Query Evidence Lab
Build the evidence before you tell the story.
Start with an empty workspace, record what an AI interface actually showed, and carry the same fields into every retest. This protocol is a measurement foundation, not a visibility score or a results claim.
Empty by design · Tenten has not run or published the observations in your browser · No benchmark, uplift, or causal conclusion is produced
Local workspace
Record bounded prompt observations
Create up to 100 manual observations. Import a matching CSV, export your local rows, or download the blank header-only template.
Prompt text, notes, and URLs stay in this browser. They are never placed in analytics events or submitted to Tenten.
Rows are saved to this browser's local storage. Clearing site data or using another browser removes access unless you export a copy.
Rows
0
Engines represented
0
Locales represented
0
Protocol-complete rows
0
Completeness band
Empty
0/0 required cells complete. These are documentation summaries only. They do not measure visibility, performance, uplift, or causality.
No observations yet
Add a blank row or import the stable CSV template. No sample data is prefilled.
Observation vocabulary
Four different claims. Keep them separate.
A single response can mention without citing, cite without recommending, or recommend without exposing why it retrieved a source.
01
Retrieval
A system using material in generation. Hidden retrieval cannot be proven from the response alone, so this workspace does not offer a retrieval-status field.
02
Mention
The named entity appears in the generated body. A mention is not automatically positive, cited, or recommended.
03
Citation
The interface visibly attaches a source card, link, or attribution. Record the displayed source; do not infer unseen sources.
04
Recommendation
The response explicitly endorses, selects, or ranks an entity for the user's need. Neutral inclusion is not a recommendation.
Evidence protocol
One observation, one visible response, one row
Freeze the conditions you can control, record only what is shown, and preserve ambiguity instead of resolving it with assumptions.
- 01
Define the prompt set
Choose a bounded, decision-relevant set before running it. Keep labels and exact text stable.
- 02
Hold conditions
Record engine/interface, locale, date, run number, and material account, location, or session conditions in notes.
- 03
Observe separately
Code body mention, visible citation, source, and recommendation as distinct fields. Use unclear or not checked when warranted.
- 04
Repeat without averaging away variance
Run repeated observations under documented conditions. Keep each run as its own row.
- 05
Link changes to retests
Name the intervention, keep the original observation, and connect a later run without claiming the change caused the result.
Method limitations
What this workspace cannot prove
Use the rows as an audit trail, not as a market benchmark or causal model.
- AI answers can vary by time, model, interface, account, location, session, and personalization.
- A visible citation does not reveal every retrieved source or prove that a page caused the answer.
- Manual coding introduces interpretation and transcription error; ambiguous states should remain marked unclear.
- A bounded prompt set describes only its own coverage. It does not represent all demand or all users.
- Local browser storage is not a backup, shared database, monitoring service, or Tenten-held dataset.
- Before-and-after rows show sequence, not uplift or causality; other changes may have occurred between runs.
Reproducibility checklist
Make a retest interpretable
Record enough context for another operator to repeat the run without pretending every engine condition can be frozen.
- Preserve the exact prompt or a version-controlled prompt ID.
- Name engine, interface, and visible model/version when available.
- Record locale, market/location, date, run number, and relevant login or session state.
- Keep screenshots or source evidence in your approved system; link the reference rather than uploading it here.
- Use the same coding definitions and retain unclear/not-checked states.
- Document the intervention before retesting and keep both original and follow-up rows.
- Have a second reviewer check high-stakes or ambiguous observations.
Change log
Protocol history
v1.0.0 · 2026-08-12 · Initial empty, browser-only protocol and CSV workspace. No first-party result dataset is published.
FAQ
What this lab does, and does not do
Does Tenten receive my prompts or URLs?
No. Rows stay in this browser's local storage and CSV files you choose to create. Interaction analytics use aggregate row, engine, locale, and completeness-band counts only.
Is this a Tenten benchmark or research dataset?
No. The workspace opens empty, contains no invented sample rows, and does not claim Tenten has run these observations. It is a protocol for creating your own evidence trail.
Why is there no visibility score?
A single score would hide prompt selection, engine variance, mentions, citations, and recommendations behind false comparability. The lab reports coverage and documentation completeness only.
Can these rows prove an intervention worked?
No. Linked before-and-after observations can support investigation, but sequence alone does not establish uplift or causality.
Turn a clean evidence trail into an operating decision.
Use the blank templates for adjacent work, the manual for cadence and ownership, or the URL Snapshot for a separate page-level technical review.
Local evidence remains yours · URL Snapshot is a separate workflow · No performance guarantee