Exploring “GuidedRAG: Semantic Steering of Retrieval-Augmented Generation” and Why SEO/GEO Professionals Should Care (Hamsterdam Research)

By Ethan Lazuk

Welcome to a new edition of Hamsterdam Research! 🐹

This is where we look at recent AI research papers to learn just what the heck they’re talking about and explore their hypothetical implications for the future of search and SEO/GEO strategies.

This time, we’ll look at a research paper called, “GuidedRAG: Semantic Steering of Retrieval-Augmented Generation.”

It was published on July 10, 2026, and its authors are Matthijs Jansen op de Haar, Tobias Stähle, and Lorenzo Gatti.

Before we dive in, why should SEO/GEO professionals care about this research paper?

GuidedRAG shows that AI systems may retrieve better information by first narrowing the search space using explicit semantics, such as entities, categories, relationships, dates, or question types, rather than relying on broad similarity to the user’s prompt alone.

In the paper, semantic steering improved retrieval relevance, brought useful passages higher in the rankings, reduced irrelevant results, and made retrieval less sensitive to how differently users phrase the same intent.

The practical implication of GuidedRAG is that visibility in AI-generated answers may depend not only on writing content that sounds relevant, but also on making its entities, topics, relationships, and metadata clear enough for retrieval systems to correctly classify and select it.

Having said all of that, the paper evaluates a controlled knowledge base rather than the open web, so it offers a useful model for thinking about SEO/GEO rather than direct proof of how commercial search engines currently rank or cite webpages.

To start, we’ll review the paper’s abstract.

Here’s the abstract (with my highlights):

“In this work, we propose GuidedRAG, a novel extension to traditional Retrieval-Augmented Generation (RAG) that introduces a dedicated selection stage and semantic steering during retrieval. In contrast to current state-of-the-art RAG approaches, which depend on increasingly complex retrieval and knowledge structures, GuidedRAG constrains the knowledge base using semantics before retrieval, aligning the retrieval space with user intent while substantially reducing the search space. Our evaluation shows that GuidedRAG improves retrieval relevance by 14.0-15.8%, mitigates a 19.7-27.4% loss in retrieval precision, and reduces retrieval overhead by orders of magnitude. Moreover, relevant chunks are consistently retrieved earlier in the ranking process, while alignment with user intent improves by 31.8-36.8%. We further show that GuidedRAG achieves full coverage across 15 diverse RAG variants, demonstrating generalizability across the literature. Together, these findings establish semantic steering and selections as a powerful and generalizable paradigm for improving the current state-of-the-art in RAG.”

Next, let’s break down the key vocabulary terms from the research paper.

• Retrieval-augmented generation (RAG) is a process in which an AI system retrieves information from an external knowledge base and adds that information to the model’s prompt before generating an answer. RAG allows the response to use information beyond what was stored in the model during training.
• GuidedRAG, the framework proposed in the paper, extends the typical RAG process by adding a dedicated selection stage and using semantic information to guide multiple parts of the system.
• Semantic steering is the use of explicit meaning-based signals to direct retrieval. For instance, an AI system might be told that a question concerns the entity “Rome” and is a fact-based question. Those signals help the system determine where and how to search instead of relying entirely on the wording of the prompt.
• Selection stage is an additional step that GuidedRAG performs before retrieval. During this stage, the system uses semantic information to narrow the full knowledge base to the areas most likely to contain the answer.
• Selection space is the smaller subset of the knowledge base created during the selection stage. Then, retrieval and ranking are performed inside this constrained space rather than across the entire collection. This concept is particularly important for SEO/GEO professionals because it suggests that being correctly associated with the appropriate entities, topics, categories, and relationships may help content enter the right retrieval space.
• Chunk refers to a passage or section of a document that can be stored, retrieved, and added to an AI model’s context. RAG systems generally retrieve chunks rather than entire documents, making passage-level clarity and relevance potentially important for AI visibility.
• Top-k ranking is the ordered list of the highest-scoring chunks returned by the retrieval system; k represents the number of chunks selected. The paper finds that GuidedRAG tends to place relevant chunks earlier in the top-k ranking, allowing a system to use fewer passages while still finding useful information.
• Retrieval relevance is the degree to which the retrieved chunks actually contain information that helps answer the question. Higher retrieval relevance means less irrelevant context is passed to the language model, increasing the likelihood of an accurate and focused response.
• Retrieval precision is the proportion of retrieved content that is genuinely relevant. An AI system has poor precision when it returns superficially similar passages even though they do not answer the question. GuidedRAG attempts to improve precision by excluding semantically unrelated areas before ranking begins.
• Retrieval turbulence in the paper measures how much the retrieved results change when the same underlying question is phrased in different ways. Lower turbulence means that novice, intermediate, and expert formulations of a similar intent are more likely to retrieve consistent information. The authors found that semantic steering reduced this variation, suggesting that retrieval became less dependent on exact prompt wording.

Cool. Let’s take a deeper look at the research paper.

If you want to follow along, you can grab a PDF or the HTML version on arXiv.

The full paper has seven sections, including: 1) Introduction, 2) Background, 3) Framework, 4) Methodology, 5) Evaluation, 6) Discussion, and 7) Conclusion.

We’ll summarize each of them below.

1. Introduction

The paper starts with Figure 1:

This figure shows the five-stage GuidedRAG workflow, from configuring the system to generating an answer:

1. Configuration: The system defines the knowledge base, the available semantic information, and how much control the user has.
2. Selection: The user supplies a prompt and semantic guidance, such as an entity, category, relationship, date, or question type. The system uses those semantics to narrow the full knowledge base to a smaller, more relevant subset.
3. Retrieval: The system ranks and retrieves the most relevant chunks from that selected subset rather than searching the entire knowledge base.
4. Augmentation: The original prompt, retrieved chunks, and semantic information are combined into an expanded prompt.
5. Generation: The large language model uses that expanded prompt to produce its response.

The colored dotted lines show that semantics influence several stages, not just retrieval, while the prompt travels through the process and the final response returns to the user.

The main point of Figure 1 is that GuidedRAG tries to understand the relevant semantic area before ranking content, which can reduce irrelevant retrieval and better align the answer with the user’s intent.

The notable idea for SEO/GEO professionals is that clearly defined entities, topics, categories, and relationships may help content enter the correct retrieval space (keeping in mind the paper tests a controlled knowledge base rather than open-web search).

The authors start out the paper by explaining the benefits of RAG:

“The rapid advancement of Large Language Models (LLMs) has contributed significantly to natural language understanding and generation across a wide range of applications, including Information Retrieval (IR) Systems (Zhai, 2024; Joho and Jose, 2025; Zhu et al., 2025). However, despite their impressive capabilities, LLMs remain limited by their static training data and susceptibility to generating factually incorrect or outdated information (Lewis et al., 2020; Brown et al., 2020). Retrieval-Augmented Generation (RAG) seeks to mitigate these issues by augmenting prompts with external knowledge retrieved from a knowledge base at inference time, thereby improving factual accuracy and adherence to domain-specific knowledge (Lewis et al., 2020).”

However, “While RAG has demonstrated substantial benefits,” they write, “its effectiveness is fundamentally dependent on the quality and relevance of the retrieval process.”

“Traditional RAG (Lewis et al., 2020) systems often rely on broad semantic similarity searches over large document collections, which can introduce irrelevant context, increase computational costs, and reduce retrieval precision (Amiraz et al., 2025; Zhao et al., 2024). As knowledge bases continue to grow in size and complexity, these limitations become increasingly pronounced, negatively impacting both efficiency and response quality. Furthermore, approaches such as GraphRAG (Edge et al., 2025) rely heavily on generated entity annotations. As a result, inaccuracies or omissions in these annotations can introduce a semantic gap between the user’s intent and the knowledge retrieved (Volkanovska, 2025).”

Hence, the authors introduce GuidedRAG as a solution.

GuidedRAG “extends traditional RAG approaches by introducing a (i) selection stage and (ii) semantic steering during the canonical RAG process.”

“In the selection stage, a knowledge base (KB) is scoped to form a selection space using semantic information, such as categories or entities. This reduces retrieval overhead by narrowing the search scope, thereby making RAG systems more scalable. Furthermore, it improves retrieval relevance by grounding the search space in high-level semantics. As a result, retrieval focuses on semantically relevant portions of the corpus, reducing the inclusion of irrelevant context. Moreover, while standard approaches rely solely on a user prompt to steer retrieval, GuidedRAG improves user control by allowing pre-specified semantics to steer retrieval. This results in closer alignment between user intent and machine interpretation. In this work, we demonstrate how GuidedRAG can improve efficiency, accuracy, and precision versus traditional RAG approaches.”

In short, LLMs can improve factual accuracy by using RAG to access external knowledge, but traditional RAG systems often retrieve irrelevant information, become less efficient as knowledge bases grow, and may misalign retrieved content with user intent.

GuidedRAG addresses these problems by using semantic information to narrow the search space before retrieval, with the goal of improving relevance, efficiency, precision, and alignment between the user’s intent and the generated response.

2. Background

The authors first discuss the limitations of LLMs:

“Since knowledge is stored in pre-trained parameters, LLMs cannot directly access information unavailable during training (Lewis et al., 2020; Roberts et al., 2020). Furthermore, domain-specific knowledge may be underrepresented, resulting in inaccurate or incomplete responses (Lewis et al., 2020). This can additionally result in hallucinations, where LLMs make up facts or claims (Anh-Hoang et al., 2025). These limitations motivate the use of external knowledge sources during inference, thereby grounding knowledge and further augmenting the context window and understanding of LLMs (Lewis et al., 2020).”

Next they talk about RAG:

“Retrieval-Augmented Generation (RAG) (Lewis et al., 2020) addresses the limitations of static knowledge by retrieving information from an external KB and incorporating it alongside a user-generated query provided to an LLM. As a result, generated responses can be grounded in information beyond the model’s pre-trained parameters (Lewis et al., 2020).”

They give a good explanation of how traditional RAG works, which I think is beneficial to SEO/GEO professionals:

“Traditional RAG systems rely on embedding models and dense retrieval. Documents, i.e., chunks of text, are transformed into embeddings and stored in a (vector) database, while user queries are converted into embeddings using the same embedding model according to the vector space model (Salton et al., 1975). Relevant chunks are then retrieved by comparing embedding similarity (Lewis et al., 2020).”

They also talk about evolutions on RAG, principally GraphRAG:

“In recent years, the definition of what constitutes a RAG system has grown to include alternative approaches such as GraphRAG (Edge et al., 2025). These approaches generally change one or multiple components in the canonical RAG process of retrieval, augmentation, and generation. In GraphRAG, a knowledge graph (KG) (Hogan et al., 2021) is used to provide relevant semantic context during retrieval to incorporate relational context during the generation stage.“

Next, the authors talk about limitations of RAG:

“Although RAG certainly improves access to external knowledge, retrieval quality often declines as the knowledge base grows (Izacard et al., 2023). Larger search spaces increase computational costs and can introduce irrelevant and misleading context into the retrieval process (Izacard et al., 2023). Furthermore, retrieval noise can cascade into the generation stage, leading to hallucinations (Anh-Hoang et al., 2025), incorrect or flawed responses, and increased token usage, all of which persist even with large context windows (Du et al., 2025; Lewis et al., 2020).”

In short, “The effectiveness of a RAG-based system is largely determined by the quality of its retrieval process.” But there’s an art to this: “Retrieving too much information can cause information to be diluted or the context window to overflow, while retrieving too little might cause relevant context to be lost during inference.”

This hints at the purpose of their research with GuidedRAG: “it is important to evaluate strategies that achieve higher retrieval relevance and that score relevant chunks more favorably. This will allow for the inclusion of fewer chunks during augmentation, reducing token usage and generation delays.”

They also discuss the selection stage, in particular:

“These requirements motivate the use of a selection stage in GuidedRAG; this stage narrows the KB to include semantically relevant regions during retrieval. This reduces retrieval overhead, as embedding similarity is compared over a smaller corpus, and improves response quality by considering only semantically relevant chunks during retrieval.”

Followed by a mention of the semantic steering process:

“Furthermore, by leveraging semantics, the canonical RAG process can be steered to employ retrieval or augmentation techniques tailored to the scenario, such as vector similarity (Salton et al., 1975) or maximal marginal relevance (Carbonell and Goldstein, 1998). The use of semantic steering and selections remains underexplored in existing RAG literature and the current state-of-the-art approaches.”

3. Framework

This section explains the GuidedRAG workflow, which adds semantic guidance before and during retrieval.

First off, the AI system defines the knowledge base, prompt, and semantic signals, such as an entity or question type.

Next, the system uses those signals to narrow the knowledge base to a relevant subset, ranks the best chunks within that smaller space, combines those chunks and semantics with the original prompt, and sends the expanded prompt to the language model for generation.

The main idea is that searching a smaller, meaningfully selected area can produce more relevant answers, reduce false positives, and use fewer tokens than searching the entire knowledge base.

Figure 2 in this section shows the different ways GuidedRAG can narrow a knowledge base before retrieval:

• Value selection: choose one specific item, such as a particular entity.
• Categorical selection: choose every item belonging to a category or type.
• Metric selection: choose items within a numerical range, such as dates from 2002 to 2026.
• Relation selection: choose items connected through a particular relationship.
• Manual selection: let the user or system choose a custom group of items.
• Full selection: use the entire knowledge base, which is essentially standard RAG without narrowing.

The shapes represent different values or entities, while the highlighted shapes show what remains in the selected subset.

The main point of Figure 2 is that GuidedRAG can use different kinds of semantic information to restrict retrieval to the most relevant part of the knowledge base instead of searching everything.

4. Methodology

This section explains how the researchers tested GuidedRAG against standard RAG and GraphRAG. They used the All Relations Lead to Rome (ARLtR) dataset, which contains information from 300 Wikipedia articles about the Roman Empire, along with entities, relationships, text chunks, 6,000 answerable questions, and 2,400 unanswerable questions. The questions vary by type, complexity, relationship focus, and user expertise, allowing the researchers to measure relevance, consistency, and false positives.

The authors compared standard RAG, GuidedRAG, GraphRAG, and a guided version of GraphRAG. While all systems use the same similarity-based retrieval method, GuidedRAG narrows the knowledge base using predefined entities or relationships before ranking content.

The evaluation focuses mainly on selection and retrieval, testing whether searching a smaller, semantically relevant area produces more precise and efficient results than searching the entire knowledge base or broadly expanding through a knowledge graph.

Figure 3 in this section compares how standard RAG, GuidedRAG, GraphRAG, and Guided GraphRAG move through the same stages — configuration, semantics, selection, retrieval, augmentation, and generation:

Standard RAG searches the entire vector-based knowledge base and retrieves top-ranked chunks using cosine similarity.

GuidedRAG first uses user-specified entities to create a smaller subset, then searches within it and adds both the retrieved chunks and semantic tags to the prompt.

GraphRAG uses an LLM to identify entities and expands outward through nearby connections using k-hops.

Guided GraphRAG uses predefined entities and specific relationship types to perform more targeted relation hops.

The main point of Figure 3 is that the guided versions narrow retrieval more deliberately, while the traditional versions search more broadly or depend more heavily on automated entity identification.

5. Evaluation

This section includes several charts showing the results. I’ll skip including them here and instead just give a summary of the results.

This section evaluates whether GuidedRAG improves entity identification, retrieval relevance, efficiency, consistency, and precision.

The authors first find that automated entity tagging is unreliable: cosine similarity correctly identifies only 26.7% of entities, while LLM-based and hybrid methods reach about 55%–57%. Pre-specified semantic tags avoid this problem entirely, which is one reason GuidedRAG performs better (although this assumes the supplied semantics are accurate).

On retrieval relevance, both guided approaches outperform their traditional counterparts. Standard RAG retrieves a relevant chunk for 59.7% of questions, GuidedRAG reaches 68.2%, GraphRAG reaches 66.1%, and Guided GraphRAG performs best at 76.6%.

Relevant chunks also appear earlier in the rankings, meaning the system can retrieve fewer chunks, reduce token use, and lower the risk that useful information gets buried in a long prompt. GuidedRAG also reduces the search space dramatically and produces more stable results when the same intent is phrased differently by novice, intermediate, or expert users.

The precision tests show another major weakness of traditional RAG: even for questions with no answer in the knowledge base, standard RAG and GraphRAG still retrieve false-positive chunks about 19%–28% of the time at the chosen similarity threshold. GuidedRAG can avoid this weakness by determining that the requested entity or semantic category is absent before retrieval begins.

Overall, the evaluation section suggests that narrowing the knowledge base with explicit semantics can improve relevance, speed, consistency, and precision (though the results come from a controlled Roman Empire dataset rather than open-web search).

6. Discussion

The authors first talk about retrieval relevance:

“We observed that GuidedRAG improves (i) retrieval relevance by considering semantically relevant chunks, and increasing recall by ranking these chunks more highly in the $top-k$ ranking. Consequently, RAG systems can generate more accurate responses across different domains and KBs. Furthermore, we have shown that GuidedRAG outperforms state-of-the-art approaches, such as GraphRAG ($+$15.8%), even when only extending base RAG ($+$2.1%). These gains enable more scalable systems without sacrificing accuracy or recall, as state-of-the-art approaches typically incur additional time and space complexity to achieve comparable or lower improvements.”

Next, they talk about retrieval efficiency:

“The results indicate a substantial improvement in (ii) retrieval efficiency for GuidedRAG, as embedding distances are calculated over smaller search spaces. This means that retrieval is faster in RAG systems and, consequently, that corpus size can scale substantially without slowing retrieval. The latter is explained by retrieval speed scaling with the average number of chunks per search space (e.g., per entity) rather than with the corpus size. This drastically improves scalability, facilitating larger systems that can answer a wider range of questions. This effectively gives LLMs more domain-specific knowledge, enabling them to provide better responses to users. Furthermore, as token usage is reduced through the smaller $top-k$ ranking or avoided as unanswerable questions can be mitigated prior to retrieval, this can make the operation of RAG systems considerably more affordable. The smaller $top-k$ ranking can also encourage the use of small open-source LLMs with reduced context windows, which is particularly useful for RAG systems that handle privacy-sensitive data, such as in the medical or financial domains.”

Thirdly, the authors talk about retrieval precision:

“The evaluation suggests that GuidedRAG avoids the loss of (iii) retrieval precision observed in traditional RAG approaches (19%–28%) when using unanswerable questions. A striking result is that the similarity threshold underlines a precision-recall dilemma, inherent to traditional RAG approaches. This dilemma is circumvented through semantic steering by identifying and excluding non-existent semantics prior to retrieval, facilitating lower threshold values, and improving recall. This may lead to new use cases for RAG systems, as designers are less constrained by the precision-recall dilemma. In addition, through semantic steering, it would be possible to use alternative retrieval, augmentation, or generation methods to further improve performance. This would, for example, allow the use of MMR instead of cosine similarity for open questions.”

Lastly, the authors talk about the design of GuidedRAG:

“Another benefit of GuidedRAG is implicit in its design. Traditional RAG systems rely solely on user-generated prompts to steer IR. This means that steering in IR relies solely on variations in prompt formulation. Meanwhile, in GuidedRAG, a user can additionally steer IR by specifying semantics. This directly improves user control by giving users two steering methods. Furthermore, by reducing retrieval turbulence caused by different user formulations, GuidedRAG more closely aligns with user intent than traditional RAG approaches. This accommodates a wider range of user expertise levels. Nevertheless, there is a contrast: introducing semantic steering may require experience, since users need know-how to choose relevant semantics. Despite this gap, systems can guide users to select appropriate semantics through user interactions. As an example, a user may select a location on a map, which the system then uses as a semantic to apply GuidedRAG. In addition, while we evaluated entity selection, alternative selection methods (i.e., see Figure 2) can further increase GuidedRAG’s flexibility, making it easier to adapt to diverse applications while still using the same underlying KB.”

In terms of generalizability, the authors argue that GuidedRAG is broadly applicable because they successfully mapped 15 different RAG approaches onto its framework. GuidedRAG is not intended to replace those systems, but rather it provides a common structure that can extend them with semantic selection. When no narrowing is applied and the full knowledge base is selected, the system simply behaves like the original RAG approach.

But there are limitations to this work, as well.

GuidedRAG depends heavily on the availability and accuracy of semantic information, such as entities, categories, dates, and relationships. Fine-grained annotations can be expensive and noisy, especially when generated by an LLM. The evaluation also used only one controlled Roman Empire dataset and primarily tested cosine-similarity retrieval, so the reported improvements may not transfer directly to other domains, datasets, or retrieval pipelines.

The authors suggest that future research should test other semantic-selection methods, annotation strategies, retrieval and reranking techniques, and the effects of semantic steering during augmentation and answer generation, not just retrieval.

7. Conclusion

Let’s look at the authors’ conclusion in full:

“In this paper, we investigated the impact of introducing semantic steering and adding selections to traditional RAG approaches through GuidedRAG. Traditional approaches rely on overhead in space and time complexity to improve performance in the Retrieval, Augmentation, and Generation stages. Furthermore, said approaches perform poorly on precision and recall, as chunks with high similarity scores are naively added to the context window, even when they lack semantic relevance to the underlying question, thereby bloating retrieval. GuidedRAG mitigates these effects drastically by introducing a Selection stage that makes retrieval and generation far more efficient, precise, and accurate. In our evaluation, we determined that computational overhead across the entire canonical RAG process can be reduced by leveraging semantics, avoiding automated entity tagging, and reducing the search space by orders of magnitude. Moreover, compared to state-of-the-art RAG approaches, a noticeable improvement in retrieval relevance is observed due to the removal of semantically irrelevant chunks from the search space. In addition, GuidedRAG provides a solution to the precision-recall dilemma by simultaneously avoiding loss of precision while improving recall. Overall, these benefits allow for more scalable systems. Besides performance metrics, GuidedRAG offers a way to increase user control, allowing users to steer the inference process using semantics while reducing turbulence. This makes GuidedRAG a powerful tool for aligning user intent with machine interpretation.”

So, why should SEO/GEO professionals care about “GuidedRAG: Semantic Steering of Retrieval-Augmented Generation”?

Based on everything we’ve learned, I’ve extracted several key takeaways based on my interpretation:

For retrieval-grounded AI answers, visibility starts with retrieval.

Before an AI system can mention, summarize, or cite content, it first has to retrieve the right passage. GuidedRAG focuses on improving that selection process.

Clear semantics may help content enter the right search space.

The paper uses signals such as entities, categories, relationships, dates, and question types to narrow the knowledge base before ranking content. For SEO/GEO, this reinforces the value of making a page’s subject, entities, attributes, and relationships unmistakable.

Passage-level relevance matters.

RAG systems retrieve chunks, not necessarily entire webpages. Content should therefore contain focused, self-contained sections that clearly answer specific questions without requiring excessive surrounding context.

Exact prompt wording may matter less when intent is well defined.

GuidedRAG produced more consistent retrieval when users phrased the same intent differently. This suggests that strong semantic alignment may be more durable than optimizing content around one exact query or prompt formulation.

More content is not always better.

The study found that narrowing retrieval to a smaller, relevant subset brought useful chunks higher in the rankings and reduced irrelevant context. For content creators, tightly focused coverage may be more useful than adding broad but weakly related information.

Entities and relationships deserve special attention.

The framework can select information based not only on an entity, but also on how entities are connected. This supports building content that clearly explains relationships between brands, products, people, locations, features, and concepts.

Better retrieval can reduce hallucinations and false positives.

GuidedRAG can sometimes determine that relevant information is absent instead of returning the closest available passage. Clear sourcing, factual specificity, and well-defined claims may therefore improve the usefulness of content as grounding material.

Note: The study used a controlled Roman Empire knowledge base, not the open web. Its value for SEO/GEO is as a model for understanding retrieval behavior, not as evidence that adding certain markup or wording will directly produce citations.

Outro

I hope you’ve enjoyed this edition of Hamsterdam Research! 🐹

Feel free to comment below or contact me with your feedback.

Stay tuned for another new article, hopefully next week, or check out related research posts below.

Until next time, enjoy the vibes:

Thanks for reading. Happy optimizing!