Sentiment is the polarity of language towards its subject: favourable, unfavourable, or neither. In advertising the word rarely describes a feeling. It describes a label produced by software that has read a web page, a podcast transcript, a product review or an answer generated by a chatbot, judged whether the subject is being praised, criticised or merely described, and attached a value that can then be targeted, avoided, priced or reported to a client.

The label exists because subject matter alone is a weak predictor of commercial value. A travel segment can recount a dream holiday or a cancelled flight. Both carry the same Interactive Advertising Bureau (IAB) content category. One suits an airline, the other a travel insurer. Topic classification cannot separate them, and the industry has spent two decades building systems that try.

How a score is produced

Every implementation begins by converting content into text. Web pages are crawled, podcast and streaming audio is transcribed by automatic speech recognition, video frames are read by computer vision models, and answers from AI assistants are collected by issuing queries and capturing what comes back. A classifier then assigns polarity, either from a lexicon of scored words adjusted for negation and intensifiers, or from a model trained on labelled examples.

Output formats diverge, the first obstacle to comparing vendors. Google's Cloud Natural Language API returns two numbers per document: a score between -1.0 and 1.0 for direction, and a magnitude beginning at 0.0 and rising without a ceiling for the quantity of emotional language present. The pairing matters because a score near zero is ambiguous. Google's documentation works the point through film reviews. A negative review returns a score of -0.6 against a magnitude of 3.3. A mixed review returns 0 against 4.7. An emotionally flat review returns -0.1 against 1.8. Strip out the second number and the mixed review becomes indistinguishable from the dull one.

Granularity is the second variable. Google's documentation uses the sentence "I liked the sushi but the service was terrible" to show why a single document score can be useless, and offers entity-level analysis, scoring each subject separately, as the remedy.

Advertising inherits that problem at scale. Sounder, the audio intelligence unit Triton Digital acquired on March 26, 2024, began applying one of four labels to every IAB category detected in a podcast episode on July 2, 2026: positive, negative, neutral or mixed. Scoring runs at segment level first, so a finance episode can register negatively on market volatility and positively on investment advice, then aggregates upward. The company describes the episode label as reflecting the overall distribution of segment signals, without disclosing the weighting.

Scales vary. HubSpot's answer engine optimisation product, launched in April 2026 at $50 a month, scores brand descriptions across ChatGPT, Gemini and Perplexity from -100 to +100. SISTRIX reports a balance of counted instances instead: its July 2026 example for Porsche recorded 726 instances of praise against 271 of criticism, a net score of positive 46, with Google AI Mode at positive 67 and ChatGPT at positive 7 on the same brand.

Where the label travels

OpenRTB, the real-time bidding specification, carries no sentiment field. Its content object holds category arrays and, since version 2.6 in April 2022, a cattax field declaring which taxonomy those arrays use, but nothing for tone. Classification reaches the auction indirectly, as third-party segment identifiers on a bid request or as a filter applied inside a demand-side platform (DSP) before bidding.

Integral Ad Science (IAS) built the best-documented buy-side implementation. The company acquired contextual specialist ADmantX in November 2019, then launched Context Control in April 2022 with sentiment and underlying emotional classification attached to page-level analysis. The pitch was homonym resolution: separating a gun shot from a basketball shot, which keyword blocking cannot do. IAS now offers more than 380 contextual segments activated across major DSPs, and reported that Teads recorded a 99% suitability pass rate after integrating the avoidance products in January 2023.

Sell-side packaging runs the other way. Publishers on Triton's platform pair an IAB category with a tone label to build inventory bundles, the published examples being Travel-Negative for insurance advertisers and Adult Education-Positive for course providers. That stack reached The Trade Desk as pre-bid signals on May 20, 2026.

Formats without text need an extra step. Comscore extended transcript-level classification across Spotify, SiriusXM, Triton Digital, Acast and Libsyn on July 22, 2026. In live video, Amazon Web Services runs computer vision models alongside encoding to identify objects, actions, brands, on-screen text and sentiment, mapping results to the IAB taxonomy and the categories written by the Global Alliance for Responsible Media.

Origin and evolution

Two papers from 2002 opened the field. Peter Turney's ACL paper applied semantic orientation to unsupervised review classification, and Bo Pang, Lillian Lee and Shivakumar Vaithyanathan tested machine learning classifiers on the same task at EMNLP. The term is generally credited to Tetsuya Nasukawa and Jeonghee Yi of IBM, whose paper at the Knowledge Capture conference in Florida, held from October 23 to 25, 2003, reported precision between 75% and 95% on web pages and news articles.

Commercial adoption ran through social listening, where tone became a standard column in brand monitoring dashboards covering forums, reviews and social platforms. Contextual advertising absorbed the technique as third-party cookies came under pressure, which is the arc the IAS acquisition and product launches trace. Audio followed roughly four years behind display, once transcription quality made spoken-word classification practical. The newest destination is the answer engine. Meltwater extended brand monitoring across AI platforms in August 2025, and Similarweb shipped sentiment features inside an AI search tracking toolkit the same summer.

Why the label moves money

Sentiment decides where budget lands, and news absorbs most of the consequence. Vodafone reported in July 2025 that revised settings cut its block rates by 41% and opened 10% more news inventory, partly by switching to exact-match keywords so that a blocked "die" stopped catching "diet". Tone classification is the mechanism sold as the alternative to those lists.

Platform controls use the same vocabulary. Microsoft Advertising, which began rolling out exclusions for up to 1,000 page title terms in August 2026, defines its natural disasters exclusion as covering negative news or sentiment around hurricanes and tropical storms. An advertiser switches off a tone with a checkbox and a publisher loses inventory.

Measurement has now formalised it. IAB published Measuring Visibility in the AI Era on August 3, 2026, placing sentiment inside a Portrayal tier described as a brand safety dimension unique to AI and requiring providers to disclose classification methodology, framing taxonomy and accuracy benchmarks. That exposure has no equivalent in display buying, since a brand cannot buy its way out of being characterised badly in an answer it never paid for.

Limitations and disputes

Accuracy claims are mostly unverifiable. Sounder published no classification method, no accuracy rate and no validation methodology for its four-way labels. The AWS materials do not say how sentiment is scored in live video. IAS stated at launch that its semantic technology was 42% more accurate than the next best offering, a vendor-supplied figure with no published benchmark behind it.

IAB's own framework is the sharpest published critique. It records that sentiment carries most meaning in informational queries while recommendation and commercial contexts tend towards neutral descriptive language, and that classification accuracy varies across tools and is particularly unreliable for nuanced or comparative language. Sarcasm, negation and comparison remain the standing failure modes, as they were in 2003.

Negative is not the same as unsuitable, and conflating the two is the oldest complaint in the category. Reporting on a hurricane, a court case or a recession scores negatively and is ordinary journalism. The dispute has since turned legal: GARM, whose categories still populate platform documentation, was disbanded in August 2024 days after X filed an antitrust suit alleging an illegal boycott.

Measurement bodies have drawn one line. As set out in PPC Land's explainer on brand suitability, the Media Rating Council ruled on October 18, 2025 that vendors cannot describe property-level or keyword-level checks as brand safety unless they also analyse images, video and audio, with a grace period running to April 18, 2026.

Not the same as

Brand suitability is a policy decision about which content a given advertiser will accept. Sentiment is one input to that decision, not the decision itself, and suitability frameworks are organised by risk category rather than by tone.

Emotion classification goes further than polarity, sorting content into anger, joy, fear or sadness. Google's documentation is explicit that its API reports angry or sad text only as negative, without naming the emotion.

Consumer sentiment in economics is a survey index, not a text classification. The University of Michigan measure, created by George Katona in 1946 and benchmarked to the first quarter of 1966 at 100, read 44.8 in May 2026 on the unadjusted series. Advertising research uses surveys the same way: FreeWheel found in 2026 that 48% of buyers assumed viewers disliked AI-generated creative while 10% of viewers said so.

Share of voice counts how often a brand appears relative to competitors. It is a volume metric that says nothing about whether the mentions are flattering.

Recent developments

Classification of AI answers is where activity has concentrated, but the older surfaces keep moving. DoubleVerify's February 2026 surveys found 43% of North American consumers reporting negative feeling towards low-quality AI-generated advertising against 23% positive, a measure of audience attitude rather than content tone. In podcasting, Basis brought episode-level suitability controls into its DSP on August 13, 2026, continuing a run of pre-bid integrations that began that January.

Timeline

  • 2002 - Turney publishes semantic orientation work at ACL; Pang, Lee and Vaithyanathan publish machine learning sentiment classification at EMNLP
  • October 23 to 25, 2003 - Nasukawa and Yi present the paper credited with naming sentiment analysis, reporting 75% to 95% precision
  • November 2019 - IAS acquires ADmantX
  • April 2022 - IAS launches Context Control with sentiment and emotional classification; OpenRTB 2.6 adds the cattax field
  • January 2023 - Teads reports a 99% suitability pass rate using IAS avoidance classification
  • March 26, 2024 - Triton Digital acquires Sounder
  • August 2024 - GARM is disbanded following the X antitrust complaint
  • October 18, 2025 - The MRC restricts brand safety claims to vendors analysing images, video and audio
  • April 2026 - HubSpot ships a -100 to +100 brand sentiment score across three AI platforms
  • May 20, 2026 - Sounder contextual and suitability models go live inside The Trade Desk
  • July 2, 2026 - Sounder adds four-way tone labels to IAB category classification
  • July 22, 2026 - Comscore extends transcript-level classification across five audio platforms
  • August 3, 2026 - IAB publishes Measuring Visibility in the AI Era, defining sentiment as a Portrayal metric
  • August 2026 - Microsoft Advertising rolls out page title exclusions covering up to 1,000 terms
  • September 2026 - AWS documents computer vision sentiment detection in live video

Summary

Who: Classification vendors including IAS, DoubleVerify, Peer39, Comscore and Sounder produce the labels; platform providers including Google, Microsoft and AWS embed them in buying and encoding tools; IAB defines the measurement vocabulary; advertisers, agencies and publishers act on the output.

What: A machine-assigned judgement of whether content or a brand mention is positive, negative, neutral or mixed, expressed variously as a bounded score, a four-way label or a balance of counted instances, and used to target, avoid or package inventory and to monitor brand portrayal.

When: Named in academic work in 2003, commercialised through social listening in the following decade, added to contextual advertising products from 2019 onward, extended to podcast audio in 2026 and formalised as an AI visibility metric on August 3, 2026.

Where: Applied to web pages, podcast and streaming transcripts, live video frames, social and review text, and answers generated by ChatGPT, Gemini, Perplexity and comparable assistants, activated inside DSPs, publisher packaging tools and monitoring dashboards.

Why: Topic classification cannot distinguish content that flatters a subject from content that attacks it, and the distinction changes what inventory is worth, which advertisers will buy it, and how a brand is represented in answers it does not control.