How can a list of numbers hold meaning?
Picture a map where every sentence has a position, and sentences about the same thing sit near each other. "Wet weather backup" and "what if it rains" are neighbours; "minimum spend on a Saturday" is across town. An embedding is that position, except the map has hundreds or thousands of directions instead of two. OpenAI's embeddings guide, for example, lists a model that returns 1,536 numbers for every piece of text.
The positions are learned from how words are used together across a huge amount of text. The idea became famous in 2013, when Tomas Mikolov and colleagues at Google published word2vec and showed that the arithmetic works on meaning: take the numbers for "king", subtract "man", add "woman", and the closest word is "queen".
Where do embeddings already turn up in your marketing?
In search, mostly. When Google announced in 2019 that a model called BERT was helping it read search queries, the point was matching what people meant instead of the exact words they typed (Google, "Understanding searches better than ever before"). Broad match in Google Ads works on related meaning as well: Google Ads Help describes it as showing your ad on searches related to your keyword, including searches that do not contain its words. And any AI tool that answers from your documents uses embeddings to find the right passage first, which is retrieval-augmented generation.
How does a wedding venue match questions to answers? (example figures)
A wedding venue near Cessnock gets about 120 enquiry emails a month, and most ask one of 25 questions it has answered many times. It stores those 25 answers with an embedding for each. Every new email is embedded too, and the tool finds the closest saved answer and scores the match from 0 to 1.
| Question in the email | Closest saved answer | Score | What happens |
|---|---|---|---|
| "Can we still do the ceremony outside if it pours?" | Wet weather plan: the barrel room | 0.86 | Draft reply offered |
| "Is there a minimum spend on a Saturday in October?" | Peak season pricing, from $14,000 | 0.81 | Draft reply offered |
| "Are sparklers OK for the exit?" | Sparklers and confetti policy | 0.88 | Draft reply offered |
| "Can our dog be the ring bearer?" | Children and guests | 0.41 | Sent to the owner |
Below a cut-off the owner sets, here 0.6, nothing is drafted and the email goes straight to a person. If 90 of the 120 emails get a usable draft and each reply drops from 10 minutes to 3, that is 10.5 hours a month, about $470 of time at $45 an hour.
What can embeddings not tell you?
Closeness is not correctness. "We allow sparklers" and "we do not allow sparklers" sit almost on top of each other, because they are about the same thing and the word "not" moves the numbers very little. A matched answer still needs the check a person would give it.
Embeddings also go stale. When the October pricing changes, the saved answer and its embedding have to be redone, or the tool keeps matching questions to the old figure. And the score is a measure of similarity that the owner has to calibrate: a cut-off that suits one inbox will be too loose or too strict for another.
Do you need marketing jargon to talk to Cora?
No: for anyone using an AI like CoraVision, matching on meaning is why you can talk about your business without marketing vocabulary. Say "people who came once and never came back" and the model reads it much as it reads "lapsed customers". In the onboarding call, Cora asks for your goal in your own words, and plain words are what the model is built to read. That goal then becomes the one number your 12-month plan is built around, with weekly check-ins against it.
Questions
Are embeddings the same as keywords?
They do different jobs. A keyword search finds the letters you typed, so a search for "rain" misses "wet weather"; an embedding search finds nearby meaning and catches it. Good search tools use both, because exact words still count for product codes, names and prices.
Can someone read my text back from its embedding?
Partly, and sometimes closely. Researchers at Cornell showed in 2023 that text can often be largely rebuilt from its embedding, in a paper titled "Text Embeddings Reveal (Almost) As Much As Text". Treat embeddings of customer emails with the same care as the emails themselves.