Monitoring

CiteRank supports 8 AI engines. Answers are captured per engine, because engines disagree: the same prompt frequently produces different recommendations and different citations depending on the model and its retrieval layer.

• Engine coverage is identical on every paid plan; plans differ by prompt volume, replay depth, competitor count, reporting and services.
• Engine-level breakdowns are shown for every metric, so you can see where visibility is concentrated.
• Engines that change behaviour mid-period are flagged on the run record rather than silently averaged.

A monitoring run is one execution of a prompt set across the engines at a point in time. Runs are the unit of comparison — trends are always run-over-run, never averaged across arbitrary windows.

• Each run stores the prompt text, engine, timestamp, raw answer and extracted entities.
• Run cadence depends on plan: weekly or monthly on standard plans, more frequent on managed programmes.
• Ad-hoc runs can be triggered after a content or PR change to check for movement.

Generative answers are non-deterministic: asking the same question twice can produce different brands. A replay is a repeat execution of the same prompt on the same engine within a run, used to establish how stable a result is.

• Typical depth is 10–30 replays per prompt per engine, depending on plan.
• A brand named in 9 of 10 replays is treated very differently from one named in 2 of 10.
• Replay counts are shown alongside every metric so a number can be traced to its sample.

Confidence expresses how much weight a result deserves given its replay sample and variance. It is a measurement quality signal, not a performance score.

• High confidence — consistent across replays and engines; safe to act on.
• Medium confidence — directionally useful; verify before making large commitments.
• Low confidence — small sample or high variance; treat as observation only.

Engines rate-limit, time out and occasionally refuse. CiteRank records these outcomes rather than discarding them, because a silent drop would bias the resulting score.

• Failed prompt executions are retried within the run window.
• Prompts that still fail are excluded from scoring and shown as excluded on the run record.
• A run with substantial exclusions is labelled partial, and comparisons flag the gap.
• Refusals — where an engine declines to answer — are stored as a distinct outcome, not as a zero.