Question corpus
Brand, expertise, comparison, risk and decision questions tested consistently.
GUIDE 1401
LLMO & citations in AI answers
Observe what assistants claim, understand the sources they use and strengthen a verifiable public representation.
Operational briefLLMOreadability
CITEprovenance
EVALmeasurement
Operational brief
An AI answer is a synthesis, not a primary source.
Models may cite, paraphrase, combine or omit. Strategy is not about writing for an abstract machine, but making important facts available through coherent, identifiable and trustworthy sources. Measurement should track presence, accuracy, citations and stability over time.
AR / GUIDE 1401INDEPENDENT ANALYSIS
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.
AI-generated illustration
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.
A citation per answer does not verify every claim
Four answers contain six checkable claims. Three answers cite a source; passage review supports four of six claims.
| Observed unit | Fictional data |
|---|---|
| Answers with a citation | 3 / 4 = 75 % |
| Claims supported by a passage | 4 / 6 ≈ 66.67 % |
Why not compare these two rates directly?
One denominator counts answers, the other claims. One answer may contain several claims and a reference may support only some. Retain claim-to-passage mappings.
References and method
- Google Search Central · AI features
- Google — Generative AI performance report (Search)
- Bing — AI Performance
Documentation consulted on .
CITATION ARCHIVE
Provenance turns an answer into information.
A citation is useful only when its source, date, context and limits can be recovered. The observatory compares answers over time and documents gaps rather than drawing conclusions from one screenshot.
Cards, pages and markers connected in a documentary composition. Illustrative scene.
02 / Intervention areas
Intervention areas
Multi-model observation
Same wording, dates, assistants and search modes to identify variation.
Citation analysis
Displayed sources, probable sources, indirect reuse, freshness, authority and contradiction.
Entity accuracy
Name, role, locations, expertise, relationships, timeline and separation from namesakes.
Evidence strengthening
Official pages, verified profiles, reference publications, authors, structured data and external confirmation.
Correction loop
Detected error, probable origin, source correction, recrawl, new measurement and documented result.
03 / FIELD NOTES
Signals of unstable representation
An omission or error in an assistant often reveals a broader public-source weakness.
01OMISSION
The entity does not appear
Available sources do not establish relevance, authority or the relationship with the question.
02CONFUSION
Two entities are combined
Identifiers, biographies, profiles, domains and relationships do not support robust disambiguation.
03OBSOLESCENCE
Old information persists
The recent correction is less visible, connected or credible than the historical source.
04WEAK ATTRIBUTION
The answer is right but unattributed
The system synthesises information without preserving an identifiable brand or source.
04 / OUTPUTS
Generative-answer observatory
Useful measurement must be replayable and comparable; one screenshot does not describe a system.
Versioned corpus
Questions, language, intent, user profile, date and test conditions.
Answer journal
Model, mode, answer, citations, captures, errors and confidence level.
Representation score
Presence, accuracy, attribution, completeness, stability and risk.
Source map
Cited domains, dominant sources, relays, gaps and contradiction.
Correction backlog
Actions ranked by impact, control, propagation delay and verifiability.
Change report
Wave comparison, confirmed changes, uncertainty and next decisions.
WORKING RECORDS
Make the dossier verifiable
Three practical records connecting diagnosis to a traceable decision.
Question panel
Minimum content : Fix intents, phrasings, languages, interfaces and dates for repeatable observations.
Check : Does the panel reflect the target audience’s questions?
Citation table
Minimum content : Link answers to cited URLs, exact passages, entities and error types.
Check : Is a citation present, relevant and faithful to its passage?
Iteration log
Minimum content : Record page changes, reasons, dates and later observations without premature causal claims.
Check : Is the change stable across comparable observations?
05 / Method
Measure before influencing
The protocol limits premature conclusions: answers vary and depend on context, language and search mode.
- 01Fix the corpus
- 02Define conditions
- 03Test several systems
- 04Archive citations
- 05Qualify errors
- 06Trace source origins
- 07Correct publicly
- 08Replay measurement
DECISION CHECKS
Decisions to document
Two checks that change the next decision.
The brand appears with the wrong source
Record the exact claim and open the cited page. Correct an error at source; if the page is accurate but misread, make its context more explicit.
One platform differs from the others
Repeat the same question and variants on nearby dates, then compare links, wording and language. One divergence calls for observation, not an immediate site rebuild.
Two approaches to distinguish
Two approaches to distinguish
The right method depends on the question and the decision it must inform.
Entity mention
A name appears in an answer, sometimes without a link or correct context. Save the question and check identity, role and associated entities.
Usable citation
A link points to a relevant source page and the passage actually supports the claim. Check summary fidelity and whether the reader can verify the fact.
Proposed work plan
Adapt this outline to the agreed scope. It describes a proposed method, not a completed mission, measured result or guaranteed schedule. Record your observations and decisions as checks proceed.
AI-generated illustration
Prepare before starting
A question panel, authorized reference pages, languages and tested interfaces. Define what counts as a useful citation.
Scope the work
Check answers the pages actually provide. Identify the stable passage, author, date and limits for each fact.
Apply the method
Make references readable and consistent with visible content. Check links and structured data without adding unsupported promises.
Verify and hand over
Log answer, reference and checked passage for each question. Distinguish name mention, relevant link and actual support for the claim.
Pause or revision conditions
Occasional appearance in an answer guarantees no future citation. Suspend the conclusion if the cited passage cannot be checked.
06 / FAQ
Frequently asked questions
Clear answers on scope, limits, methods and engagement conditions.
Can an AI-assistant citation be guaranteed?
No. Clarity, authority and source availability can be improved, while a third-party model remains outside direct control.
Is one screenshot enough?
No. Answers vary by model, date, language, search mode, wording and sometimes user.
What does answer share measure?
How frequently an entity or source appears within a defined corpus under documented test conditions.
How can a hallucination be corrected?
Identify the claim, investigate its probable source, publish or strengthen the correction, then measure again after propagation.
Should content be produced specifically for LLMs?
Content should primarily be useful, explicit, sourced and technically accessible, with structure preserving meaning during extraction.
Are unlinked mentions useful?
They may support entity understanding, depending on context, source quality and disambiguation.
Which assistants should be monitored?
Those actually used by the relevant audiences, with several systems when the decision matters.
Is llms.txt sufficient?
No. Google Search says it does not use this file, including for its generative features. For another system, maintaining it is justified only by an identified, documented use. HTML pages remain the reference.
When should measurement be repeated?
After major correction, migration, model change or at regular intervals for priority subjects.
Methodological references
Official documents explaining methods, review criteria and their limits. These references do not certify the service or guarantee search rankings.
08 / CONTACT
Open a private channel
State the context and objective. You will receive a clear initial reading.
Turn a need into a defined assignment
A few interventions relevant to this topic. Each amount covers the stated reference scope; the quote confirms priorities and overlapping tasks.
AI representation audit
Reference questions, observations and documented gaps.
Entity : 1 · 6 h
From 480 € excl. VAT / 576 € incl. VAT
Preparation and delivery checks
- What to prepare
- Entity, queries, languages, surfaces and comparable observation conditions.
- Suggested process
- Keep the protocol, record answers and check each citation before comparison.
- Delivery to agree
- Dated answers, checked citations and findings separated from test variability.
- Review for this service
- Link each recorded answer to its query, surface, language and test date.
- Scope boundaries
- Results remain tied to test conditions; future citations are not guaranteed.
Indicative working scope, to confirm in the quotation.
Entity, evidence & public consistency
Identifiers, sources and information alignment.
Entity : 1 · 6 h
From 480 € excl. VAT / 576 € incl. VAT
Preparation and delivery checks
- What to prepare
- Names and variants, official identifiers, reference pages and conflicting information.
- Suggested process
- Keep the protocol, record answers and check each citation before comparison.
- Delivery to agree
- Sourced entity record, distinguished variants and public corrections awaiting approval.
- Review for this service
- Check names, identifiers and associated evidence without merging namesake entities.
- Scope boundaries
- Results remain tied to test conditions; future citations are not guaranteed.
Indicative working scope, to confirm in the quotation.
AEO/GEO answer pages
Direct answers, definitions, limits and references.
Pages : 3 · 6 h
From 480 € excl. VAT / 576 € incl. VAT
Preparation and delivery checks
- What to prepare
- Real questions, expected answer, available evidence and qualifications to retain.
- Suggested process
- Keep the protocol, record answers and check each citation before comparison.
- Delivery to agree
- Pages answering questions directly, with visible references and limits.
- Review for this service
- Review each short answer with its question, references and necessary qualifications.
- Scope boundaries
- Results remain tied to test conditions; future citations are not guaranteed.
Indicative working scope, to confirm in the quotation.
Prepare and review a service
Use these notes to scope the request and review delivery. The quotation defines the agreed work, volumes and schedule.
GEO, AEO & AI visibility
AI-generated illustration
Make information citable and review observed answers in a defined sample.
- What to prepare
- Define questions, languages, tested surfaces and reference sources.
- Suggested process
- Keep the protocol, record answers and check each citation before comparison.
- Delivery to agree
- Agree a record of questions, answers, citations, dates and comparison limits.
- Scope boundaries
- Results remain tied to test conditions; future citations are not guaranteed.
Acceptance check to prepare
When an assistant-facing answer cites a table, preserve cell address, dimensions, unit and observation status. Accurately copied values can still be assigned to the wrong scope.
See scope and pricingDEEPER READING / PRACTICE
Interpret an AI citation within its limits
A assistant takes a page back but fails to attribute a data or omits its reserve. The presence of the link is not enough: it is necessary to examine which proposal is supported by the aforementioned source.
Updated on .
Priority checks
Keep the question, the complete answer, the links displayed and the test date. Open the source to compare the aforementioned formulation to its context and its explicit limits.
Interpretation pitfalls
A quote may be relevant for one phrase and not enough for the rest of the answer. A isolated test does not demonstrate a lasting preference for a source.
Working record to retain
A report distinguishes the aforementioned source, associated proposal, concordance and missing information. Repeating observations with a stable protocol allows you to compare the differences without generalizing them abusively.