Multilingual brand safety: how does one policy hold up across languages?
Part of Brand suitability
The short answer
Multilingual brand safety applies one written tolerance to content in several languages. The tolerance stays the same; the review changes. Translated keyword lists miss inflected forms, local slang and words that mean something else in another language, so each language needs a fluent reviewer, shared category labels and a record of who judged what, when.
Key takeaways
- Keep one tolerance per category for the brand, and adjust how each language and market is reviewed, not what the brand accepts.
- Country decides what is sensitive and legal; language decides whether the reviewer and the classifier can read the page correctly.
- A translated keyword list fails on exact match spelling, inflection, double meanings and regional vocabulary.
- A review log that records the language, the locale, the reviewer and the policy version makes multilingual decisions auditable.
What is multilingual brand safety?
Definition
- Multilingual brand safety
- Applying one brand's safety floor and suitability tolerances to content in more than one language, with review methods that read each language on its own terms rather than through a translation of the English rules.
For a Canadian plan this is the ordinary case, not an edge case. The Canadian Heritage statistics page (opens in a new tab), drawing on the 2021 Census, counts 8,066,633 French-speaking people, 22.0 percent of Canada's population, including 84.1 percent of Quebec's population, and 18.0 percent of Canadians bilingual in English and French. Add community languages in the large cities and a single national campaign can reach pages in four or five languages.
Standards have expected this for years. The Media Rating Council's 2018 supplement on content level brand safety (opens in a new tab), section 5, says "Different languages and language-related nuances should be considered when analyzing metadata including dialects" and that processes should account for "different or multiple languages present."
How should country level suitability differ from language level review?
Country and language are two separate layers, and a plan can go wrong when one stands in for the other. A French page read in Quebec and the same page read in France sit under different laws and different public debates. An English page in Toronto and a French page in Montreal can sit under the same law and need different readers.
| Question | Country or market layer | Language layer |
|---|---|---|
| What it decides | What is sensitive, legal or debated where the ad runs | Whether a reviewer or classifier can read the page correctly |
| Example from Google | Display & Video 360 says sensitive social issues include those "controversial in specific countries and markets" | Display & Video 360 lists an unsupported language among the causes of classification errors |
| Who owns it | The brand and its market leads | The reviewer pool and the verification partner |
| What changes it | Elections, new laws, local news events | New languages on the plan, classifier coverage, reviewer turnover |
| Example failure | A US list of debated issues applied to a Canadian plan | An English reviewer judging French pages through machine translation |
The Display & Video 360 sensitive categories page (opens in a new tab) names both layers in its own words. It says classification errors can happen when "the site is in an unsupported language," and it defines sensitive social issues to include ones that are controversial in specific countries and markets. A setup that handles one layer and not the other leaves a gap the platform has already described.
Why do translated keyword lists fail?
A translated list carries the English list's logic into a language that does not share its grammar or its slang. Google's content suitability page (opens in a new tab) states the mechanics: "Excluded content keywords only support exact matches, so they have to be spelled correctly," and "some words have multiple meanings." Both sentences bite harder outside English.
| Failure | What happens | Example |
|---|---|---|
| Inflection | Exact match needs each gender, plural and verb form listed separately | French adjectives and participles change with gender and number, so one English word becomes several French entries |
| Accents | An accented and an unaccented spelling are different strings | A list typed without accents misses pages that use them, and the reverse |
| Cross language collision | A word on one language's list is harmless in another | "Pain" on an English health list matches French bakery pages, where pain means bread |
| Regional vocabulary | The translation uses the standard word, while local writing uses another | A term from France misses the Canadian French word for the same thing |
| Different hot topics | The subjects that draw toxic content differ by region | Researchers found equivalent search terms returned different results in English, French and Arabic |
searching for equivalent terms in the three languages led to different results at first.
That 2019 study, collecting English, French and Arabic posts, revised its search words three times per language and reported code switching, with Hindi, Spanish and French words appearing inside English posts. It is research on social media, not on advertising inventory, but it shows the same problem a translated blocklist has: the words that mark a risky page are not the same words in each language. For the English side of list design, see the keyword blocklists guide.
How do you review suitability in French content?
Give French pages a reviewer who reads French natively or fluently, working from the same tolerance document as the English reviewer. The 2019 study did the equivalent for its labels: it wrote three aligned annotation guidelines in English, French and Modern Standard Arabic and "requested native speakers to annotate the data."
A French suitability review in five steps
- Translate the policy, not the keywords. Give the reviewer the tolerance per category in French, with the brand's own examples rewritten in French rather than translated word for word.
- Use shared category labels. Record decisions against the same category names or taxonomy IDs used for English, so results compare across languages. IAB Tech Lab lists Content Taxonomy 3.1 with a December 2024 release date.
- Confirm the locale. Check whether a French site serves Canadian readers or readers in Europe before judging which debates and laws apply, as the French Canadian advertising guide explains.
- Calibrate with a bilingual reviewer. Have one bilingual person grade a shared sample in both languages, so the French and English reviewers apply the same line.
- Log disagreements. When reviewers split on a page, record both views and the final call, since those cases show where the written tolerance is unclear.
Should tolerance differ by language?
No. The tolerance describes the brand, and a brand does not become more or less comfortable with graphic violence because a page is in French. What differs is calibration: which local debates count as sensitive, which slang counts as strong language, and how well the tools read each language. Research shows why calibration needs people from each market.
In Multi³Hate (opens in a new tab), presented at NAACL in April 2025, Bui, von der Wense and Lauscher built 300 parallel meme samples across five languages and found that "cultural background significantly affects" hate speech annotation, with average agreement between country groups of 74 percent. The vision and language models they tested aligned "more closely with annotations from the US than with those from other cultures," even when the content was in another language. French was not one of the five languages, so the study shows a pattern, not a Canadian accuracy rate.
How do you document multilingual review?
The Trustworthy Accountability Group's Brand Safety Certified Guidelines (opens in a new tab), version 5.0 of July 2026, require certified buyers and intermediaries to document how inclusion and exclusion lists are maintained, to give TAG the method for how keywords, content categories and properties are added and removed, and to review those lists quarterly. A multilingual plan meets the spirit of that by keeping one log across languages.
What a multilingual review log records
- The page or site, its language and its locale, such as Canadian French or European French.
- The category, the level assigned and the version of the tolerance document used.
- The reviewer's name and the languages that reviewer is qualified to judge.
- Whether the decision came from a classifier, a person or both, and any disagreement.
- The date, and the next review date for sites kept with a caveat.
- Keywords added or removed per language, with the reason, so each language's list has its own history.
What does multilingual review look like on a national plan?
Fictional example
A parcel courier's holiday shipping campaign across Canada
A fictional national parcel courier runs a holiday shipping campaign across Canada in English, Canadian French, Simplified Chinese and Tagalog. The earlier plan used an English blocklist machine translated into the other three languages, and the French version removed recipe and bakery sites while missing a French forum thread full of local slang the English list never had.
The new setup keeps one tolerance document for all four languages. A Montreal based reviewer handles French with a French list written from scratch, and two contract reviewers handle Simplified Chinese and Tagalog community news sites from a shared sample. A bilingual planner grades twenty pages in English and French to calibrate the reviewers. Each decision goes into one log with the language, locale and reviewer, and the lists are reviewed each quarter.
How does PlacementGraph handle suitability in several languages?
Frequently asked questions
Can a classifier's confidence score be trusted equally in each language?
Not without checking. Google's own Display & Video 360 page lists an unsupported language as a cause of classification errors, and the Multi³Hate study found tested models leaned toward US judgments. Ask the verification partner which languages it supports and spot check the rest by hand.
Does a brand need a separate suitability profile for each language?
No. Keep one profile so the brand's tolerance stays consistent, then attach a short annex per language with local examples, the reviewer pool and that language's keyword list. A separate profile per language invites the tolerances to drift apart.
How many reviewers does each language need?
At least one qualified reviewer per language on the plan, plus one bilingual person for calibration between the two largest languages. Where a single reviewer covers a language, a second reader for disputed pages keeps one person's judgment from becoming the policy.
Should community language news sites be held to a stricter standard?
No. They are judged by the same tolerance as any other publisher, read by someone fluent in the language. Holding them to a stricter line because they are harder to review removes local publishers for a reason that has nothing to do with their content.
Sources
- Statistics on official languages in Canada (opens in a new tab)Canadian Heritage
- MRC Supplement to IAB Guidelines for the Conduct of Ad Verification: Enhanced Content Level Context and Brand Safety (opens in a new tab)Media Rating Council, September 2018
- Sensitive categories in Display & Video 360 brand safety targeting (opens in a new tab)Display & Video 360 Help
- About content suitability (opens in a new tab)Google Ads Help
- Multilingual and Multi-Aspect Hate Speech Analysis (opens in a new tab)Association for Computational Linguistics, EMNLP-IJCNLP 2019
- Multi³Hate: Multimodal, Multilingual, and Multicultural Hate Speech Detection with Vision Language Models (opens in a new tab)Association for Computational Linguistics, NAACL 2025
- Brand Safety Certified Guidelines, version 5.0 (opens in a new tab)Trustworthy Accountability Group, July 2026
- Content Taxonomy (opens in a new tab)IAB Tech Lab
Researched and written by PlacementGraph Research. Read how we research, source and update these pages. Corrections: diana@placementgraph.com.