Build the panel
Choose questions by intent: discovery, comparison, decision, verification and brand. Preserve exact wording.
GUIDE / PROTOCOL
AI citation observatory
Measure answers observed in a defined context without turning a screenshot into a visibility promise.
Direct answer
How can AI citations be measured credibly?
Use a stable question panel, date, identified interface, exact answer and links actually visible. Then compare accuracy, citations and change over time.
01 / Review method
Review method
Capture context
Record date, country, language, interface, relevant settings and whether an account was used.
Qualify the answer
Separate brand mention, cited link, source actually used, answer position and factual accuracy.
Compare series
Observe differences across engines and periods; inspect source pages, corrections and possible causes.
MEASUREMENT / VISIBILITY SOURCES
Impression, citation, visit: three different observations
Appearing in an AI answer and receiving a website visit happen at different stages. Choose the signal that answers your question before comparing periods.
- Be accessible
Crawling may make a page available for search.
- Be shown or cited
A link can appear in an answer without receiving a click.
- Receive a visit
The site records an arrival within the limits of its measurement setup.
| Observation source | What is measured | Interpretation limit |
|---|---|---|
| Google · Generative AI in Search Console Official documentation | Link impressions in AI Overviews and AI Mode, with dimensions including page, country and device. | Days use Pacific Time (PT). Property totals can differ from the sum of page rows because of aggregation. |
| Bing Webmaster Tools · AI Performance Official documentation | Visible citations and cited pages in supported AI experiences; grounding queries are grouped. | A citation is neither a click, a ranking nor a measure of page importance. Grouped queries do not reproduce individual prompts. |
| Website measurement | Arrivals and actions recorded under your measurement configuration. | An attributed visit does not represent every impression. Missing referrers or blocked measurement limit attribution. |
| Documented question panel | Citation presence and fidelity across a set of observed answers. | The result depends on questions, mode, language and date. It does not describe all users. |
Google AI report impressions are included in the Search Console Web scope: do not add them to the Web total. Compare visibility, citations and visits separately.
Compare the scope of two observationsDocumentation accessed on .
MEASUREMENT WORKSHOP
Being cited and being cited accurately are separate measures
Count answers citing the target domain separately from references whose contents support the claim. One answer can contain several links, so the two rates have different denominators.
Fictional example · no measurement of this site or a client
| Indicator | Calculation or threshold | What the result establishes |
|---|---|---|
| Presence in answers | 12 / 40 × 100 = 30 % | 12 of 40 answers cite the domain: 30% presence in this sample. |
| Support among checked references | 15 / 20 × 100 = 75 % | 15 of 20 references support the claim: 75% of checked references. This says nothing about unchecked references. |
From observation to decision
Fix the panel
Keep questions, language, service, mode, dates and repetitions. Include answers with no citation.
Open each reference
Match the exact passage to the claim; record inaccessible links and citations that do not support it.
Compare a constant scope
A change requires the same panel and protocol. Publish counts and exclusions too.
Ten unreadable answers change the meaning of the rate
Of 30 attempts in one programme, 20 answers are readable and 12 cite the target. The other 10 results remain unclassified.
| Observed unit | Fictional data |
|---|---|
| Attempts / readable answers / citations | 30 / 20 / 12 |
| Rate among readable answers | 12 / 20 = 60 % |
| Bounds across all attempts | 12 / 30 = 40 % ; 22 / 30 ≈ 73.33 % |
Does the observed 60% apply to all 30 attempts?
60% applies to the 20 readable answers. Across 30 attempts, possible assignments of 10 unknowns give about 40–73.33%. This descriptive range is not a confidence interval.
References and method
- Google Search Central · AI features
- Google — Generative AI performance report (Search)
- Bing — AI Performance
Documentation consulted on .
02 / Evidence to keep
Evidence to keep
Stable question
ID, exact wording, intent and check frequency.
Dated answer
Text, capture, interface, language and context.
Verified citation
Clickable URL, accessible page and actual support for the claim.
Recorded gap
Error, absence, entity confusion or uninterpretable variation.
PRACTICAL CHECK
Before reaching a conclusion
A frequent misreading and the check that helps avoid it.
Premature conclusion
A single citation proves neither stable visibility nor system preference. Answers, models, languages and interfaces can vary.
Decisive check
Repeat a fixed query panel, record settings and date results. Separate presence, accuracy, source links and any actually observable traffic effect.
AFTER THE REVIEW
Decide from the evidence
The brand is mentioned without a link
Count presence, not a citation. Keep the answer and check separately whether facts are accurate and assigned to the right entity.
A citation appears then disappears
Rerun the same questions in the same context at several times. Record variability before attributing the change to an edited page.
Decision and review trail
Decision and review trail
What the check must help decide and what to retain so the choice can be reviewed.
Sufficient evidence
Keep a fixed query set, languages, dates, services, full answers and cited links. Measure appearance, accuracy and verifiable citation separately.
Next decision
Correct false information under your control. Repeat the observation after the fix without calling one answer a trend or visibility guarantee.
Detailed inspection cases
Typical situations for preparing a check. They do not describe completed assignments or actual observations.
The model changes mid-series
What to inspect
Compare recorded model identifiers, modes and interfaces across campaigns.
Decision and action
Separate periods before and after the change; repeat a control panel rather than attribute every difference to the site.
Record to retain
Displayed version, control questions and transition dates.
An answer is inaccessible
What to inspect
Separate no citation in a readable answer, generation failure and unavailable service.
Decision and action
Count valid observations separately from unavailable attempts; access failure is not an observed missing citation.
Record to retain
Attempts, readable answers and exclusion reasons.
The reference does not support the answer
What to inspect
Open the cited destination and locate support for each studied claim.
Decision and action
Separate citation presence from genuinely relevant evidence.
Record to retain
Query, answer, consulted URL and passage correspondence.
Review before deciding
These questions complement the method. They help establish what is actually supported while retaining qualifications in the record.
Reproducible comparison
Do before and after records use the same question panel, language, context and collection protocol? Separate site changes from engine variation.
View a worked decision
Fictional example: the state applies only to the criterion and scope described.
- Fictional observation
- The first observation uses 10 French questions; the second uses 12 reworded English questions.
- Declared state
- Observed discrepancy
- Decision and next check
- Do not attribute the difference to the site. Rerun a shared panel and retain the two series as distinct contexts.
Relevant citation
Is each cited URL accessible and does it support the attributed passage? Retain wrong citations and uncited answers, not only favourable results.
View a worked decision
Fictional example: the state applies only to the criterion and scope described.
- Fictional observation
- The assistant cites an accessible page, but the passage says 2023 while the answer claims 2024.
- Declared state
- Observed discrepancy
- Decision and next check
- Retain the answer, URL and conflicting passage. A present citation is not counted as a faithful citation.
Unknowns in the results
Are inaccessible or unreadable answers separated from missing citations? Report counts and the denominator rule.
Documentation consulted : NIST · AI Risk Management Framework ·
View a worked decision
Fictional example: the state applies only to the criterion and scope described.
- Fictional observation
- Of 10 attempts, 3 citations, 5 answers without citations and 2 inaccessible answers are separated.
- Declared state
- Verified within this scope
- Decision and next check
- Report 3/10 attempts or 3/8 readable answers depending on the question, naming the denominator and 2 unknowns.
TOOLS / REGISTERS
Information to record
Count citations only within the observed panel. Retain the number of answers, answers without citations and repeated runs before calculating any proportion.
- Access date and time zone
- Exact URL
- Exact query
- Interface / mode / account
- Language / region
- Observed passage / value
- Independent source
- Decision / next action
- Review owner
- Next review
03 / FAQ
Frequently asked questions
Does llms.txt guarantee citation?
No. Google Search does not use it, including in its generative features. For other systems, check their documentation and observed use: having the file demonstrates neither retrieval nor citation.
Does a mention without a link count?
Yes as brand presence, but it must be separate from a verifiable citation.
How many questions should be tracked?
Enough to cover decisive intents; a stable, reviewed panel is better than a large unverified one.
Can two interfaces be compared from one screenshot?
No. Keep exact question, language, date, interface and context, then repeat before interpreting a difference.
How is a wrong citation counted?
Separate link presence from claim accuracy. A visible citation may support a different statement.
04 / Official references
Official references
05 / Continue reading
Continue reading
Describe a context
DEEPER READING / PRACTICE
Organizing a series of comparable observations
The captures are collected through the water on various assistants. To get a useful reading of it, each observation must be linked to a protocol and an identifiable question.
Updated on .
Priority checks
Identify questions and executions. Keep the answer, sources, platform, language and conditions known, and then use a stable grid to classify the results.
Interpretation pitfalls
A apparent evolution may come from a change of question or protocol. A missing answer and a lack of quotation are two distinct events.
Working record to retain
The observation journal allows you to reconstruct each test and compare compatible series. The modifications of the protocol are recorded with their date and their effect on the interpretation.
EXAMPLE / READING THE NUMBERS
Presence does not always mean accurate representation
Fictional data illustrating the method; not client results or measurements of this website.
| Observation | Count | Interpretation |
|---|---|---|
| Panel questions | 40 | Fixed comparison base. |
| Answers mentioning the brand | 12 | 12 / 40 = 30% of the panel. |
| Answers describing the offer correctly | 8 | 8 / 40 = 20% of the panel; 8 / 12 ≈ 66.7% of mentions. |
Show the interpretation limit
What these numbers do not prove
These observations measure neither market share nor sales. A mention may be inaccurate. Report the panel, date, language and repetitions alongside the result.
PRACTICAL CASE / WORKED EXAMPLE
A quote disappears between two observations.
This pedagogical case shows how to qualify a research result or a generated response. It distinguishes the punctual observation, its verification and the conclusion that the test conditions actually support.
Example context
A source appears in a response and is no longer cited in the next test. The team doubts between a loss of visibility and a variation of the response generated.
Suggested approach
Check question, platform, language and known conditions. Repeat according to a documented rule and compare the proposals and sources. Consider the changes made in the Protocol.
Conclusion and next step
The observation is preserved as a variation of the tested series. A sustainable conclusion requires a compatible set of executions, without transforming two different answers into a general trend.
Questions for deeper practice
How often do you repeat the observations?
Choose a frequency related to the need and the changes that are sought. Documenting this rule and the exceptions allows you to interpret the series without confusing a more intensive collection with a real evolution.
DETAILED WORKSHOP / APPLICATION
Follow a series without silently changing the question
01 / Prepare the necessary information
Assemble the list of questions, execution identifiers and known conditions. Identify the formulation or platform changes that occur between the series.
02 / Carry out the approach
Compare the exact questions before the results. Separate the versions of the protocol and keep the repetitions. Qualify the quotations with the sources and proposals associated.
03 / Examine an alternative situation
If a new question improves the test but changes the intention, open a separate series. The ancient history remains useful with its scope and its initial formulation.
04 / Decide on the next step
The series is consistent when comparisons indicate common conditions and differences. Protocol changes are visible in the same way as results.
05 / Hand over the verification records
Transmit a journal question, execution, answer and judgment. Identifiers make it easier to review a ranking without losing the relationship to original observations.
ANNOTATED WORKING DOCUMENT / REVIEW GRID
A series journal that keeps the protocol changes
The journal combines question, version, execution, answer and judgment. It allows you to follow the quotes while distinguishing observed variation and modification of the test conditions.
On a small screen, scroll the grid horizontally to read all three columns.
| Observation | Supported interpretation | Next step |
|---|---|---|
| Questions and conditions are comparable. | Observations can be approached according to the defined categories. | Examine sources and proposals before describing the differences. |
| The formula changes between two series. | Some of the variation may come from the new protocol. | Open a separate version and keep the old series with its scope. |
| An isolated response is presented as a trend. | The staff and coverage do not support this generalization. | Repeat according to an explicit rule and qualify the unknown. |
Reviewing the working document
The final verification verifies the identifiers and conditions before the graphs. Protocol changes must be visible as well as results changes to avoid misleading comparisons.
REPORTING / EXAMPLE WORDING
Presentation of results: Organize a series of comparable observations
Point to clarify
A new question formulation can open another series rather than silently extend the old one.
Before handing over
Name the versions and explain the dimensions that remain comparable.
Example of a qualified conclusion
Compared observations use the indicated version of the protocol. The results obtained after changing the question are kept in a separate series.