Running scans and reading the evidence

A scan asks every approved prompt of every provider you enabled, once each, and stores the full answer exactly as it was returned.

Providers

Citerva can query OpenAI, Anthropic, Google Gemini and Perplexity. Only providers with credentials configured can be selected — one without a key cannot appear in a scan, and is therefore never reported as checked.

They differ in one way that affects your numbers:

Providers that cannot return sources sit outside the citation-rate denominator. They are not counted as having cited you zero times, because they were never in a position to cite anyone.

Why Gemini has a condition

Assistants can be reached two ways: directly, with that vendor's own key, or through a single gateway account that covers all four. Gemini returns its grounding sources on a direct key and returns none through the gateway.

So the same assistant, asked the same question, is either inside or outside your citation rate depending on how this deployment is configured — and Citerva follows the configuration rather than the label. If Gemini is reached through the gateway, its runs are excluded from the citation denominator instead of being recorded as having cited nobody. Perplexity returns sources either way.

This documentation said "yes, when grounding is enabled" until 2026-09-05, which was true of the assistant and not of the product.

Partial scans are normal

If Perplexity answers and Gemini is rate-limited, the scan finishes as partial. That is a first-class outcome, not an error: some of it worked.

Scan health is shown above every figure, listing which provider failed, how many runs it affected and why. Those runs are excluded from every rate rather than counted as a missing mention, because there is no way to distinguish "not recommended" from "not asked".

Reading the evidence

Every answer is stored verbatim and shown against the prompt that produced it, with:

• whether your brand was named, and at what position in the answer
• which competitors were named
• every source cited, marked as declared (the provider said it used this) or found in text (we found a link in the prose)

A declared source is much stronger evidence than a URL typed into a paragraph, so the two are counted separately and never summed without the split available.

If a number ever looks wrong, open the prompt and read the answer it came from. That is the point of storing them.

Raw answers are never overwritten

Mention detection improves over time. Because the original answers are kept, an improved parser re-derives historical results from the same observations instead of history quietly changing meaning underneath you. Every derived row records the parser version that produced it.

Scheduling

Scheduling is where most of the value is. A single scan tells you where you stand today with meaningful uncertainty; twelve weekly scans tell you which direction you are moving, which is the question worth acting on.

Trends are only drawn between scans that asked the same prompt set version of the same providers. Change either and the chart marks a break rather than drawing a misleading slope.

Cadences, what a schedule costs and how to set one up are on their own page: Scheduling scans. The cadence table lives there and only there, so the two cannot drift apart.