not a record of attention. Property and page aggregation differ, and AI Mode follow-up questions count as new queries.

Use the answer panel to understand the kind of credit appearing alongside that exposure. If links increase while author mentions do not, the evidence supports source visibility more strongly than personal-name recognition. If the source is misrepresented, address the ambiguous passage and retain the old and new versions. If neither exposure nor citations change, inspect the sources actually selected before producing another batch of pages.

For an awareness objective, a useful citation with no click can count as progress. State that progress in its measured unit. A saved answer naming the author establishes one observed attribution; it does not establish that a reader remembered the name.

Next step: [Read the Google AI impression report](https://www.iamgeorgekelly.com/field-guide/search-console-generative-ai-report) — Preserve property, dates, filters and aggregation when reviewing provider-reported exposure.

Sources: [Search Console: What are impressions, position, and clicks?](https://support.google.com/webmasters/answer/7042828?hl=en)

## Sources and scope

- [OpenAI API: Web search](https://developers.openai.com/api/docs/guides/tools-web-search) — Distinguishes consulted sources from citation annotations in API output. Does not establish a consumer ChatGPT observation or a site's actual citation volume. Checked 2026-09-22.

- [Search Console: Generative AI performance report (Search)](https://support.google.com/webmasters/answer/16984139) — Defines AI link impressions, Search product scope, Web-report overlap and report aggregation. The answer-observation rubric here is editorial. Checked 2026-09-22.

- [Search Console: What are impressions, position, and clicks?](https://support.google.com/webmasters/answer/7042828?hl=en) — Explains seen or potentially seen links, standard impression rules, aggregation and AI Mode follow-up queries. Impressions do not measure human recall. Checked 2026-09-22.

## Related reading

- [Measure AI visibility without confusing it with traffic.](https://www.iamgeorgekelly.com/field-guide/ai-visibility-measurement/index.md)

- [AI visibility metrics: impressions, citations, referrals, and outcomes](https://www.iamgeorgekelly.com/field-guide/ai-visibility-metrics-reference/index.md)

- [How to read Search Console’s generative AI report](https://www.iamgeorgekelly.com/field-guide/search-console-generative-ai-report/index.md)

- [How to find referral traffic in GA4, including AI](https://www.iamgeorgekelly.com/field-guide/ga4-ai-referral-traffic/index.md)

## Cite this guide

George Kelly. How to measure AI citations and visible brand attribution iamgeorgekelly. Updated 2026-09-22. https://www.iamgeorgekelly.com/field-guide/ai-citation-observation-protocol

Library navigation: [Choose a task](https://www.iamgeorgekelly.com/field-guide/start-here) · [Source directory](https://www.iamgeorgekelly.com/field-guide/sources) · [Editorial policy](https://www.iamgeorgekelly.com/for-agents#editorial-policy)

## Editorial note

AI-assisted research and drafting. Provider-specific claims link to primary sources. Frameworks are editorial proposals; worked examples are illustrative and are not employer performance results.