Discovery Science · established evidence

Retrieval, Not Ranking: A Working Vocabulary for the AI Search Era

Last reviewed 2026-07-20. Written by Chandranshu Kumar, Founder, Raveneye Global. · 11 min read

To understand how AI search works, start with one correction: a generative engine does not rank your page, it retrieves passages and writes an answer from them. ChatGPT, Perplexity, and Google AI Overviews all run a pipeline that most buyer-facing marketing collapses into a single word. There are at least five distinct steps, retrieval, embedding, reranking, grounding, and citation, and each one is a separate place a business can be present or absent. Conflating them is why so much advice about being seen in AI answers stays vague. This is a working vocabulary for the retrieval era, drawn from the founding research on retrieval-augmented generation and the survey literature that formalized it. Naming the steps precisely is not academic hair-splitting. It is the foundation for measuring where you actually stand, and for telling a measured improvement plan apart from a hopeful one.

The vocabulary problem in AI search

Most writing about "AI visibility" reaches for one verb, usually rank, and applies it to a process that no longer ranks anything the way classic search did. A ten-blue-links results page ordered documents. A generative answer engine does something structurally different: it assembles a small set of candidate passages, reasons over them, writes an answer, and attaches a few sources. Each of those actions is a separate mechanism with its own literature, its own failure modes, and its own pull.

When the steps are treated as one, the advice that follows is unfalsifiable. "Improve for AI" can mean improving what is retrieved, or what is ranked within the retrieved set, or whether the answer is anchored to real evidence, or whether the credit line points to your page. These are not the same job, and a tactic that helps one can be irrelevant to another. The purpose of this piece is narrow and deliberate: to define the terms precisely enough that every downstream recommendation can be checked against them.

How AI search works: retrieval, not ranking

The architecture underneath almost every generative answer engine is retrieval-augmented generation, or RAG. The founding study paired a pretrained language model with a separate, searchable index of documents that the model queries at the moment a question arrives, rather than relying only on what it memorized during training. The reported result was that this hybrid produced more specific, diverse and factual language than a model working from its parameters alone, and that its knowledge could be updated by changing the index instead of retraining the model.

That design choice is the whole reason "retrieval" belongs at the center of this vocabulary. The engine does not carry a ranked list of your pages in its head. It reaches out to an index, pulls back passages that look relevant to the specific question, and generates from them. Classic ranking orders results for a human to choose among. Retrieval selects raw material for a machine to write from. The two overlap in tactics, but the underlying operation is not the same, which is why importing SEO intuitions wholesale into AI search quietly misleads.

Retrieval: fetching candidates from a corpus

Retrieval is the step where the engine, given a question, pulls a set of candidate passages from an index. It is a fetch, not a verdict. The output is a shortlist of raw material, often a few dozen passages assembled into a query-specific working set, from which the answer will later be written.

This reframes what is actually being improved. In classic search the unit was a document competing for a position. In retrieval the unit is a passage: a self-contained chunk of text that either does or does not get pulled into the candidate set for a given question. A page can hold a passage that retrieves well for one phrasing of a question and poorly for another, because retrieval is matched to the meaning of the specific query, not to a fixed page-level rank. The practical implication is that "being retrievable" is decided at the level of clear, self-sufficient statements, not at the level of the whole page.

Embedding: turning meaning into a vector

Embedding is the mechanism that makes meaning-based retrieval possible. Text is converted into a numerical vector, a long list of numbers positioned so that passages with similar meaning sit near each other in that space. When a question arrives, it is embedded the same way, and the engine retrieves the passages whose vectors are closest to the question's vector. The founding RAG work called this a non-parametric dense-vector index: a store of these vectors that lives outside the model and is searched at query time.

The difference from keyword matching matters for anyone writing content. Embedding-based retrieval can surface a passage that never uses the exact words of the question, as long as it means the same thing. It can also fail to surface a passage stuffed with the right keywords if the passage does not actually address the question's meaning. "Embedding versus ranking" is the useful contrast here: ranking sorts a known set by signals, embedding decides what enters the set at all by proximity of meaning. Tuning for the second is about being unambiguous and on-topic in plain language, not about density of terms.

Reranking: the second, stricter pass

A first retrieval pass favors recall: it casts wide to avoid missing anything relevant, which means it also pulls in passages that are only loosely on point. Reranking is the corrective second pass. A more precise, and usually more expensive, model re-scores the retrieved candidates against the question and reorders them, so that the passages the answer is actually written from are the strongest of the set rather than merely the first ones fetched.

The survey literature that formalized RAG treats retrieval and this kind of post-retrieval refinement as separate, independently improvable stages. That separation is why reranking earns its own word. A passage can clear retrieval and still be demoted at rerank because a stricter model judged a rival passage more directly responsive. For content, the lesson is that surviving to the answer is a two-gate process: be retrievable enough to enter the set, then be the most directly responsive passage in it. Vague, hedged, or padded text tends to lose at the second gate even when it passes the first.

Grounding: anchoring the answer to its evidence

Grounding is the requirement that the generated answer actually rest on the retrieved passages rather than on the model's unverified recall. A well-grounded answer is one whose claims can be traced back to the evidence that was in front of the model. Grounding is the property that RAG was meant to improve in the first place, and it is also where the research is most sober about how far the problem is from solved.

The current state of the field is that grounding is a goal the systems approximate, not a guarantee they deliver. Work on grounded attribution documents that in the dominant production patterns the answer and its supposed evidence can drift apart, and a related attribution survey reports that over 95 percent of answers from the open-source models it tested contained at least one sentence with no supporting citation at all. Grounding, in other words, is a spectrum, and treating "the engine answered" as "the engine was grounded in my content" is an assumption the evidence does not support.

Citation: the credit line, and why it can mislead

Citation is the visible credit line: the named sources an engine attaches to its answer. It is the step buyers see and the one they most want to win, because a citation is the closest thing AI search offers to a link. It is also the step where precise vocabulary matters most, because a citation is not proof of grounding.

The faithfulness literature is direct about this. In common pipelines the citation is attached to the answer independently of the passages that actually produced it, which means a cited source is not reliably the source the model reasoned from, and a page far outside the top of classic results can still appear in the credit line. So "we got cited" is worth having but is not evidence that your content shaped the answer, and it is separately true that most AI-summary sessions never click any citation at all. This is why the responsible measurement target is presence and mention, named in the answer, rather than borrowed organic-traffic metrics that were built for a different channel.

What actually moves citation visibility

The one controlled result in this space is worth stating precisely. In a benchmark of roughly ten thousand queries across nine datasets, the content changes that most raised a source's visibility inside generated answers were adding citations to credible sources, including direct quotations, and replacing vague claims with specific statistics, together worth a 30 to 40 percent relative lift on the study's visibility metric, with citing authoritative sources the strongest single lever. That is a real, peer-reviewed finding about what makes a passage citable, and it is also the ceiling of what is proven: it measured visibility inside the tested engines, not booked customers.

Naive, Advanced, and Modular: the map of the whole pipeline

These five terms are not a loose list. The survey that consolidated the RAG literature organizes the field into three generations, Naive, Advanced, and Modular RAG, precisely so that the stages can be reasoned about separately. Naive RAG is the basic retrieve-then-generate loop. Advanced RAG adds pre-retrieval and post-retrieval steps, the query handling and reranking that sharpen what reaches the model. Modular RAG treats the whole thing as reconfigurable components that can be rearranged and swapped.

For a business the taxonomy is a diagnostic, not trivia. It says that when an answer engine fails to feature you, the failure has a location. You were not retrieved, or you were retrieved and reranked out, or you were used but not credited, or you were credited but the answer was not really grounded in you. A serious plan names which stage is failing before it prescribes a fix. Advice that cannot say where in the pipeline it intervenes is advice that cannot be checked.

Why this vocabulary changes how you measure visibility

The reason to hold these words apart is measurement. Classic rank and AI citation are now decoupled steps in different processes, so a single blended "visibility" number hides more than it reveals. Ranked and cited are two different outcomes: you can rank well and never be retrieved into an answer, and you can be cited from a page that ranks nowhere near the top. Reporting them as one figure is how oversold claims survive.

The entity idea underneath all of this is older than the current wave, which is part of why it is trustworthy rather than faddish. Google's Knowledge Graph launched in 2012 with 500 million entities and 3.5 billion facts under the banner of indexing things, not strings, and that entity substrate is still what answer systems reason over. The continuity is the point: retrieval-era visibility is a real, measurable discipline with a decade of foundations behind it, not a rebrand of ranking. Measuring it accurately means scoring the steps separately and refusing to promise a number the mechanism cannot deliver.

The evidence

Key findings, with their sources

Retrieval-augmented generation pairs a pretrained language model with a separate dense-vector index queried at inference time, and the founding study reports it produces "more specific, diverse and factual" language while letting the model update its knowledge without retraining.

established Lewis, P. et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks", arXiv:2005.11401, NeurIPS 2020.
The RAG literature has settled into a formal taxonomy, Naive, Advanced, and Modular RAG, that separates retrieval, augmentation, and generation as distinct stages a pipeline can improve independently.

established Gao, Y. et al., "Retrieval-Augmented Generation for Large Language Models: A Survey", arXiv:2312.10997, 2023 to 2024.
In a controlled benchmark of roughly 10,000 queries across nine datasets, adding cited sources, direct quotations, and specific statistics produced a 30 to 40 percent relative lift on a position-adjusted visibility metric, with citing authoritative sources the single strongest lever.

established Aggarwal, P. et al., "GEO: Generative Engine Optimization", arXiv:2311.09735, ACM SIGKDD 2024.
Citation faithfulness is unresolved: in the dominant production pipelines a cited source is not reliably the source the model reasoned from, and an attribution survey reports over 95 percent of answers from the open-source models tested contained at least one unattributed sentence.

established Grounded-attribution study, arXiv:2409.11242, 2024; "Attribution Techniques for Mitigating Hallucinated Information in RAG Systems: A Survey", arXiv:2601.19927, 2025 to 2026.
Structured entity representation predates generative AI: Google's Knowledge Graph launched in 2012 with 500 million entities and 3.5 billion facts under the thesis of indexing "things, not strings", and it is still the entity substrate answer engines reason over.

established Singhal, A., "Introducing the Knowledge Graph: things, not strings", Official Google Blog, 2012.
Being retrieved is not being read: in real Google sessions users clicked a traditional result in about 8 percent of searches with an AI summary present, versus 15 percent without, and clicked a link inside the summary only about 1 percent of the time.

established Pew Research Center, "Do people click on links in Google AI summaries?", 2025 (68,879 searches).

Calibration

What is proven, what is promising, what is unproven

Evidence tier	Tactics	What the evidence says
established	The RAG architecture (retrieval, embedding, generation), the Naive/Advanced/Modular taxonomy, the GEO content levers, the entity/knowledge-graph substrate, and AI-summary click suppression.	Lewis 2020 (NeurIPS); Gao et al. survey 2312.10997; Aggarwal et al. GEO 2311.09735 (KDD 2024); Google Knowledge Graph 2012; Pew Research 2025.
emerging	Which mitigation actually closes the citation-faithfulness gap, and the exact two-stage retrieve-then-cite mechanics of specific engines.	Grounded-attribution work arXiv:2409.11242 and the attribution survey establish the gap; the fixes are not yet settled, and engine-specific pipeline mechanics rest on secondary technical analysis pending primary-source verification.
contested	The precise historical lineage separating AEO (pre-LLM featured-snippet tuning) from GEO (LLM-era retrieval and citation).	Industry-consensus narrative rather than a peer-reviewed or standards-body account; write it with attributed hedging, not as settled fact.

Reference

Glossary

Retrieval-augmented generation (RAG): The architecture underneath most answer engines: a language model paired with a searchable index it queries at answer time, so it can generate from fetched passages instead of memory alone.
Retrieval: The step that fetches a shortlist of candidate passages from an index in response to a question. A selection of raw material, not a ranked verdict for a human to choose among.
Passage: A self-contained chunk of text that is the actual unit retrieved, rather than a whole page or document. The thing that does or does not get pulled into the candidate set.
Embedding: A numerical vector representing a piece of text so that similar meanings sit close together, letting the engine retrieve by meaning rather than exact keyword match.
Dense-vector index: The store of embeddings that lives outside the model and is searched at query time; the non-parametric memory in a RAG system.
Reranking: A stricter second pass that re-scores and reorders the retrieved candidates so the answer is written from the most directly responsive passages, not merely the first ones fetched.
Grounding: The property that an answer actually rests on its retrieved evidence rather than the model's unverified recall. A goal RAG approximates, not a guarantee it delivers.
Citation: The visible credit line naming an answer's sources. Because it can be attached independently of the passages that produced the answer, a citation is not proof of grounding.
Naive / Advanced / Modular RAG: The survey taxonomy of RAG pipelines: the basic retrieve-then-generate loop, the same loop with pre- and post-retrieval refinement, and a reconfigurable component-based design.
Entity: A canonical thing (a business, a person, a place) that search systems reason over, as distinct from the strings of text that mention it. The substrate the Knowledge Graph introduced in 2012.

Straight answers

Frequently asked questions

What is the difference between retrieval and ranking?

Ranking orders a known set of documents so a person can choose among them, the classic ten-blue-links model. Retrieval fetches candidate passages from an index so a machine can write an answer from them. They share tactics but are different operations, which is why AI-answer visibility is not the same job as ranking in classic search.

What does grounding mean in AI search?

Grounding is whether an answer actually rests on the evidence the engine retrieved, rather than on the model's unverified memory. It is the property retrieval-augmented generation was designed to improve, but the research shows it is a spectrum: answers and their cited evidence can drift apart, so "the engine answered" does not prove the answer was grounded in your content.

Is being retrieved the same as being cited?

No. Retrieval decides whether your passage enters the working set the answer is written from. Citation is the credit line the engine displays afterward, and in common pipelines it can be attached independently of the passages that produced the answer. You can be used without being cited, and cited without having shaped the answer, which is why they are measured as separate steps.

Does tuning for AI search replace SEO?

It does not replace it, and it is not a rebrand of it either. Classic rank and AI citation are now decoupled steps in different processes, so they need separate measurement. A business can rank well and never be retrieved into an answer, or be cited from a page that ranks nowhere near the top. Reporting them as one blended number hides the gap.

What is a reranker?

A reranker is the model that runs the stricter second pass over retrieved candidates. The first pass casts wide to avoid missing relevant passages; the reranker re-scores those candidates against the question and reorders them so the answer is written from the most directly responsive material. Surviving to the answer therefore means clearing two gates, not one.

Provenance

Sources

Lewis, P. et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks", arXiv:2005.11401, NeurIPS 2020 (established)arxiv.org
Gao, Y. et al., "Retrieval-Augmented Generation for Large Language Models: A Survey", arXiv:2312.10997, 2023 to 2024 (established)arxiv.org
Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., Deshpande, A., "GEO: Generative Engine Optimization", arXiv:2311.09735, ACM SIGKDD 2024 (established)arxiv.org
Grounded-attribution study, "Measuring and Enhancing Trustworthiness of LLMs in RAG through Grounded Attributions and Learning to Refuse", arXiv:2409.11242, 2024 (established finding of the faithfulness gap)arxiv.org
"Attribution Techniques for Mitigating Hallucinated Information in RAG Systems: A Survey", arXiv:2601.19927, 2025 to 2026 (emerging on which mitigation closes the gap)arxiv.org
Singhal, A., "Introducing the Knowledge Graph: things, not strings", Official Google Blog, 2012 (established)
Pew Research Center, "Do people click on links in Google AI summaries?", 2025 (established)pewresearch.org

Every figure above is attributed to a real, dated source and tagged with its evidence tier. Where a claim could not be verified to a primary source, it is not stated as fact.