Reranking: definition and how it works

Search systems usually work in two passes. The first pass, retrieval, is fast and broad. It pulls a pool of candidate documents using keyword matching, vector search, or both.

The second pass is reranking. A reranker scores each candidate against the query more carefully, then reorders the pool and often cuts it down. In general information retrieval, the reranker is typically a slower, more precise model that reads the query and the document together.

The output is the short, ordered list that lands in the model's context window. That list is what the model reads when it writes the answer.

A grounded AI answer usually moves through these steps:

1. Query rewriting. OpenAI says ChatGPT search typically rewrites your query into one or more targeted queries. Google says AI Mode may fan out into multiple related searches.
2. Retrieval. Each query pulls candidates from an index.
3. Reranking. The operator orders and trims those candidates.
4. Generation and citation. The model writes the answer and picks which sources to cite.

Operators don't say whether they run a separate reranker or which signals it weighs. OpenAI's full disclosure is that ChatGPT ranks search results using multiple factors, and that placement is not guaranteed. The OpenAI help article on searching the web is the closest first-party description available.

Retrieval is also discretionary. An engine can answer without searching, in which case nothing is retrieved and nothing is reranked.

Order inside the context window affects what gets cited. C-SEO Bench (NeurIPS 2025) found that moving a source higher in the LLM context was significantly more effective than rewriting its content. A 252,000-trial factorial study found topical relevance and list position were the biggest drivers of being cited first. A July 2026 critical survey of 45 GEO studies named the same two levers as the most reproducible.

The catch: those studies placed documents at position 1 by hand. No live search engine was queried. No publisher can do that on ChatGPT, Perplexity or Google.

So reranking is where a big part of the outcome is likely decided, and it sits outside your control. What you can control is earlier in the four stages of AI search:

• Server-render the facts that matter, so crawlers that don't execute JavaScript can read them.
• Allow retrieval bots such as OAI-SearchBot in robots.txt, and confirm your firewall or CDN doesn't block them.
• Check noindex, nosnippet and data-nosnippet aren't hiding content by accident.
• Answer the buyer's actual question on the page, since topical fit is the one content factor the studies agree on.

"You can optimize for the reranker"

Nobody outside the operator can observe it. C-SEO Bench found most conversational SEO rewriting methods were largely ineffective and often hurt ranking.

"Structure or markup lifts reranker scores"

There is no published evidence for this on any live engine. Google says its AI features need no special optimization beyond normal Search eligibility.

"Trackers show your reranked position"

No engine exposes candidate order. A tracker sees the final answer, not the list behind it.

"Reranking works the same everywhere"

Google says AI Overviews and AI Mode may use different models and techniques. Each engine runs its own pipeline.

• Retrieval: the first pass that builds the candidate pool.
• Vector search: one common way to fetch candidates by meaning.
• Agentic search: pipelines where the model issues several searches itself.
• Answer engine: the product that turns reranked sources into a written answer.

Is reranking the same as ranking in Google Search? Not exactly. Classic ranking orders results for a person to scan. Reranking in an AI pipeline orders candidates for a model to read, and usually trims them to a handful. Some engines may reuse search rankings as an input, but none publishes how the two relate. Treat organic rank as a loose proxy, not as your position inside an AI answer.

Can I see where my page landed after reranking? No. No engine exposes the candidate list or its order, and Google's AI reporting shows impressions only. You can see the outcome: whether an engine names or cites you for a given question. The free checker on this site shows that for up to three buyer questions, as a snapshot rather than a position.

Does every AI engine rerank the same way? No. Each operator runs its own retrieval and ordering, and Google says even AI Overviews and AI Mode may use different models and techniques. A page can do well on one engine and be absent on another. See [whether what works on one AI engine works on the others](/questions/geo-across-ai-engines) for the practical consequences.

Can a reranker read content that loads with JavaScript? Only if the crawler captured it. Vercel's crawler study found ChatGPT and Claude crawlers fetch JavaScript files but don't execute them, so client-side content can be missing from what they store. A reranker can only score text that reached the index. Put prices, product names and key claims in the initial HTML.