Sentiment analysis in advertising: what a tone score can tell a buyer
Part of Contextual targeting
The short answer
Sentiment analysis in advertising scores whether the language on a page reads as positive, negative or neutral. It measures tone, not subject, quality or risk. A storm preparation guide can score negative and still be the best page for an insurance ad, so sentiment works as a planning hint and is a weak reason to block.
Key takeaways
- A sentiment score describes the emotional direction of the words on a page, not what the page is about or whether it suits a brand.
- Negative tone can come from useful content, such as repair guides, safety advice and honest reviews, that sits close to what a buyer needs.
- Sarcasm, quoted speech and mixed articles are the classic cases where a tone score points the wrong way.
- Sentiment helps as a secondary signal that matches creative mood to page mood, checked against sampled pages.
- A blocking rule built on sentiment alone removes good inventory without removing the content a brand actually fears.
What does sentiment analysis measure on a page?
Definition
- Sentiment analysis
- An automated reading of text that estimates whether its language expresses a positive, negative or neutral attitude, and sometimes how strongly. In advertising it is applied to article text, titles or transcripts around an ad slot.
A sentiment tool can return a direction and a strength. Google Cloud's Natural Language documentation (opens in a new tab) is a clear public example: its score runs from minus 1.0 for negative to 1.0 for positive, and a separate magnitude value shows the overall strength of emotion in the text. Google notes that magnitude is not normalized for a whole document, so a long article can carry a large magnitude without being unusually emotional.
The same documentation explains why a score near zero is ambiguous. A neutral score may mean a calm, low emotion document, or it may mean a document full of feeling in both directions. Google says truly neutral documents have low magnitude while mixed documents have higher magnitude, and that the tool does not identify specific emotions. It also recommends that users define a threshold that works for them, because clearly positive and clearly negative vary by use case. That advice matters for buyers: there is no industry standard cut-off that makes a page negative.
| Question a buyer asks | Does sentiment answer it? | Where the answer comes from instead |
|---|---|---|
| Is the language upbeat or downbeat? | Yes, as an estimate | The score itself, read with its magnitude |
| What is the page about? | No | Keywords, categories or meaning-based matching |
| Is the page news, opinion or a review? | No | Purpose labels such as the IAB Tech Lab descriptive vectors |
| Does the content cross the brand's tolerance? | No | A brand suitability review by category |
| Is the page well made and read by real people? | No | Quality, audience and supply checks |
The industry's shared vocabulary reflects this split. When we read the IAB Tech Lab Content Taxonomy 3.0 descriptive vectors file (opens in a new tab) in October 2026, it described purpose with values such as News, Opinion & Op-Ed, Review and Instructional Content, and listed no sentiment or tone vector. Tone, where a plan uses it, comes from a separate tool.
Why is a negative tone not the same as unsuitable content?
Negative words follow problems, and some advertisers sell solutions to problems. A guide to clearing a blocked drain, a review that explains why a budget laptop disappointed, a column on managing a first hard winter of heating bills and a checklist for preparing a house before hail season all read as negative to a tone model. Each is also a page where a plumber, a laptop maker, an energy retailer or an insurer has something useful to say.
Researchers who study news make the same distinction. A 2010 paper by Balahur and colleagues, Sentiment Analysis in the News (opens in a new tab), argues that opinion mining on news must separate good or bad news content from the sentiment expressed toward the subject, and that an article can be read from the author's, the reader's and the text's point of view. A report on falling crime figures concerns crime and still carries good news. A glowing profile can sit on top of a grim subject.
| Page type | Why the tone reads negative | Advertisers it can still suit |
|---|---|---|
| Repair and troubleshooting guides | Words such as broken, leak, failure and damage | Home services, tools, appliance makers, warranty providers |
| Safety and preparedness advice | Risk language about storms, fire or theft | Insurers, home security, generators, roofing |
| Critical product reviews | Complaints, faults and disappointment | Rival products in the category, retailers, repair services |
| Health condition explainers | Symptoms, pain and treatment side effects | Pharmacies, clinics, health insurers, within health ad rules |
| Personal finance advice | Debt, fees, losses and stress | Banks, credit unions, tax and budgeting services |
How do sarcasm and quoted speech confuse sentiment scoring?
Sarcasm flips the meaning of strong words. Riloff and colleagues at the University of Utah, in a 2013 paper titled Sarcasm as Contrast between a Positive Sentiment and Negative Situation (opens in a new tab), note that sentiment analysis can be easily misled by words with strong polarity used sarcastically. Their examples pair a positive word with an unpleasant situation, such as loving being ignored or thoroughly enjoying shoveling the driveway. A model that counts the word love scores the line as positive.
Quoted and reported speech raises a different problem: whose sentiment is it? A product recall story may quote an angry customer, a neutral regulator and a hopeful company spokesperson in the same article. A page level score blends all three voices into one number that belongs to none of them. Satire, humor columns and fiction add a third layer, because the tone is deliberately at odds with the subject.
- Sarcasm and irony: positive words describing a bad experience, found in comments, reviews and humor writing.
- Quotations: strong language from a source the author is reporting on, not agreeing with.
- Mixed articles: a balanced review with praise and complaints that averages out near zero.
- Domain vocabulary: words such as cancer in a medical page or attack in a sports report that carry negative weight in a general lexicon.
- Model fit: tools tuned on one kind of text read another poorly. The authors of VADER (opens in a new tab), a rule-based model, describe tuning its lexicon for microblog-like contexts, which differ from long form journalism.
Balahur and colleagues reported that ignoring domain specific vocabulary improved their results on news, which is the domain vocabulary problem in research form. For a buyer, the practical lesson is to ask any vendor which text the model was trained or tuned on, and whether it scores the article body alone or also the comments and headlines around it.
When does sentiment help a contextual plan?
Sentiment earns its place when the question is mood, not permission. A light, celebratory creative may sit better beside upbeat content, and a reassuring service message may sit better beside problem solving content. In both cases the topic match comes first and the tone is a tiebreaker or a split between creative versions.
Using sentiment as a secondary signal
- Match topic first. Build the candidate list from what the pages are about, using the methods on our contextual targeting hub, before any tone data is applied.
- Decide what tone is for. Write down whether sentiment will choose a creative version, order a test, or flag pages for a human read. Avoid using it to remove sites.
- Read the score with its strength. Treat a near zero score with high magnitude as mixed content that needs reading, not as neutral content.
- Sample the pages behind the number. Open several scored pages per site and check that the tone label matches what a person reads.
- Test tone where it matters. Rotate creative versions across positive and negative pages within the same topic and compare them on the brand's own measures.
Fictional example, no results implied
A home insurance broker in Iowa and Nebraska
An unnamed independent home insurance broker in Des Moines and Omaha wants English display ads beside home maintenance and severe weather preparedness content in Iowa and Nebraska. A first sentiment pass marks most hail, wind and roof damage articles as strongly negative, and a rule that blocked negative pages would have removed nearly the whole topic.
The planner keeps the topic list built from meaning and uses tone only to pick a creative. Preparation checklists and claims explainers get a calm message about reviewing coverage before storm season. A handful of upbeat pages on spring garden and patio projects get a lighter message about protecting outdoor additions. Pages reporting injuries from a specific storm are reviewed by a person under the suitability settings, not by tone.
Should sentiment ever be used as a blocking rule?
Not on its own. A block on negative sentiment fails in both directions. It removes useful pages whose tone comes from the problem they solve, and it lets through calmly written content that a brand would rule out on subject. Because thresholds are set by each user, as Google's own documentation recommends, the same page can be blocked by one setup and allowed by another.
Before any sentiment rule touches a live campaign
- The rule changes which creative serves or which pages get read first, and does not remove sites from the plan.
- The vendor has named the text it scores, such as article body only or body plus comments.
- A sample of pages on each side of the threshold was read by a person in the page's language.
- Mixed pages, with a near zero score and high strength, go to review rather than to either bucket.
- Content the brand rules out is handled by suitability settings, not by the tone threshold.
- The threshold and the date it was set are written into the plan so a later change is visible.
How does PlacementGraph read tone and subject on a page?
Frequently asked questions
Is sentiment analysis the same as emotion detection?
Not quite. Sentiment reports direction, positive or negative, and some tools report strength. Emotion detection tries to name feelings such as anger, fear or joy. Google's Natural Language documentation says its sentiment score does not identify specific positive and negative emotions. Ask a vendor which of the two it offers and how each label was defined.
Does a positive sentiment score mean a page is brand safe?
No. Tone and risk are measured differently. An enthusiastic page can promote gambling, an upbeat forum thread can include adult content, and a cheerful review can sit on a low quality site. Treat a positive score as a statement about the language only, and run suitability, quality and supply checks as separate steps.
Can sentiment analysis read French or Spanish pages?
Some tools support several languages, but accuracy can vary by language and by kind of text. Idioms, irony and regional usage, such as Canadian French expressions, trip models trained mostly on English. Ask which languages were evaluated, on what text, and have a fluent reader check a sample of scored pages before the score influences a bilingual plan.
Should comments sections count toward a page's sentiment?
Decide deliberately and write it down. Comments can be sharper in tone than the article and can shift a page level score, while the ad slot may sit beside the article itself. One workable choice is to score the article body alone and treat heavy comment activity as a separate question about moderation and user generated content.
Sources
- Natural Language API basics: sentiment analysis (opens in a new tab)Google Cloud documentation
- Sarcasm as Contrast between a Positive Sentiment and Negative Situation (EMNLP 2013) (opens in a new tab)Riloff et al., ACL Anthology
- Sentiment Analysis in the News (LREC 2010) (opens in a new tab)Balahur et al., arXiv
- VADER: A Parsimonious Rule-Based Model for Sentiment Analysis of Social Media Text (ICWSM 2014) (opens in a new tab)Hutto and Gilbert, AAAI
- Content Taxonomy 3.0 Descriptive Vectors (opens in a new tab)IAB Tech Lab, GitHub
Researched and written by PlacementGraph Research. Read how we research, source and update these pages. Corrections: diana@placementgraph.com.