These paths are proposed workflows, not completed assignments or certification. References inform the checks; facts in your own case still need verification.
Verify a citation produced by AI A response contains a number, attribution or reference. Start with the exact claim, find the document and bound what it supports.
Gather before starting Retain the question, response, language, service and date; isolate one claim at a time.
Separate response from evidence Record the passage and displayed link. Fluent wording or several links do not verify a claim.
Open the method or tool · Separate response from evidence Return to the document Compare title, author, identifier, passage, version and any correction. For a scientific publication, check its status too.
Open the method or tool · Return to the document Write a usable reference Cite the page or passage actually consulted. Include the date and qualifications; the tool formats a reference without verifying the fact.
Open the method or tool · Write a usable reference Outcome to retain A claim linked to a passage and version, with a bounded conclusion and outstanding checks.
When to leave the conclusion open If the document cannot be consulted or lacks the expected passage, mark the claim as unverified.
Useful source profiles Method reference : Crossref · Crossmark 2026-10-09
Prepare a CSV before comparison or joining A file may open correctly while containing ambiguous keys, duplicates or missing values. Check its structure before matching records.
Gather before starting Use a copy; record producer, version, licence, delimiter and column definitions.
Profile the file locally Inspect rows, columns, single or composite keys and duplicates. Profiling verifies neither facts nor representativeness.
Open the method or tool · Profile the file locally Interpret blanks Use the producer’s dictionary to distinguish uncollected, non-applicable and unknown values. Do not automatically turn blanks into zero.
Open the method or tool · Interpret blanks Check the join Check keys on both sides and multiple matches. Compare counts before and after; retain unmatched rows.
Open the method or tool · Check the join Outcome to retain A structural profile and explicit join rule, documenting missing values and losses.
When to leave the conclusion open If a column’s meaning or a key remains unknown, pause interpretation instead of improvising a conversion.
Useful source profiles Method reference : W3C · Tabular data model 2026-10-09
Identify a company without confusing entities A trading name, brand and legal unit can refer to different objects. Start with the jurisdiction and identifier you need.
Gather before starting Specify country, reference date, known name and document to check. French registers do not cover all jurisdictions.
Find the correct record For France, search for the legal unit and establishments using appropriate identifiers. Retain status, date and source.
Open the method or tool · Find the correct record Build an identity record Support each name, domain and relationship with evidence. Separate legal unit, establishment, brand and person; a shared name does not prove identity.
Open the method or tool · Build an identity record Cross-check origins Distinguish independent documents from copies of one database. Compare dates; registration alone does not prove authorisation or solvency.
Open the method or tool · Cross-check origins Outcome to retain A bounded identity, its identifiers, consulted documents and uncertain relationships.
When to leave the conclusion open Keep ambiguous results separate; do not infer a relationship merely from similar names.
Useful source profiles Method reference : INSEE · Consulter et télécharger la base Sirene 2026-10-09
Diagnose weak visibility in Search and AI Separate technical access, answer quality and observed visibility. A valid sitemap or markup does not guarantee appearance.
Gather before starting Choose specific pages and questions; record language, period and available measurements before changes.
Check access and indexability Check status, canonical, robots, rendering and links. Distinguish accessibility from actual indexing.
Open the method or tool · Check access and indexability Connect intent and answer Ensure promise, answer, evidence and next step serve the same task. Strengthen complementary pages rather than multiplying variants.
Open the method or tool · Connect intent and answer Observe within a constant scope Retain queries, services, dates and citations. Compare equivalent observations; one citation or a technical score is not a trend.
Open the method or tool · Observe within a constant scope Outcome to retain A diagnosis separating fixable defects, editorial hypotheses and observed visibility.
When to leave the conclusion open Without search or traffic data, label priorities as hypotheses; no ranking gain has been established.
Method reference : Google Search Central · AI features 2026-10-09
Prepare a migration preserving languages and journeys Continuity concerns URLs, anchors, downloads, translations and forms. Prepare mappings before switching.
Gather before starting Inventory versions, useful URLs, direct entry points and features. Keep a restorable copy of affected files and data.
Map equivalents Record content, language, status and expected destination. A home-page redirect does not replace an equivalent page.
Open the method or tool · Map equivalents Prepare mappings Match each old URL to the correct version, retaining useful parameters and fragments. Check targets and rollback before publication.
Open the method or tool · Prepare mappings Test journeys after switching Test direct access, language, keyboard, search, deep links and files. Compare served versions with the inventory, including an existing cache.
Open the method or tool · Test journeys after switching Outcome to retain A mapping table and checks per language, with discrepancies, owners and rollback conditions.
When to leave the conclusion open Keep the switch pending if an equivalent destination, backup or essential feature is missing.
Method reference : Google Search Central · Site move 2026-10-09
Evaluate an agent beyond a successful demonstration One successful task does not validate all permissions, quality or recovery. Define an acceptable result and authorised actions.
Gather before starting Use tests without real effects; record model, version, tools, limits and acceptance criteria.
Define cases and failures Test normal results, expected refusal, missing data, unavailable tools and recovery. Keep failures in the report.
Open the method or tool · Define cases and failures Map dependencies Identify data, providers, permissions and fallback options. Test the fallback within its stated scope.
Open the method or tool · Map dependencies Measure the whole journey Separate waiting, tool calls, validation and usable output. Record units, repetitions, timeouts and conditions.
Open the method or tool · Measure the whole journey Outcome to retain A record of quality, permissions, dependencies and timing, including unmet criteria and a decision bounded by the test.
When to leave the conclusion open If a test requires a real external action or unauthorised access, prepare a simulation before continuing.
Method reference : NIST · AI Risk Management Framework 2026-10-09
