Sentiment analysis in PR: Definition, formulas, and limits

In this article

Sentiment analysis in PR classifies the evaluative tone of analyzed content, usually as positive, neutral, or negative. It describes the content. It does not measure what an audience thinks or feels.

That distinction sounds simple, but many reports blur it. A positive article is not evidence of positive reputation. It is evidence that the article met the report's rule for positive content.

Overall media tone and brand sentiment are different

Before choosing a formula, decide what is being scored.

Overall media tone asks whether the item as a whole is positive, neutral, or negative. A report about an industry downturn may have a negative overall tone even if it describes one company favorably.

Brand sentiment asks how the item treats a named brand. The same article can be positive toward Brand A, negative toward Brand B, and neutral toward Brand C.

These are separate coding questions. A report should not switch between them without saying so.

Sentiment is an output metric

The AMEC Taxonomy of Evaluation places tone, sentiment, and favorability among communication outputs. They describe published or distributed content.

Audience attitude belongs at the outcome level and needs audience evidence. Surveys, interviews, experiments, or behavioral data can test whether people trust, prefer, or intend to choose a brand. Media sentiment alone cannot.

The same boundary applies to AI answers. A positive AI description is an observed content property, not proof that the user adopted that view.

Sentiment distribution

The clearest report shows the complete distribution.

Positive share = positive scored items ÷ all scored eligible items × 100

Neutral share = neutral scored items ÷ all scored eligible items × 100

Negative share = negative scored items ÷ all scored eligible items × 100

Suppose 100 eligible clippings were collected. The system scored 90 of them: 45 positive, 30 neutral, and 15 negative.

The distribution among scored items is:

• positive: 45 ÷ 90 = 50%;
• neutral: 30 ÷ 90 = 33.3%;
• negative: 15 ÷ 90 = 16.7%.

Ten items remain unscored. That missing share belongs in the report.

Report scoring coverage

Excluding unknown items from the sentiment denominator is reasonable only when the report also shows how many items were scored.

Scoring coverage = scored eligible items ÷ all eligible items × 100

In the example, scoring coverage is 90%. A favorable average can be misleading when it comes from a small or biased subset.

Another valid method includes "unknown" as a fourth category. Choose one method, disclose it, and keep it stable.

Net sentiment

Net sentiment compresses the distribution into one number:

Net sentiment = positive share minus negative share

Using the example above, net sentiment is 50 minus 16.7, or +33.3 percentage points. The range runs from -100 to +100 when the inputs are percentages.

Net sentiment is a custom summary, not a universal AMEC standard. Different distributions can produce the same value. A sample with 50% positive and 20% negative has the same net score as one with 30% positive and no negative items, yet the communication situations differ.

Always show the underlying positive, neutral, negative, and unknown shares.

Build the codebook before analysis

The coding rule determines the result. Define at least:

1. the unit, such as whole article, paragraph, quotation, brand appearance, or AI answer;
2. the target, either overall item tone or stance toward a named brand;
3. the categories and their thresholds;
4. the treatment of mixed coverage;
5. the treatment of quoted criticism, irony, negation, and factual allegations;
6. the rule for unknown or insufficient context.

For example, a factual report of declining revenue might count as negative toward the company even when the journalist's language is neutral. Another codebook may reserve "negative" for explicit criticism. Both choices can be used, but they cannot be compared as if they were identical.

Human and automated coding

Human coders understand context but can disagree. Train them on the same codebook, test a shared sample, and report intercoder reliability when the analysis supports important decisions.

Automated classification can process much larger volumes. It still needs validation against human-coded material from the relevant languages, channels, and topics. Sarcasm, mixed viewpoints, quotations, legal reporting, and local phrasing remain difficult.

Revalidate after changing the model, prompt, threshold, or codebook. A historical trend can break because the measurement changed, even if the coverage did not.

Use automation to scale a defined method, not to avoid defining one.

Sentiment in aclipp

aclipp keeps two concepts separate.

Clipping Sentiment scores the overall tone of a media item on a 0 to 100 scale. The product displays negative, neutral, and positive bands. The score applies to the clipping as a whole and can exist even when no tracked brand appears.

AI Brand Sentiment scores the stance toward one tracked brand in one answer run. aclipp averages the available brand-level scores under the selected filters. Unscored rows are excluded, so the scored count matters.

These thresholds and averages are aclipp product conventions, not industry standards. The general AI visibility guide explains how sentiment fits beside mentions, citations, and Share of Voice.

Read sentiment with other content metrics

Sentiment becomes more useful when it answers a specific communication question.

• Read it with Share of Voice to see whether high presence came from favorable or critical coverage.
• Read it with key message pull-through to separate message presence from the way that message was framed.
• Review target media and prominence so a flood of minor neutral mentions does not hide a small number of important critical stories.

Do not multiply these factors into an unexplained reputation score. Keep the components visible unless a published decision rule genuinely requires a composite.

The boundary of sentiment data

Content sentiment does not establish:

• audience emotion or attitude;
• trust or reputation;
• recall or understanding;
• purchase intent or behavior;
• the cause of a business result.

It can show whether analyzed content became more positive or negative under a stable method. The next question is whether the right audience encountered that content and whether anything changed. Those questions need different evidence.

Our PR KPI guide places sentiment at the output level and shows which methods belong to outtakes, outcomes, and organizational impact.