AI search intelligence for agencies and consultantsSee how we measure
Product help

Understand collection status before retrying

Read queued, submitting, collecting and terminal states. Keep failed coverage, saved evidence and point settlement separate when deciding what to do next.

Six platforms have passed local answer-and-source collection tests: ChatGPT, Gemini, Google AI Mode and Google AI Overviews for website answers; Perplexity and Claude for model API replies. Production rollout is pending. Region and search activity must be confirmed per sample before including it in comparable visibility metrics. View platform readiness.

LLM Citely4 min readReviewed Read as Markdown
Inspect the original run before retrying. Application status and evidence quality are separate checks.
Application status and evidence quality are separate checks. A failed run is unresolved coverage, not a valid answer without a brand mention. Original explanatory diagram; illustrative counts and states are not customer measurements.Open the figure for a closer look.

Read the run you already started

Find the original run and its latest status before starting another. A collection is an operation with a saved question context, observed surface and lifecycle; it is not just a button that immediately produces a score.

StatusWhat it describesUseful next check
queuedWork is waiting to be submittedCurrent platform availability and allowance
submittingA provider submission is in progressWait for the existing operation to settle
collectingProvider work or result collection is in progressLatest update and returned evidence
succeededA result was returnedReadability, surface and citation completeness
failedThe attempt ended without a usable successful collectionRecorded failure reason and usage settlement
timed_outA submitted collection exceeded its allowed lifecycleOriginal run, later evidence and settlement

Provider-specific terms can differ from the saved application status. Preserve the recorded status rather than guessing from a vendor dashboard.

A returned answer can still have limited evidence

A succeeded collection is the start of review. Confirm whether the answer is readable and whether the expected surface and search context were established. Citation evidence can also be incomplete even when text is present.

This matters because a readable literal brand mention and a confirmed search-metric observation are different evidence. Do not fill an unavailable metric with zero or treat a missing link list as a confirmed absence of citations.

The comparison workflow shows how answer counts and stricter metric eligibility remain distinct. The measurement walkthrough defines the denominators.

Read the failure before choosing a response

Check whether the problem concerns configuration, provider availability, timeout or returned evidence. Repeated attempts cannot fix an unsupported surface or a missing quote. Review platform readiness before assuming another request will work.

Keep the run identifier, question version, surface and update time. If a provider failure is recorded, do not reinterpret it as a negative answer about the brand. If the original operation is still active, preserve it and avoid overlapping duplicate requests.

This public guide does not access private run details. The walkthrough image uses illustrative lifecycle states.

Keep points and result status side by side

Open the usage entry associated with the operation. A reservation, final deduction and returned points describe different stages. The supplier can also report its own expense; that internal liability is separate from the customer point balance.

Use points and usage to read these fields together. A retry is a deliberate new attempt with its own identifier and possibly a new quote. Do not assume that a failed provider response has already settled every related accounting entry.

If a result arrives late, preserve its provenance and the original run status. Avoid adding the same answer twice to a reporting cohort.

Keep the failed attempt in coverage

When a retry is appropriate, preserve the exact question, market, observation surface and brand rules. If any of these changes, label the new attempt accordingly. Connect it to the original investigation rather than rewriting the old outcome.

Illustrative report: 12 planned runs, 10 valid answers and two failed attempts have 83.3% coverage. If three valid answers mention the brand, the mention rate is 30%, not 25%. A later retry can change the coverage cohort; explain how it is counted.

Use the client report template to keep missing coverage visible, and the action plan for unresolved collection work.

Questions, answered.

Is a failed run an answer without a brand mention?

No. It belongs in unresolved coverage, not the valid-answer denominator.

Should I immediately rerun a collecting request?

First inspect the original operation. Overlapping attempts create separate work and make the evidence harder to interpret.

Sources and further reading

Examples marked illustrative are not customer measurements.

Put this into practice

Start with your website

Add your brand and customer questions, then choose a platform for your first check.

Read the related point settlement

Your questions are saved before you start. Review availability and point costs before confirming a check.