Google Community Perspectives: An Attribution and Context Test Protocol
Google says AI Search can preview public discussions and first-hand sources. This preregistered protocol tests attribution, context, quote support, and destination quality.
Research status: Google says generative-AI Search responses can preview perspectives from public discussions, social media and other first-hand sources, with context such as a creator name, handle or community name. This page preregisters a test of whether the displayed advice preserves attribution, source context and a useful destination.
No query sample has been completed for this publication. The protocol makes no claim about coverage, accuracy, platform preference or the frequency of Community Perspectives.
State the research question
The primary question is: when Google extracts first-hand advice into an AI response, can a reader identify who said it, where it was published, the boundary of the advice and the page that contains the original context?
Secondary questions cover source type, quote match, qualification loss, link destination and market variation. The test does not attempt to reverse-engineer ranking or estimate all eligible discussions.
| Layer | Field | Pass condition |
|---|---|---|
| Attribution | Name, handle or community | Visible and source-matched |
| Passage | Displayed advice or quote | Meaning preserved |
| Context | Conditions and limitation | Material boundary retained |
| Destination | Clicked URL and fragment | Resolves to supporting discussion |
Build a fixed query sample
Create task families where first-hand experience is useful: troubleshooting, local conditions, product durability, accessibility, travel constraints, hobby technique and professional workflow. Exclude sensitive health, legal and safety advice from the first study unless qualified review and stronger safeguards are available.
For each family, save exact queries, phrasing variants, country, language, device, browser, signed-in state, personalization state, date and time. Include absent Community Perspectives as data. Do not keep rerunning only until the feature appears.
Predefine source classifications such as forum, social platform, independent community, publisher comments and specialist first-hand article. Record the public accessibility of the source and whether login or consent interrupts the destination.
Compare the preview with the source
- Capture the complete AI response and visible Community Perspectives treatment.
- Record the displayed creator, handle, community, passage and link.
- Open the destination and locate the supporting source text.
- Compare the subject, action, conditions, date and limitation.
- Classify the match as supported, qualified, mismatched or unresolved.
- Have a second reviewer check material mismatches before reporting them.
Paraphrase does not require word-for-word identity. The test asks whether the meaning and boundary survive. A short preview that drops “only on version 3,” “in Canada,” or “after disabling the cache” can materially change otherwise accurate advice.
Report results without a platform ranking
Publish the sample size, query families, environment, date range, absent results, source mix and classification rules. Show counts and examples only after preserving the raw observation and reviewer decision.
A finding can describe the tested queries and markets. It cannot establish that Google prefers Reddit, forums, creators or one publisher across the whole product. Search availability, personalization and interface labels may change during the study.
Use the citation-ready passage test to evaluate source boundaries and the small experiment method for the saved baseline and stopping rule.
Ask a question or join the discussion