How tracking works

Tracking is the process of sending each active prompt to each enabled model channel and saving the resulting answer and evidence.

The collection matrix

The planned work for a run is:

active prompts × enabled model channels

Each successful answer is unique for one prompt, one engine, and one project-local date. Running tracking again on the same day skips successful combinations that already exist. Failed combinations can be retried.

This makes daily collection idempotent while keeping failures visible.

Model channel and model version

Selfwin stores two model identities when available:

• Model channel is the stable surface, such as ChatGPT UI, OpenAI API, Gemini UI, or Gemini API.
• Model ID is the particular underlying model version reported for that answer.

Model channels keep a trend continuous when a provider changes the model behind a consumer product. The model ID preserves the technical detail needed to explain a change.

Browser and API collection

Selfwin supports provider APIs and separate browser-based collection.

Provider API

API collection uses the project’s configured provider credentials. It is generally direct and structured, but the answer or sources can differ from what a logged-out consumer sees in a web product.

Browser collection

Browser collection uses a fresh session for supported model websites. It can preserve visible sources and observed background searches. Production scheduling requires configured cloud-browser infrastructure; local Chrome is a development fallback.

Do not merge browser and API channels when you need a like-for-like comparison.

What a completed answer contains

Depending on the model response, Selfwin stores:

• Full answer text and raw provider metadata.
• Mentioned tracked brands and newly observed brands.
• The number of textual occurrences for each brand.
• First-mention position, sentiment, and a supporting snippet.
• Retrieved URLs and domains.
• Citation counts and citation positions.
• Observed fanout queries.
• Response features, maps, ads, and products when present in the collector output.

Analysis happens after collection

After an answer is stored, Selfwin analyzes it against the project’s brand definitions. Literal tracked names and aliases are matched case-insensitively. Custom matching expressions are case-sensitive.

The result is recalculable. Changing a brand name, alias, domain, or matching expression can trigger historical analysis so old answers follow the current brand identity.

Scheduling

Projects can use daily or weekly tracking. Scheduled runs use the project’s time zone and selected collection method. Manual tracking is available from the application header.

The scheduled Visibility Agent and Slack brief are separate jobs from answer collection. They read stored tracking evidence; they do not replace it.

Coverage and comparability

Two visibility percentages are comparable only when their underlying populations are comparable. Check:

• Date range.
• Active prompts represented.
• Model channels.
• Countries.
• Topics and tags.
• Failed or newly added prompt-engine combinations.

The Overview shows response and prompt coverage so a thin data set is not mistaken for a complete market measurement.