What should you request when AI visibility shows no progress?
Request the attempt history, the selection rule agreed before collection and the current judgment for each reported cell. These records can confirm a genuine lack of progress or reveal changes hidden by an aggregate. If selection and judgment are sound and comparable observations show no improvement, report that result as unchanged and apply the agreed continuation or stopping conditions. Ask the agency to identify which condition governs the next decision.
A planned observation receives a cell_id; each request receives an attempt_id; each assessment of an answer receives a judgment_id. A scheduled repeat has its own cell. A recovery request belongs to the cell it attempts to complete.
Record the cutoff and eligibility conditions before collection. A rule such as “first eligible answer before the cutoff” requires checking attempts in time order. Eligibility cannot depend on whether the target brand appears. A different selection policy needs an explicit version and prospective use, rather than a change made after seeing the answers.
Which records should Doubao, Qwen and DeepSeek return separately?
For Doubao, Qwen and DeepSeek, request a separate cell list, attempt history and judgment history for each engine. Each list should identify the rule version, selected attempt and exclusion reasons. Verify the selection against that rule before reconciling the engine's current states; a correction in one engine must not silently alter another engine's result.
Include the question-bank version, question, route and planned repeat in the cell identity. Record the actual model, search conditions and collection time alongside the measurement environment record. An out-of-scope attempt remains in history without automatically entering the comparable set.
Seven states make the selected record reproducible
Evaluate the following states in order. A report reads one current state per cell while retaining the underlying attempts and judgments.
| State | Entry condition |
|---|---|
NOT_RUN | No collection attempt |
COLLECTION_ERROR | Attempts exist, but none produced a complete readable answer |
INELIGIBLE | Readable answers exist, but none meet the agreed conditions |
JUDGMENT_PENDING | An eligible answer is selected and awaits assessment |
JUDGMENT_ERROR | Assessment of the selected answer failed |
MENTION_YES | Completed assessment confirms the target brand appears |
MENTION_NO | Completed assessment confirms the target brand does not appear |
The denominator guide covers the underlying distinction between an unavailable observation and a valid negative result. At export, preserve missing-field and unknown-value distinctions too. JSON Schema distinguishes a property containing null from an absent property. JSON Schema: Objects
The ledger can require the result field on every judgment, use a null value with an unresolved state, and reserve false for a completed negative assessment. Keep the state next to blank CSV values.
Reconciliation rules for the report version
Let P represent planned cells for one engine. The seven state counts sum to P; attempt counts remain outside that equation. If Y and N count cells whose selected answer has a current completed positive or negative judgment, J = Y + N.
Use fixed labels: judgment completion for J/P, mention rate among judged observations for Y/J, and share of planned cells with a confirmed mention for Y/P, if retained. A ratio with a zero denominator is not applicable. These are definitions, not measured performance figures.
Corrections should reveal what changed
Create a new judgment when a reviewer changes an assessment of the same answer. Connect it to the previous judgment_id, with the old value, replacement, supporting passage, reason, reviewer and time. If the selected answer also changes, record the attempt-selection change separately.
Attach the original report version, revised version and affected cells to the correction. Compare selected attempts and judgments cell by cell, then recalculate states and measures. The agency report audit guide provides context for inspecting the evidence returned with a report.
An answer collected after the cutoff belongs to a later observation or an explicit correction version. Check its conditions before using it in that correction, disclose both the original cutoff and actual collection time. A reviewer should be able to reproduce either version.
Want to see how AI engines describe your brand today?
Get a free growth audit