Docs / Intelligence

Intelligence

After every scan, Sightlabs runs the full mention set through an AI pipeline. This page explains what each output means and how to read it.

Note

All analysis is generated fresh on every scan using state-of-the-art LLMs. Results are never recycled from a previous run, even for the same query.

Sentiment scoring

Every mention receives one of three sentiment labels:

Positive

The mention expresses approval, enthusiasm, recommendation, or satisfaction toward the query.

Neutral

The mention references the query without a clear positive or negative stance. Informational posts, questions, and news items typically land here.

Negative

The mention expresses criticism, frustration, complaint, or warning about the query.

The sentiment breakdown shown at the top of each scan result is the percentage split across all mentions. Individual mention sentiment tags appear in the Top Mentions panel.

Warning

Sentiment scoring is probabilistic. Sarcasm, irony, and mixed-sentiment posts are difficult to classify reliably. Treat the aggregate percentages as a directional signal, not a precise measurement.

Action items

Sightlabs reads through all mentions and surfaces the ones that suggest something you should do. Each action item has a priority level and a recommended next step.

Reply opportunityHigh

A question, comparison request, or direct mention that would benefit from a response. Typically a support question or feature inquiry in a public thread.

PR riskMedium

A negative thread gaining traction, a complaint that has attracted replies, or a viral criticism. The recommendation includes a suggested response approach.

Engagement spikeLow

A positive post performing well that you could amplify with a reply, repost, or quote. Low urgency but high upside.

Themes

The AI reads all mentions and clusters them into recurring topics. A typical scan surfaces 3-8 themes. Common examples include: pricing, onboarding, support quality, API reliability, and competitor mentions.

Themes are not a fixed taxonomy. They are generated fresh from the actual mention content, so they reflect what people are genuinely discussing right now. A theme like "pricing confusion" appearing for three scans in a row is a stronger signal than one appearance.

About LLM-backed analysis

All intelligence features use large language models. This means:

Results are probabilistic, not deterministic. Two identical scans may produce slightly different action items or theme labels.
Analysis quality depends on mention quality. If the source data is low-signal (short posts, lots of abbreviations), the output reflects that.
The model does not have memory across scans. Each scan is analyzed independently.

Use the intelligence output as a decision-support layer, not a replacement for reading the actual mentions. The Top Mentions panel always shows you the raw source text.