Sentiment monitoring

Sentiment monitoring runs a tracked prompt set against AI platforms — ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Perplexity, and Grok — and classifies each response that mentions your brand on a sentiment scale (commonly −100 to +100, or negative / neutral / positive buckets).

Three signals are extracted from each response:

• Polarity: is the model leaning positive, neutral, or negative when it talks about you?

• Stated reasons: what specifically is the model saying — "fast onboarding", "expensive", "buggy on mobile", "trusted by Fortune 500"? These reasons come from the corpus the model was trained on or retrieved from at answer time.

• Cited sources: when sentiment is grounded in real-time retrieval (Perplexity, Bing Chat, Google AI Overviews), the citations expose which page on the public web shaped the negative or positive framing. This is the most actionable signal in the entire workflow.

A mature sentiment monitor aggregates polarity into a single brand sentiment score, breaks it down per platform, and surfaces the cited URLs driving any negative skew so the brand can act on the source — not just the symptom.

Mention frequency tells you how often AI assistants talk about your brand. Sentiment monitoring tells you how they talk about it.

A brand can have high AI share of voice and damaging sentiment — a category leader being cited as "the expensive option" or "the one with reliability issues" loses pipeline despite high visibility. Conversely, a smaller brand with low mention frequency but uniformly positive AI sentiment can outperform on conversion because every AI mention is a recommendation.

The two metrics are complementary. Mention frequency is the reach metric; sentiment monitoring is the quality metric. Reporting one without the other paints an incomplete picture.

AI assistants increasingly act as a pre-purchase research layer. A buyer evaluating CRM tools may ask Claude to summarize the strengths and weaknesses of three vendors before ever visiting their websites. The way the model frames each vendor heavily influences which one ends up on the shortlist.

Three operational reasons sentiment monitoring matters:

1. Catch crises early. A viral negative review on Reddit, a G2 1-star, or a critical Hacker News thread can flip AI sentiment within days as platforms re-crawl. Monitoring surfaces the swing before it metastasizes.

2. Trace negative framing to its source. Cited sources in Perplexity, Gemini, and AI Overviews reveal exactly which URL shaped a negative answer. This makes sentiment a fixable problem (publish a rebuttal, request a correction, ship a product fix) rather than a vague brand-health complaint.

3. Benchmark against competitors. A brand with +30 sentiment while a direct competitor sits at +65 is losing the AI narrative regardless of who has more mentions. The competitive delta is the leading indicator.

Five steps for credible AI sentiment monitoring:

1. Define a tracked prompt set tied to commercial intent. Mix category prompts ("best Y for Z"), comparison prompts ("[brand A] vs [brand B]"), and open-ended brand prompts ("what do people say about [brand]"). 100–300 prompts is the practical baseline.

2. Run prompts on every relevant platform. Sentiment diverges sharply across ChatGPT, Claude, Gemini, Perplexity, and Grok because each model is trained on different data and retrieves differently at inference time. Per-platform tracking is non-negotiable.

3. Score on a continuous scale, not just buckets. A −100 to +100 scale captures intensity. "Slightly negative" and "scathing" should not produce the same number. Indexly uses embeddings plus an LLM grader against a defined rubric.

4. Always pair sentiment with the cited source. When a response carries citations, capture them. The cited URL is the lever — without it, sentiment alerts are observational, not actionable.

5. Track per-platform deltas weekly. Single-day snapshots are noisy. 7-day moving averages with a 30-day baseline catch genuine shifts without false alarms. Trigger crisis alerts only on multi-platform negative shifts of 10+ points.