Cross-encoder
A cross-encoder is a model that reads a question and a passage together to score how well the passage answers it.
A cross-encoder is a model that reads a question and a passage of text together, and scores how well the passage answers the question. Because it reads the two side by side, it's more accurate than faster methods. It's also slower, so it's used on a shortlist, after a quicker step has narrowed things down.
You won't run one yourself. It's worth knowing about because it explains why being on the right subject isn't always enough to get your page into an answer.
Two steps to pick a passage
Many search systems pick passages in two steps. First, a fast method turns the question and millions of passages into lists of numbers and keeps the ones that look closest, often by cosine similarity. That gives a shortlist of maybe a hundred passages.
Then a cross-encoder reads each passage on the shortlist together with the question, and reorders them by how well they actually answer it. The top few go into the answer. This second step is often called reranking.
It's a bit like hiring. The first step reads job titles and keeps everyone who looks about right. The second step reads each application properly.
What it means for your pages
The first step rewards being about the right subject. The second rewards actually answering the question.
For example, someone asks "how long do roasted coffee beans stay fresh?" Fernhill's page says "Freshness matters to us, which is why we roast to order." Kestrel's page says "Whole beans are at their best for about four weeks after roasting, if you keep them sealed and away from light." Both pass the first step. The cross-encoder puts Kestrel's passage on top, because it answers the question.
AI platforms don't publish exactly how they choose passages, and each does it differently. But writing passages that answer a question directly, in plain words, helps under any method.