Different Label, Right Ad: How Vectors Match Ads Across Categories

Contextual advertising is an old idea: put the ad next to content it fits. Companies such as Peer39, GumGum and Seedtag built whole businesses on it. Their machine learning models run alongside the DSP, read the page and tell the buyer which pages suit which ad.
Without those models, a buyer mostly has IAB categories. A category says roughly which section of the internet a page sits in. It says little about what the page is actually about, and nothing about how well it fits a particular ad. The new Agentic Audiences spec could change that quickly, because it lets the meaning of the page travel inside the bid request itself, as a vector.
Here is what that looks like on one page and one ad.
Picture a general article on a site for older readers: Technology Grandparents Actually Use: Big Buttons, No Cables. It is not a phone review. It talks about what older people keep asking for: large, separate buttons on the TV remote and on the phone, no fiddly cables and tiny charging sockets, loud and clear sound, one button that calls family.
Now picture EasyCall, a brand that makes exactly that phone: big buttons, a large screen, loud sound, an SOS key. That article is one of the best places on the internet for its ad.
With today's IAB labels, the ad will not appear there. With vectors, it will.
Two ways to describe a page
When a page has an ad slot to sell, the ad auction needs to tell buyers what the page is about. There are two ways to do it: give the page a category, or give it a vector.
The IAB Content Taxonomy is a tree of a few hundred topics. Each page is filed under one of them.
"cat": ["IAB6-9"]An embedding model turns the text into a long list of numbers: an arrow in a space of meaning. Texts about similar things point in similar directions.
Here is what that difference means for our article and EasyCall's ad, in one picture:
What labels see
Ad auctions describe a page with a category from the IAB Content Taxonomy, and advertisers pick the categories they want to appear in.
- The article is about older people, so it is labelled Eldercare.
- EasyCall sells phones, so it targets Cell Phones.
Eldercare is not Cell Phones. By label, the ad and the article have nothing in common, and EasyCall never gets to bid on a page that practically describes its product.
What vectors see
An embedding model turns a piece of text into a vector: a long list of numbers that captures what the text means. Texts about similar things get similar vectors, whatever words or labels they use. The IAB Tech Lab's Agentic Audiences work proposes sending such a vector with the bid request. We wrote about how that works in Text to Vector to Match.
So we tried it, with one of the most widely used embedding models, OpenAI text-embedding-3-large. We wrote the article, turned it into a vector, turned EasyCall's campaign brief into a vector, and compared the two.
Here is the article. Flip the switch from Labels to Vectors and watch the ad slot:
Technology Grandparents Actually Use: Big Buttons, No Cables
What do readers in their seventies and eighties want from everyday technology? That it works the first time, every time.
TV remotes have dozens of tiny keys, and phones hide everything behind a touchscreen. Older hands want large, separate buttons: one for the volume, one to call a grandchild.
Fiddly chargers and small charging sockets are a daily struggle. A charging stand, or a battery that lasts a week, fixes that.
Loud, clear sound that works with hearing aids, big bright text, and an emergency key that calls family with one press.
IAB6-9 EldercareEasyCall targets:
IAB19-6 Cell PhonesThe article is general. It never names a phone brand and spends as much time on TV remotes and chargers as on phones. But it talks about big buttons, loud sound and a button that calls family, for older people, and so does EasyCall's brief. The labels are different, the meaning is the same, and the vectors pick that up: a cosine similarity of 0.58, a fit of 80%. For comparison, the same brief scores around 0.1 against a typical post on this blog.
That is the whole trick: the match comes from what the page says, not from which box it was filed in.
Under the hood
For the auction, the change is small. The bid request, the message an ad slot sends to buyers, keeps its label and gets one more block with the page's vector:
{"id": "req-1324","imp": [ … ],"site": {"page": "silverdays.example/living/…","cat": ["IAB6-9"],"cattax": 1}}
- id
- The ID of this auction.
- imp
- The ad slot for sale, here a 300 × 600 banner.
- site.page
- The page the slot is on.
- cat
- The page's category: IAB6-9 Eldercare. One code, and buyers can only check whether it is on their list.
- cattax
- Which list the code comes from: 1 = IAB Content Taxonomy 1.0.
{"id": "req-1324","imp": [ … ],"site": {"page": "silverdays.example/living/…","cat": ["IAB6-9"],"cattax": 1},+ "user": { "data": [{+ "name": "agenticAudience",+ "segment": [{ "ext": { "aa": {+ "model": "text-embedding-3-large",+ "dimension": 3072,+ "type": [2],+ "vector": "AEAIPAAgsrwAwK…"}}}]}]}}
- agenticAudience
- The new block from the proposed IAB Tech Lab Agentic Audiences extension. The label above stays as it was.
- model
- Which embedding model made the vector, so buyers can embed their ad brief the same way.
- dimension
- How many numbers the vector has: 3,072.
- type
- 2 = contextual: the vector describes the page, not the person reading it.
- vector
- The page’s meaning: 3,072 numbers packed into one string of about 16 KB.
A buyer turns its campaign brief into a vector with the same model and compares the two. That comparison is the 0.58 you saw in the example.
Why this matters
None of this replaces labels overnight. Categories are still useful for brand safety and for broad buys. But a label tells you the neighbourhood, not the street, and vectors tell you the street.
At Appliscale we build the infrastructure behind this: bidders, SSP and DSP integrations, Prebid setups and the data pipelines that turn pages and briefs into vectors. If you want to see how embedding-based matching would work on your own inventory, we are happy to talk.