Glossary · Beginner
What Is a Sentiment Score? Definition and Example
A sentiment score measures how positively, neutrally or negatively AI answers describe a brand. In GEO, it's usually calculated by classifying each brand mention in a set of AI answers, then reporting the share of each tone. It shows how AI talks about you, which mention counts alone can't.
On this page
- How is it calculated?
- Why does it matter?
- What are the limits?
- How do you improve it?
- From MentionX data: why sentiment needs mention context
- Worked example: Ledgerline reads its negatives
- How reliable is automatic classification?
- Common questions
- Related
Checked on 31 October 2026. "Our view" marks MentionX opinion. Ledgerline is a fictional company used for illustration.
How is it calculated?
A common approach:
- Run a prompt set across AI engines.
- Find each passage that mentions the brand.
- Classify it as positive, neutral or negative, by a model, a person or both.
- Report the shares, or a net score (positive minus negative).
| Answer excerpt (illustrative) | Classification |
|---|---|
| "Known for fast onboarding and friendly support" | Positive |
| "Offers invoicing and expense tracking" | Neutral |
| "Some users report billing problems" | Negative |
Illustrative examples, not real data.
Why does it matter?
AI answers often summarise reviews, forums and press coverage. If those sources mention recurring complaints, the answer may repeat them to every buyer who asks. Tracking sentiment helps you spot these narratives early and trace them back to the sources that feed them.
What are the limits?
- Classifier errors: mixed or sarcastic passages are often misread. Check a sample by hand.
- Neutral dominates: many answers are simply descriptive, so small changes in positive or negative share can be noise.
- Sentiment isn't choice: a positive description followed by "but for most teams, choose X" is still a lost recommendation. Pair sentiment with recommendation rate.
How do you improve it?
You can't edit an AI answer directly. Fix the underlying causes: address real product issues, respond to reviews, correct outdated information on your own pages and earn coverage on sources AI systems cite. Then re-measure with the same prompt set.
From MentionX data: why sentiment needs mention context
Method. Read-only aggregate queries over MentionX scans from 22 August to 29 October 2026 (UTC): 2,217 engine runs (2,154 completed) across 18 anonymised projects and six engines.
A sentiment score is calculated only on answers that mention the brand, so it depends heavily on which prompts you track:
| Intent | Brand mentioned | Share of voice |
|---|---|---|
| Brand-specific | 100% | 64.3 |
| Comparison | 91.9% | 44.6 |
| Recommendation | 29.2% | 26.8 |
| Best-of | 26.6% | 22.7 |
| Problem | 5.7% | 7.7 |
Brand-specific prompts produce mentions every time and often include pros and cons, so they're where negative statements show up most. Problem prompts rarely mention a brand at all, so they contribute little to a sentiment score even though they shape early buying decisions.
Our view: always report sentiment alongside mention rate and the number of mentions it's based on. A 90% positive score from 12 mentions says much less than it seems.
Limits. A convenience sample of mostly B2B software projects, run through engine APIs, which can behave differently from the consumer apps. Perplexity, Copilot and DeepSeek figures come from a single day (26 October 2026), so treat them as a snapshot, not a trend.
Worked example: Ledgerline reads its negatives
Ledgerline is a fictional accounts-payable tool. Figures are illustrative.
Ledgerline's net sentiment looks healthy, but the team reads every passage classed as negative. Most repeat one claim: that Ledgerline "lacks multi-currency support". That was true two years ago. The cited sources are an old review and a forum thread. The team updates its documentation and pricing page, asks the review site to refresh its listing, and tracks whether the claim fades over the next two months.
How reliable is automatic classification?
Model-based classification is fast but makes mistakes, especially with mixed statements ("powerful but expensive") and with neutral facts that sound negative. Check a sample by hand each period, and keep the same classifier and instructions over time so scores stay comparable.
Common questions
Is neutral sentiment bad? No. Most factual descriptions are neutral, and accurate neutral answers are often what you want.
Should I compare my score with competitors'? Yes, if you use the same prompts, engines and method for all of them.
How often should sentiment be measured? Monthly, with your other visibility metrics. It changes slowly unless something new is published about you.
Related
Questions
How is sentiment scored?
Usually by classifying each answer, or each mention within it, as positive, neutral or negative, then summarising the share of each.
Is a high sentiment score enough?
No. An answer can describe you warmly and still recommend a competitor, so track sentiment alongside recommendation rate.
Can sentiment classification be wrong?
Yes. Automated classifiers miss sarcasm, mixed opinions and context, so spot-check a sample by hand.