What it means
A vector database stores content as embeddings, which are numerical representations of meaning, so a system can find passages by similarity rather than exact keywords. In AI search, it is the index layer: your content gets split into chunks, each chunk gets embedded, and those vectors are what a retriever searches at query time. If a passage is not indexed cleanly into a retrievable unit, it cannot be pulled into an answer.
Why it matters
The vector database is where AI search starts, which makes how your content is chunked and indexed a technical concern, not just an editorial one. Google has been moving this way for years, from matching word N-grams to understanding embeddings and vectors, and Passage Ranking was the visible step toward scoring parts of a page independently. If your pages are crawlable, parseable, and split into clean units, they enter the index in a shape a retriever can use. If they are a tangle of nav, modals, and mixed topics, they get embedded badly and surface for the wrong queries.
Take a fintech site, LedgerLoop, that hides its API rate-limit numbers inside a JavaScript accordion. The crawler never resolves the text, so the passage never gets embedded, so it never lands in the vector database. A competitor with the same fact in plain HTML gets indexed cleanly and wins the retrieval every time.
How to use this knowledge
Make every passage crawlable. Render key content in HTML, not behind clicks or scripts, so it can be fetched and embedded.
Keep one idea per section. Tight, single-topic sections embed into cleaner vectors than sprawling pages that mix unrelated topics.
Cut response time. Models retrieve like crawlers under tight time limits, so a slow server can keep your content out of the candidate pool entirely.
Use semantic HTML. Clear headings, tables, and lists give the index structured units to store and retrieve.
Growth Memo guidance
Besides being able to understand word N-grams first and then embeddings and vectors later, it seems Google was never able to understand what good content is without user signals.
Before any content enters the model's consideration (grounding) set, it must be crawled, indexed, and fetchable within milliseconds during real-time search.
Embeddings — the numerical meaning vectors a vector database stores and compares.
Chunking — how content is split into the units that get embedded and indexed.
Retriever — the component that searches the vector database for relevant chunks.
Semantic similarity — the distance measure that decides which stored vectors match a query.
Passage Ranking — Google's move to score parts of a page independently, a precursor to passage-level retrieval.