Source access
Does the answer link to accessible, relevant and dated pages? Open them: a citation does not prove that the summary is faithful.
Evidence to keep · Cited URLs and checked passages
GUIDE / COMPARISON
Compare search engines and AI assistants
The same question can produce a list of pages, a synthesis or a recommendation. This guide compares those results without confusing presence, quality and evidence.
01 / GUIDE / COMPARISON
Compare a task, not a brand
Start with a real audience question. Keep its wording stable, set the language, region and date, then inspect several interfaces under similar conditions. The result is not a universal ranking: it shows which differences matter for a specific decision.
Explore platforms02 / GUIDE / COMPARISON
Six dimensions to inspect
For each answer, keep a dated capture, the pages opened and separate notes for what is visible, inferred or still unknown.
Freshness
Are time-sensitive facts tied to an identifiable publication or update? Repeat the test after an important change.
Evidence to keep · Source date and observation dateAccuracy and limits
Separate confirmed facts, interpretation and unsupported claims. Note relevant caveats omitted by the interface.
Evidence to keep · Counterexample or primary sourceLanguage coverage
Do the question, sources and answer align by language? A fluent translation can hide a thin local evidence base.
Evidence to keep · Query and sources by languageUseful journey
Does the proposed page let someone complete the task on mobile, including navigation, accessibility and next action?
Evidence to keep · Completed task or observed barrierData context
Does the interface use the public web, its own index or connected work data? Check the settings before comparison.
Evidence to keep · Search mode and declared scope03 / GUIDE / COMPARISON
A reproducible observation
A short log makes answers comparable over time without treating one observation as a general truth.
- 01
Define the panel
Choose representative questions about entity discovery, offer comparison, fact checking, local search and multilingual needs.
- 02
Fix the context
Record date, device, language, region, account, search mode and exact wording. Keep private and connected sessions separate.
- 03
Open the evidence
Visit cited pages, record useful passages and check important claims against a primary source when possible.
- 04
Decide and follow up
Rank differences by their effect on a decision, correct pages you control and repeat the same panel afterwards.
04 / GUIDE / COMPARISON
Four common situations
The appropriate reading depends on the risk and action required.
A brand is missing
Check indexability, public identity and the queries used before attributing the absence to one assistant.
Read protocolA citation distorts a fact
Archive the answer, open the cited source and locate the passage. Correct a source you control, then observe subsequent answers.
Read protocolA market spans languages
Compare answers and sources in each target language, then check destination pages and their hreflang variants.
Read protocolA choice involves sensitive data
Identify public or connected data, granted permissions and the human recovery path before using the system.
Read protocolWorked cases
Interpret checks through worked cases
Choose a check, inspect the observation and compare the conclusion with available records. Each case links to a method for continuing.
Fictional interpretation cases: these describe neither observed outputs nor measured service performance.
Fictional case · Supporting passageA ranking absent from the passage
Starting observation
The answer says X ranks first. The cited document lists X among three actors without ranking them. The page title concerns the right sector.
Conclusion and next check
The subject matches, but the stated rank is unsupported. Remove it or find a record establishing the ranking and its method.
Record to retain
Keep the exact claim, passage and location; separate topical relevance from support for the fact.
Continue with the methodFictional case · Origin of the factThree interfaces, one report
Starting observation
Three cited pages present the same report R: a catalogue record, a summary and a reproduction. None describes a new collection.
Conclusion and next check
Three access points are not three independent confirmations. Examine report R and separate its data from added commentary.
Record to retain
Link each page to the report, version and passage used; leave any untraced origin unknown.
Continue with the methodMethod reference : W3C — PROV Model Primer · Accessed on
Fictional case · Date and versionA stable URL, a revised conclusion
Starting observation
The copy retained at T1 shows 12. At T2, the same URL shows 21 and a correction note. A current answer still cites the T1 copy.
Conclusion and next check
The older copy documents a retained state, not necessarily the current value. Read the correction and date both states before using 12 or 21.
Record to retain
Keep versions, access dates, correction and affected claim; do not silently overwrite the old copy.
Continue with the methodMethod reference : W3C — DCAT 3 · Accessed on
Fictional case · Applicable scopeA result limited to reviewed records
Starting observation
Of 300 distinct records, only 100 complete records were reviewed; 90 pass. The answer reports 90% success for all 300 records.
Conclusion and next check
The rate describes the 100 reviewed records. Outcomes for the other 200 are unknown; state coverage and selection before generalising.
Record to retain
Keep total, reviewed and passing counts, selection rule and missing states.
Continue with the methodFictional case · Verifiability and accessA reference the recipient cannot access
Starting observation
The analyst can open a private appendix cited in the answer. The recipient has only the link; their access rights have not been checked.
Conclusion and next check
The analyst’s access does not establish the recipient’s. Check the authorised review route or mark review as pending without making a public copy.
Record to retain
Record the document, intended reader, verified access and limitation; exclude credentials and secrets from public reporting.
Continue with the methodFictional case · Recorded configurationTwo settings changed in one trial
Starting observation
The first answer uses mode A and history containing a source. The second uses mode B and an empty session for the same question.
Conclusion and next check
The difference cannot be attributed to mode alone. Prepare trials changing one setting at a time while retaining observable conditions.
Record to retain
Keep the question, mode, session state and every answer; flag settings that cannot be controlled.
Continue with the methodMethod reference : NIST — AI Risk Management Framework · Accessed on
Fictional case · Controlled repetitionThe majority answer does not become evidence
Starting observation
Three repeats of the same question return 5, 7, then 5 under retained observable conditions. No supporting document has yet been reviewed.
Conclusion and next check
The series reveals disagreement. Two identical answers do not establish the correct value; review the records before retaining 5 or 7.
Record to retain
Keep all three outputs and conditions, then link the retained value to the checked passage.
Continue with the methodMethod reference : NIST — AI Risk Management Framework · Accessed on
Fictional case · System identityA client name does not identify the model
Starting observation
An API trial records MODEL-B. A consumer interface is named CLIENT-A but displays no model. The report attributes the API trial to that interface.
Conclusion and next check
The attribution mixes two systems. Describe the API result with its identifier; leave the interface’s model unknown until evidence establishes it.
Record to retain
Separate client, interface, displayed model, mode and date; retain identifiers actually visible.
Continue with the methodMethod reference : NIST — AI Risk Management Framework · Accessed on
Fictional case · Discovery and indexingAn acknowledgement does not prove indexing
Starting observation
The page returns 200. A notification receives an acknowledgement, but internal indexing status is unavailable. A dated query does not show the page.
Conclusion and next check
Access and notification are documented; indexing remains unknown. Absence from this query settles neither internal status nor every possible search.
Record to retain
Keep final URL, server response, acknowledgement, query and settings; separate established and unknown stages.
Continue with the methodMethod reference : Google Search Central — How Search works · Accessed on
Fictional case · Search dependenciesTwo rankings for the same record
Starting observation
Two interfaces order results differently. In both answers, however, the retained fact comes from the same document D. The result supplier is unestablished.
Conclusion and next check
Different rankings create neither a second origin nor an identified supplier. Compare interface, declared provenance and supporting record separately.
Record to retain
Keep both lists, final URLs, document D and explicit attributions; mark the technical dependency unknown.
Continue with the methodMethod reference : W3C — PROV Model Primer · Accessed on
Fictional case · Actions and approvalSuccess announced, state unchanged
Starting observation
The assistant announces an update. The tool returned an error and the final record shows the previous value. No other successful operation is recorded.
Conclusion and next check
These traces do not establish the claimed update. Retain the error and final state, then check permissions, exact operation and retry conditions before trying again.
Record to retain
Link request, proposal, tool call, response and final check; do not substitute a success message for the outcome.
Continue with the methodMethod reference : NIST — AI Risk Management Framework · Accessed on
Fictional case · Language and identityTransliteration brings namesakes together
Starting observation
Two translated records show the same Latin-script name but identifiers ORG-A and ORG-B. B’s original name has not been checked.
Conclusion and next check
Keep two distinct candidates. Check original spelling, register and identifiers before combining records; a translated name does not establish identity.
Record to retain
Keep original and translated names, transliteration, identifiers, jurisdiction and the passage establishing the match.
Continue with the method05 / GUIDE / COMPARISON
Official sources
Compare provenance as well as interfaces
Before counting results as confirmations, distinguish four levels. The examples below help prepare a concrete check.
Interface
The service where the question is asked: different brands may display related results.
Discovery system
The index, search provider, model or bot used. Diversity can broaden coverage without proving a fact.
Document
The page actually opened, its author, date and supporting passage.
Original data
The data issuer or initial observation. Several articles repeating a press release do not create several independent observations.
DuckDuckGo
The help distinguishes Instant Answers, its own resources and links or images largely sourced from Bing. Comparing DuckDuckGo with Bing broadens observed interfaces without ensuring two independent documentary routes.
Read the official referenceBrave Search
Brave describes its own search index. Distinguish this infrastructure, retrieved links and a synthetic answer: a different index does not establish the validity of every page found.
Read the official referenceGoogle describes Search Console as a tool for monitoring a site’s presence in results. Property data and URL inspection answer different questions from a public search alone.
Read the official referenceChatGPT
The help distinguishes web search from searching conversations. The Sources panel can contain cited references and other relevant links: these do not all play the same role in the answer.
Read the official referenceGemini
Google describes sources and related content, including supplied files or connected Workspace content. Some answers have no links; the panel does not establish completeness.
Read the official referenceMicrosoft Copilot
The reference covers Copilot in Edge: browsing context may contribute to an answer depending on the question and settings. It does not establish how Microsoft 365 Copilot behaves.
Read the official referenceThese pointers use official documents consulted on 2 October 2026. They are not performance measurements or user-account trials.
06 / FAQ
Comparison questions
Is there a best assistant?
Only for a defined task, context and date. Models, indexes, settings and sources change.
Does a citation prove accuracy?
No. Open the page and confirm that the cited passage supports the claim.
Should every interface get a separate page?
Start with accessible, original, clear and sourced pages. Then measure differences on interfaces that matter to your audience.
How often should the panel run?
Check consequential or time-sensitive questions more often. Elsewhere, use a stable schedule and retest after changes.
07 / Describe a context
Turn observations into a decision
A question panel, preserved evidence and ranked differences support a proportionate next step.
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 .
Related checks
DEEPER READING / PRACTICE
Compare results using common criteria
Two surfaces answer the same question differently. The useful comparison distinguishes discovery from sources, formulation of the answer, attribution and possibilities to continue the verification.
Updated on .
Priority checks
Keep the request accurate and known conditions. Use the same categories to observe primary source, date, relevance of the passage and possible gap between source and synthesis.
Interpretation pitfalls
Tests performed at different dates can reflect a corpus evolution. A single query does not represent information, navigation and specialized research.
Working record to retain
The comparative table retains gross results and motivated judgment. It specifies the tasks for which each surface was observed and the limits of the test.
PRACTICAL CASE / WORKED EXAMPLE
A synthetic response and a list of links are compared
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
Two surfaces provide very different forms of response. An immediate ranking promotes the most fluid presentation without examining the relevance or access to evidence.
Suggested approach
Defining a common task and appropriate criteria: document found, context, attribution and continuation. Keep raw results and conditions before making a judgment.
Conclusion and next step
The comparison is about the ability to perform the task. The form of presentation of the results becomes a criterion of use among others, without replacing the control of the background.
Questions for deeper practice
How to compare different formats?
Use criteria related to the need rather than a common length or page layout. The same issues of relevance, origin and verification may apply to several formats.
DETAILED WORKSHOP / APPLICATION
Compare the next steps available after the first answer
01 / Prepare the necessary information
Preparing a task, a precise question and relevantity criteria. Identify the document or information expected to verify the usefulness of the result.
02 / Carry out the approach
Examine answers, links, dates and passages. Test the possibility of regaining the primary piece and pursue towards a supplement. Keep the differences in format and conditions.
03 / Examine an alternative situation
If a answer is more pleasant but less verified, describe the two dimensions. Presentation and access to evidence can justify different uses depending on the task.
04 / Decide on the next step
The comparison is motivated when the common criteria are linked to the need. The conclusions remain limited to the questions and conditions effectively observed.
05 / Hand over the verification records
Transmit raw results and grid commented. A new version or other category of question requires a separate observation to extend the judgment.
ANNOTATED WORKING DOCUMENT / REVIEW GRID
Common assessment table for different answer formats
The grid compares relevance, context, attribution and access to documents. Brut results are kept so that the reader can understand the judgment beyond the presentation.
On a small screen, scroll the grid horizontally to read all three columns.
| Observation | Supported interpretation | Next step |
|---|---|---|
| The useful document is found and verified. | The surface achieves the documentary task examined. | Qualify the limits and test other relevant intentions. |
| The answer is clear, but its references are vague. | The presentation does not guarantee the support of the proposals. | Examine the passages or seek additional evidence. |
| Tests use different formulations. | A part of the gap may come from questions. | Repeat a common observation or separate the series. |
Reviewing the working document
The final verification verifies the identity of the task and the conditions. A summary and a list of links can be judged with the same use requirements without being forced in the same format.
REPORTING / EXAMPLE WORDING
Presentation of results: Comparing results with a common grid
Point to clarify
A difference in presentation should not determine the quality of documentary alone.
Before handing over
Check questions, conditions and criteria before summarizing a preference.
Example of a qualified conclusion
The results were compared according to the task and common criteria. The relevance, access to evidence and limits are distinguished from the form of response.