Hybrid search: keyword plus vector retrieval

Hybrid search, also called hybrid retrieval, combines two ways of finding relevant text in a single retrieval step.

• Lexical scoring matches the literal words in a query against the literal words in a document. BM25 is the classic function: it rewards terms that appear often in a document and are rare across the collection.
• Vector scoring turns the query and each document into embeddings and ranks documents by how close they sit in that space. This is vector search, and it finds passages that mean the same thing in different words.

Each method fails where the other is strong. Vectors can treat two similar product names as near neighbours. Keywords miss a passage that answers the question without using the query's words. Hybrid search uses both so neither gap decides the result alone.

A typical hybrid pipeline follows the same shape, whatever system runs it.

1. The system receives a query. In AI search this is often not the user's prompt: OpenAI's help page says ChatGPT search typically rewrites your query into one or more targeted queries before sending them to search providers.
2. A lexical index (for example BM25) scores candidate documents on exact term overlap.
3. A vector index scores candidates on embedding similarity.
4. A fusion step merges the two lists, usually by a weighted sum of normalised scores or by combining rank positions.
5. The top results become the candidate set the language model reads before it writes an answer.

This is the retrieval stage, the step between a crawler storing your page and a model citing it. Our guide on how AI search picks sources covers the stages on either side. As of October 2026, no major AI engine documents whether it uses hybrid retrieval, what lexical function it runs, or how it weights the two scores. Treat the mechanism as a model of how retrieval can work, not a disclosed spec.

The practical lesson is about exact facts. Your brand name, product names, prices and plan limits are precise strings. A lexical scorer can match them exactly, and a model can only quote a sentence it actually received.

That only works if the fact is in the text the crawler fetched. Vercel's crawler study found that ChatGPT and Claude crawlers fetch JavaScript files but do not execute them, so they cannot read client-side rendered content. A price that appears only after a script runs is absent from whatever index that crawler feeds, lexical or vector.

There is a small positive signal for stating facts plainly. The 252,000-trial factorial study (arXiv 2605.25517) reported that including explicit price information and a recent timestamp helped consistently. That was measured on documents already placed in the model's context, so it says nothing about getting retrieved in the first place.

The next move is concrete:

• View your page source with JavaScript off and confirm names, prices and key numbers are in the raw HTML.
• State each fact once, in a plain sentence, the same way on every page.
• Remember that access comes first. Retrieval scoring never sees a page a crawler could not reach.

"Hybrid search means keywords are back, so add more"

No. The original GEO paper (arXiv 2311.09735) found keyword stuffing lowered its own visibility metric against baseline. Lexical matching rewards the right words being present, not repetition.

"I need every long-tail variant on the page"

Google's AI optimization guide says you don't have to worry about long-tail keywords or capturing every variation of how someone might search. Vector scoring exists to bridge wording differences. See keywords vs prompts for how queries actually reach engines.

"I should chunk my content for retrieval"

Google rejects chunking as a requirement. The chunking research (arXiv 2603.06976) studies how the retrieval operator splits documents inside its own index, which publishers do not control.

"I can optimise my hybrid score"

You cannot see the weights or the rewritten queries. What the evidence does support is staying on topic: topical relevance is one of the most reproducible drivers of citation in the studies that measured it.

• Vector search: the semantic half of hybrid retrieval.
• Embeddings: the numeric representations vector search compares.
• Long-tail keywords: why exact query variants matter less than they once did.
• How AI search picks sources: retrieval in context, from crawl to citation.
• Logging retrieved vs cited sources: how to see which pages entered the candidate set through the APIs that expose it.

What is BM25? BM25 is a lexical ranking function used in many search systems. It scores a document higher when the query's terms appear in it often and when those terms are rare across the whole collection, with a dampening factor so extra repetitions add less and less. It knows nothing about meaning, which is why it pairs well with vector search in a hybrid setup.

How are keyword and vector scores combined? Two common approaches exist. One normalises both scores to a shared range and adds them with chosen weights. The other ignores raw scores and merges the two ranked lists by position, so a document near the top of both lists rises. Which method an AI engine uses is not published, so the weighting is outside a publisher's view or control.

Does Google use hybrid search for AI Overviews? Google does not describe its retrieval in those terms. What it does document, as of its AI features page, is that a page must be indexed and eligible to show with a snippet to appear as a supporting link in AI Overviews or AI Mode, and that no special optimizations are needed. Index and snippet eligibility are the checks you can actually run.

How do I check that my key facts are in the raw HTML? Open the page, view the source rather than the rendered page, and search for your price, product names and other key numbers. You can also fetch the URL with a command-line tool and search the response. If a fact appears only in the rendered page, it is being added by JavaScript, and crawlers that do not execute scripts will not see it.