AI brand sentiment: read the evidence behind a positive score

A positive AI sentiment score tells you what share of the classified brand-mention records that name your brand carry a positive label. It does not tell you whether an answer recommended you, what it praised, or what it warned about. To read the evidence behind the score, look at four things separately for each saved answer: the tone (positive, neutral or negative), the recommendation role (top recommendation, alternative option, warning or caveat, or not recommended), the strengths the answer stated, and the weaknesses it stated. An answer can be positive in tone and still caution the buyer. Everything you read is a model's classification of an answer, so it is evidence to inspect, not proof.

In short

• Sentiment is a label on a stored mention record. It is not a count of answers, customers or recommendations.
• Tone, recommendation role, and stated reasons are three different signals. They can disagree, and the disagreement is where the useful reading is.
• "Positive but cautious" answers are the ones a headline percentage hides. Find them by combining filters, then read the quotes.
• The labels are a model's reading of the answer. The quotes let you check that reading; nothing here shows why an engine produced the answer.
• The rubric below turns a review of any answer into a repeatable, five-minute task.

What does an AI sentiment score actually measure?

It measures the share of classified brand-mention records that carry a positive label. In DiscoveredBy, positive share is positive records divided by classified records (positive, neutral and negative) times 100, and unclassified records are shown separately and left out of that denominator (Sentiment).

Three details change how you should read it:

• The unit is a stored mention record, not an answer, a citation or a person. Repeating a name inside one record does not add weight. If an answer produced two stored records that resolve to the same brand, both count, and the screen does not merge them.
• Unclassified is not neutral. It means the saved record has no sentiment label. Treating it as neutral would flatter or punish you by accident.
• A dash and a 0% are different. A dash means there is no classified evidence at all. A measured 0% means there are classified records and none was positive.

The score also does not say the share of answers that recommend you. It describes sentiment among classified mentions only. If you want the definitions of mentions versus citations first, read AI visibility, brand mentions, and citations and the key terms.

Sentiment here is one signal for one channel at a time. A channel is one provider, platform, surface and collection method, and the Sentiment screen keeps channels separate, so a ChatGPT (app) result and a Gemini (API) result are never averaged together.

Why can a positive answer still hurt you?

Because tone and recommendation role are stored independently. The role is a separate saved label on the same mention record, such as Top recommendation, Alternative option, Warning or caveat, Competitor preferred or Not recommended, and the docs state plainly that a role is independent of sentiment (Sentiment).

That means an answer can describe your product warmly ("a polished, well-reviewed tool") and still slot you into a role of Alternative option, or attach a Warning or caveat ("worth it if your team is small"). Both records would read as positive in the score.

The DiscoveredBy Reputation risk filter reflects this. Its "Negative or warned against" option matches negative sentiment or a saved Warning or caveat or Not recommended role, and a record matching both counts once. The "Warned against" option includes records whose sentiment is positive, neutral or unclassified. Competitor preferred alone does not qualify as risk (Sentiment).

The four signals to keep apart

Each signal answers a different question, comes from a different place in the product, and has its own limit.

Brand reasons come from the same single model call that already lists the brands an answer names, with their sentiment and role. That call also returns up to three reasons per brand. Each has a label (one of thirteen, including pricing, features, support, security and Other reason), a direction (strength or weakness) and a quote. Before a reason is saved, its quote is looked up in the saved answer text, and a reason whose quote cannot be found is discarded (Brand reasons).

Two details matter when you interpret reasons. Strengths and weaknesses count answers, not quotes, and one answer can count as both a strength and a weakness for the same label, such as "cheapest option, but watch the add-on fees". And answers analysed before the reasons feature shipped were not re-analysed, so a long window may show fewer analysed answers than answers naming the brand.

Enlarge image

How do you find the positive-but-cautious answers?

Combine filters in the Mentions evidence list, then open the saved answer. On the Sentiment screen, the Mentions tab lists stored excerpts with sentiment, recommendation role, date and output ID, and the sentiment and role chips apply as soon as you pick a value (Sentiment).

A workable sequence:

1. Choose one window, one channel and, if relevant, one prompt segment. Keep them fixed for the whole review.
2. In Mentions, set the brand to your own brand and the sentiment chip to Positive.
3. Set the role chip to Warning or caveat. Then repeat with Alternative option, and again with Not recommended.
4. Select "View prompt and outputs" on each row and use the output ID to open the saved answer. The excerpt is capped at 500 characters, so read the full answer.
5. Open Brand reasons and select the count for a weakness label to see quotes filtered to your brand family and that label.

Filtering the evidence list changes only that list. The comparison, trend and role counts do not move, which is useful: you can inspect the odd cases without disturbing the headline.

What should you read in the quote?

Read for three things: what the answer attributes the view to, whether the caveat is conditional, and whether the claim is checkable against your own facts. A quote is copied from the saved answer, so it shows the wording of the collected answer rather than a paraphrase.

Look at whether the caveat depends on a buyer condition ("for small teams", "if you need on-premise") because a conditional caveat may point to something on a page or in your positioning that you can test. Look at whether a strength is generic ("easy to use") or specific ("supports redlining in Word"). And look at whether any statement is factually wrong; if it is, that belongs in a fact check, not a sentiment discussion.

When the same weakness label keeps recurring on tracked prompts, remember the limit: the weaknesses on the reasons screen are only those the answers happened to volunteer. Two related posts: What AI says your product is bad at: turn objections into a review backlog and Is your brand known for the attributes you want to own?.

Answer-review rubric (copy this)

Use one row per saved answer. Fill it in from the answer itself, using the platform labels as a starting point and your own reading as the check.

ANSWER-REVIEW RUBRIC
Date reviewed:            Reviewer:
Channel:                  Window:           Segment/persona/language:
Prompt:                   Output ID:

1. TONE (platform label): positive / neutral / negative / unclassified
   Your reading of the quote: agree / disagree / unclear
2. ROLE (platform label): top recommendation / alternative option /
   warning or caveat / competitor preferred / not recommended /
   unclassified
   Does the wording of the answer match the role? yes / no
3. STATED STRENGTHS (label + quote, up to 3):
   a.
   b.
4. STATED WEAKNESSES (label + quote, up to 3):
   a.
   b.
5. Is any caveat conditional on a buyer situation? yes / no
   If yes, which situation:
6. Is any statement about us factually wrong? yes / no
   If yes: route to fact check, do not score as sentiment
7. Same pattern in a competitor's record in this channel? yes / no
8. Verdict: leave as is / watch / investigate our pages / investigate
   third-party sources
9. Hypothesis to test (not a conclusion):

To keep reviews comparable, use the same rubric on the same channel and window each time, and log the verdict so that you can compare it next month.

How to read the combinations

Worked example: Quillstone

Illustrative example: Quillstone and its competitors are fictional, and the numbers are made up to show the method.

Quillstone sells document-review software to mid-sized legal and compliance teams. Its competitors are Brieflane and Clausewise. On one channel over a 28-day window, Quillstone has 26 stored mention records: 14 positive, 4 neutral, 2 negative and 6 unclassified.

• Classified records: 14 + 4 + 2 = 20.
• Positive share: 14 ÷ 20 = 70%.
• Because 20 is under 30 classified records, the figure is marked provisional. The 6 unclassified records are shown separately and are not counted as neutral.

That 70% looks like good news. Now the team reads the 14 positive records by role:

The three Warning or caveat records are positive in tone but cautious. The team opens them with the rubric. Two quotes read, in effect, "polished and well reviewed, though pricing is not published, so budget teams should request a quote." The third says the tool suits mid-sized teams but is "less proven for very large firms". The team's rubric entries:

• Tone: positive, agree with the label.
• Role: Warning or caveat, matches the wording.
• Stated strength: features, ease of use.
• Stated weakness: pricing (2 records), market presence (1 record).
• Conditional caveat: yes ("budget teams", "very large firms").
• Factually wrong: no.
• Same pattern for Brieflane: no; its records cite integrations as a weakness instead.

Reading: the headline (70% positive) is unchanged, but 3 of the 14 positive records, 21%, carry a caution, and two of them point at the same topic. The team's hypothesis is that an unpublished pricing page may be contributing to the pricing caveat. That is a hypothesis to test, for example by clarifying pricing information on Quillstone's own pages and re-reading the same channel later. It is not a finding that the pricing page caused it, because the reasons are what the answers stated, not a measured cause.

Enlarge image

Common mistakes and what this cannot tell you

• Reading tone as recommendation. A positive label is not a top-pick role. Check the role.
• Treating unclassified as neutral. It is a missing label. Fix the interpretation, not the number.
• Trusting a small sample. Fewer than 30 classified records is provisional, a sample-size flag rather than a confidence interval.
• Mixing channels. App and API answers can differ, so review one channel at a time.
• Reading the segment as fixed. The prompt segment uses each prompt's current classifications, even for historical outputs, so reclassifying a prompt changes the historical segment.
• Treating stated reasons as causes. Reasons are what an answer said. They do not show why the engine produced the answer or ranked a brand.
• Reading an empty list as safety. Missing labels and incomplete extraction can produce empty lists.
• Turning a review into a score. The view creates no risk score, case status, assignment or notification; your log has to.

What this method cannot do: verify that a claim is true, explain the model's internal choices, or predict how buyers will react. For factual errors, use a fact-check process. For persona differences, the Sentiment screen offers a by-persona view, and the comparison against one competitor restricts the analysis to answers naming both brands.

Frequently asked questions

Is a high AI sentiment score good?

It is a good sign only if the role and reasons agree with it. A high positive share with many Alternative option or Warning or caveat roles means answers speak well of you while placing you second or cautioning buyers. Check the roles and read the quotes before you report it as good news.

Why do I see records marked unclassified?

Unclassified means the saved mention record has no sentiment label. It is excluded from the positive share denominator and shown separately. A failed output, or one whose brand extraction has not finished, contributes no record at all.

Can I trust the sentiment label on each mention?

Treat it as a model's classification of the answer, not as ground truth. The excerpt lets you check it against the wording. If the label and your reading disagree often on a channel, note that in your review log rather than adjusting the number.

What is the difference between sentiment and brand reasons?

Sentiment says how an answer frames a brand: positive or negative, top recommendation or warned against. Brand reasons say what the answer itself gives as strengths and weaknesses, with a quote for each, such as pricing, features or support (Brand reasons).

Do sub-brands count in the family's sentiment?

Yes. A brand family's row already includes every sub-brand's mention records, and each sub-brand's own row overlaps the family row rather than adding to it. Read brands and sub-brands before comparing rows.

How often should I run this review?

Pick a cadence you can keep, and hold the window, channel and segment constant so reviews are comparable. Review sooner after a change to your pricing or product pages, and note the date so later comparisons are fair.

Next step: review your own positive answers

Open the Sentiment screen, pick one channel, and filter your own brand to Positive sentiment and the Warning or caveat role. Read those answers with the rubric above; then open Brand reasons for the weakness quotes. Brand perception covers both screens, and Brand reasons depends on your plan's entitlements, so check plans and limits if it is not visible to you. If you have not yet started tracking, sign in or sign up and let answers accumulate first. For a related investigation, see why AI recommends your competitor.