Sentiment Analysis

Overview

Every time an AI model responds to one of your prompts, AEO Optima automatically scores the sentiment of that response. Sentiment analysis tells you not just whether AI models mention your brand, but how they talk about it — positively, neutrally, or negatively.

A brand that appears in 80% of AI responses but is described negatively has a very different challenge than a brand that appears in 40% of responses but is described positively. Sentiment analysis gives you this critical second dimension of visibility intelligence.

How Sentiment Scoring Works

When a snapshot is captured, the AI response is scored by a dedicated language model acting as a judge (the industry-standard "LLM-as-a-judge" method). Rather than counting positive and negative words, the judge reads the whole response the way a human analyst would and assesses the stance taken toward your specific brand — not the overall tone of the text, and not how it talks about any competitor mentioned alongside you. This is a form of targeted aspect-based sentiment analysis: sentiment is always measured relative to one named entity, your brand.

The judge produces both a label and a numeric sentiment score from −100 to +100, anchored to an explicit scale:

• Positive (score above +10) — The response recommends your brand, highlights genuine strengths, or positions you as a leading option. Strong endorsements score toward +100; brief favorable mentions and faint praise score lower.
• Neutral (score −10 to +10) — The response mentions your brand factually, or balances genuine praise with genuine criticism.
• Negative (score below −10) — The response highlights weaknesses, or names your brand as the weaker option in a comparison — even when the surrounding text is full of positive-sounding words about other brands.

Why a judging model instead of a word list

Earlier sentiment scoring relied on a keyword approach: tally favorable and unfavorable words near your brand name. AI answers about brands are written in confident, flattering marketing prose, so a word-counter systematically over-rates them — it cannot tell the difference between "Brand X is the best choice" and "Brand X is capable, but trails the other two and isn't truly comparable for enterprise needs." Both sentences are full of positive words; only the second is actually unfavorable toward the brand. The judging model reads the comparison and scores the second one negative, the way a person would.

The judge is specifically instructed to get the cases a word list gets wrong:

• Comparative loss ("trails competitors", "less established", "plays in a different league") is scored negative, not positive.
• Faint praise (a brief, unenthusiastic mention) is scored only mildly favorable, not a strong positive.
• Praise of a competitor counts toward that competitor, never toward your brand.
• Mixed responses (real praise and real criticism) land near neutral, with both noted.
• If your brand isn't actually discussed in the response, it produces no sentiment signal at all — it is never counted as a fabricated "neutral."

Determinism and the judging model

The judge runs at zero temperature, so the same response always produces the same score — sentiment is reproducible, not random. The judging model is selected automatically from a reliability-proven, provider-agnostic allowlist — currently Qwen3-30B-A3B-Instruct as the primary judge, with Mistral Small 3 and Google Gemini 2.5 Flash as fallbacks — ranked cheapest-reliable-first. A model is admitted to the allowlist only after it passes a labeled probe battery, so the allowlist upgrades as newer capable models qualify. The judge is independent of the AI models you track: it is the consistent "analyst" reading every response regardless of which engine produced it.

Sentiment is assessed per snapshot, meaning the same prompt can receive different sentiment scores from different AI models or at different points in time.

Search-surface answers

From 6 September 2026, answers captured from Google AI Overviews and AI Mode are scored by the same judge and the same lexicon fallback as chat answers, and carry the same score, rationale and scoring method. Before that date a separate keyword word list scored them and wrote no numeric score — so a search answer counted toward the sentiment distribution but not toward any average, and an answer whose brand appeared only outside the AI Overview text was recorded as neutral rather than as having nothing said about the brand. Search-surface answers captured before 6 September 2026 were re-scored through the judge; the verdict each one previously held is retained alongside the new one.

Three Views of Sentiment

The Sentiment page provides three distinct views to help you understand the data from different angles.

Overview

A donut chart showing the overall distribution of positive, neutral, and negative sentiment across the answers in the selected date range that named your brand and were judged. An answer that nothing has judged is not counted as neutral (until 4 September 2026 it was). Beneath the donut, the same split is shown per question type over the whole tracked set — an answer about your brand and an answer where you are one name among many are different questions, and the four bars keep them apart; below four judged answers a bar shows the count and says the sample is too small.

Under each of those four bars, the same split is drawn over time, one point per capture cycle — a week for a weekly project, a day for a daily one, so the chart follows your own capture schedule rather than a calendar the data does not have. The denominator is unchanged: only answers that named your brand and carry a judged verdict are counted. A cycle needs at least four judged answers to be drawn, because a split taken from one or two answers swings by tens of points on a single re-run; cycles below that floor are gaps rather than zeroes, and each card says how many fell short. Where a question type has fewer than two reportable cycles there is no chart at all — one point is not a trend.

This view answers the question: "Overall, how is AI talking about my brand?"

A healthy distribution typically shows a majority of positive and neutral sentiment. A large negative segment warrants investigation.

By LLM Provider

A breakdown of sentiment per AI model. This view shows you which models are most favorable and which are least favorable toward your brand.

This view answers the question: "Which AI engine has the most positive (or negative) perception of my brand?"

It is common to see significant variation between models. For example, one model may describe your brand very positively while another is more neutral or critical. These differences often reflect variations in training data, recency of information, and model-specific behavior.

By Prompt Type

A breakdown of sentiment per prompt category or type. This view reveals how sentiment varies depending on the kind of question being asked.

This view answers the question: "What types of questions lead to the most positive or negative responses about my brand?"

Common patterns include:

• Brand awareness prompts tend to produce neutral or positive responses.
• Comparison prompts often produce more mixed or negative sentiment, because AI models attempt to present a balanced view that includes competitor strengths.
• Problem/solution prompts vary widely based on whether AI models associate your brand with the solution.

Taking Action on Sentiment Data

Understanding sentiment is only valuable if it drives action. The following table provides a framework for interpreting common sentiment patterns and responding to them.

Reading Sentiment Trends

Sentiment is not static. AI models are regularly retrained and updated, which means sentiment can shift over time. Key things to watch for:

Gradual Shifts

A slow drift from positive toward neutral may indicate that competitors are strengthening their own AI presence, causing AI models to present a more balanced view. This is a signal to reinvest in content and visibility efforts.

Sudden Changes

A sharp change in sentiment — especially toward negative — may indicate:

• A widely reported negative event or PR issue affecting your brand.
• A change in an AI model's training data or behavior.
• New competitor content that repositions the landscape.

When you spot a sudden change, use the snapshot detail view to read the actual AI responses and understand what triggered the shift.

Model-Specific Trends

Track sentiment per model over time. If one model's sentiment is improving while another's is declining, it helps you understand which information sources each model is drawing from and where to focus your efforts.

Tip: Pair sentiment analysis with your Competitor Intelligence data. If your sentiment declines while a competitor's mention share increases, the two trends may be connected — competitors may be producing content that AI models prefer.

Sentiment and Other Features