How does semantic targeting match an ad to the meaning of a page?
Part of Contextual targeting
The short answer
Semantic targeting turns the main text of a page into numbers that capture its meaning, called embeddings, and compares them with a written description of the content a brand wants. Pages whose meaning sits close to that description qualify, even when they share none of its words. Each match should come with the passages that earned it.
Key takeaways
- Semantic targeting compares the meaning of a page with the meaning of a written brief, so synonyms and paraphrases count as matches.
- The method depends on three choices that buyers may not see: which text is read, how the brief is described and where the match threshold sits.
- Keyword and category methods still win on some page types, so a balanced plan uses each where it reads best.
- A semantic score is only as trustworthy as the evidence behind it, which means the matching passages should be visible.
- Audit a semantic result by reading its best and its weakest matches, not by trusting the score.
What is semantic targeting in advertising?
Definition
- Semantic targeting
- A contextual method that places ads by comparing the meaning of a page's text with a description of the ideal content, using numeric representations of language rather than lists of words or fixed category labels.
The numeric representations are called embeddings. Google's Machine Learning Crash Course (opens in a new tab) explains that embeddings place similar items closer together, and that the distance between two points stands for how alike the items are. A paragraph about repotting a fern and a paragraph about moving a monstera into a bigger container land near each other, although they share almost no nouns.
Older word embeddings gave every word one fixed point, so a word with two senses was stuck between them. Models in the family introduced by BERT (opens in a new tab) read each word together with the words on both sides of it, which is why current semantic systems can tell a garden bed from a hotel bed. Sentence level models such as Sentence-BERT (opens in a new tab) then produce one vector per passage, which can be compared at the speed an ad system needs.
How does meaning-based matching work on a single page?
A semantic pipeline can be described in six stages, whatever the vendor or the model. Knowing them helps a buyer ask where a strange placement came from.
What happens at each stage
- Extract the article body. A page's navigation, related story links and ad slots can outweigh a short article. If they are not removed, a recipe post can score as a page about shopping because of its sidebar.
- Split long pages. One vector for a 3,000 word guide blurs its sections together. Passages keep a page about soil, pots and pests from averaging into a vague page about plants.
- Encode passages and the brief with the same model. The brief is written as a short description of the article the brand wants beside its ad, and it is converted the same way, so both live in one space.
- Score similarity. Cosine similarity is one measure. Google's guide to measuring similarity from embeddings (opens in a new tab) notes that cosine looks only at the angle between vectors and ignores popularity, while the dot product grows with vector length.
- Aggregate with a rule a person can state. For example, a page qualifies when its best passage clears the threshold, and a site qualifies when most sampled pages do.
- Keep the evidence. Store the passages that matched, so a planner can read why a page was chosen and a client can check it.
How does semantic matching compare with keyword and category methods?
Each method reads some pages better than the others. The fairest way to judge them is by page situation, because a method that is excellent on one kind of page can be weak on the next.
| Page situation | Keyword lists | Category labels | Semantic matching |
|---|---|---|---|
| The page uses different words for the brief's subject | Misses it unless every synonym is on the list | Catches it when the category is broad enough to cover the subject | Can catch it, since paraphrases sit close in meaning |
| The brief names an exact product, model or event | Strong: the exact term is either present or absent | Weak: a name may have no category of its own | Can drift to related products unless the brief is precise |
| The page is in French or Spanish | Needs a separate list per language | Works when the classifier supports that language | Works with a multilingual model; quality varies by language |
| The niche has no category of its own | Works if the niche has distinctive vocabulary | Weak: the page lands in a crowded parent category | Strong: the description can be as narrow as the brief |
| The buyer must explain the decision to a client | Strong: the matched word can be pointed to | Strong: the label is a shared industry term | Only as strong as the passages shown; a bare score explains nothing |
| A buying platform must apply it at bid time | Supported, for example as content keywords and keyword exclusions in Google Ads | Supported: OpenRTB bid requests can carry standard category IDs | Support depends on the platform or partner doing the reading |
Where does semantic matching misread pages?
Semantic systems fail in patterns. Knowing the patterns lets a planner check for them on purpose instead of finding them in a placement report.
| Misread | Usual cause | What a planner can check |
|---|---|---|
| Page scored on its template | Menus, link lists or ads were embedded with the article | Read the matching passage: is it article text or page furniture? |
| Tone missed | Satire, parody and complaint posts use the same words as sincere advice | Open a sample of top matches and read the first paragraph |
| Short pages score high | A two sentence product blurb can sit close to the brief by chance | Look at page length beside the score; thin pages need more evidence |
| Topic drift | The brief description was broad, so neighboring subjects cleared the threshold | Read the weakest matches above the threshold, not only the best |
| Uneven languages | The model is stronger in some languages than others | Have a fluent reader check matches in each language of the brief |
| One strong page carries a site | Site scores were built from the single best page | Count how many sampled pages matched, not only the top score |
The threshold is a setting buyers may never see. Set it high and the list is small and precise; set it low and reach grows while topic drift creeps in. There is no correct universal value, because similarity scores depend on the model. Ask for the threshold in plain words, such as the share of sampled pages that must match, and ask to see pages just above and just below it.
What does semantic targeting look like on a real brief?
Fictional example
An indoor plant brief for Colorado and Utah
An unnamed online houseplant shop asks for display ads in Colorado and Utah beside care guides for indoor plants. The planner writes the ideal content as a practical article on keeping tropical houseplants healthy in a dry, high altitude home, covering watering, humidity and light. The avoid description is pages about toxic plants and pet poisoning emergencies.
A keyword list built around the word houseplant would have missed some of the best pages, which talk about pothos, humidity trays and grow lights without using it. The semantic match finds them, and the planner reads the top passages to confirm they are care advice rather than product listings.
Two weak spots show up in review. A gardening forum scores well, but its matching passages are short replies, so the planner asks for more sampled pages before adding it. A lifestyle site matches on one plant care column only, so it goes on the list as a test with that reason written beside it.
How does PlacementGraph apply meaning-based matching to a brief?
How can a buyer audit a semantic targeting result?
A semantic result deserves the same scrutiny as any other list. These checks target the misreads described above and can be split among the planners on the account.
Semantic result audit before approval
- The brief's ideal content description was written as an article, not as a list of keywords.
- For a sample of sites, the matching passages were opened and read in full.
- The weakest matches that still qualified were reviewed, not only the strongest.
- Every matching passage was article text, not navigation, ads or comment threads.
- Pages in each language of the brief were checked by a fluent reader.
- Each site's reason states how many sampled pages matched.
- The threshold or match rule is written down in plain words with the list.
Frequently asked questions
Does semantic targeting work in French and Spanish?
Yes, when the system uses a multilingual model. Research on multilingual sentence embeddings, such as Reimers and Gurevych (2020) (opens in a new tab), trains models so that a sentence and its translation land in the same place in vector space. Quality still varies by language, so a fluent reader should review matches in each language the brief covers.
Can semantic targeting read video as well as text?
Partly. A semantic system that reads text reads titles, descriptions and other metadata. Google's help page about contextual targeting (opens in a new tab) says its own system considers text, metadata, visual content, audio content and channel information when it judges a YouTube video's central theme.
How close does a page have to be to count as a match?
There is no industry standard. Similarity scores depend on the model, so a value that is strict for one system is loose for another. Ask the provider to express the rule in plain terms, such as the share of sampled pages that must match, and to show pages just above and just below the line.
Why can a semantically matched site still serve ads on off-topic pages?
Because plans can target whole domains. A site can earn its place from its matching articles and still carry unrelated sections. If a buying platform targets the domain, ads may appear anywhere on it, so pair site lists with page level controls where the platform supports them and read the served placement report.
Is semantic targeting the same as artificial intelligence targeting?
Semantic targeting uses machine learning models, but it is narrower than the label AI targeting suggests. It answers one question: how close is this content in meaning to the content described in the brief? It does not predict who will buy, and it does not need to know anything about the person reading the page.
Sources
- Embedding space and static embeddings (opens in a new tab)Google for Developers, Machine Learning Crash Course
- Measuring similarity from embeddings (opens in a new tab)Google for Developers, Clustering course
- BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (opens in a new tab)Devlin, Chang, Lee and Toutanova, arXiv, 2018
- Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks (opens in a new tab)Reimers and Gurevych, EMNLP 2019, arXiv
- Making Monolingual Sentence Embeddings Multilingual using Knowledge Distillation (opens in a new tab)Reimers and Gurevych, EMNLP 2020, arXiv
- About contextual targeting (opens in a new tab)Google Ads Help
- 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.