"Relevance Engineering optimizes how content matches a query. Corpus Engineering optimizes the conditions that make matching possible in the first place."
~ Cody C. Jensen, CEO & Founder, Searchbloom
Corpus Engineering is the systems-level discipline of engineering a corpus for retrieval, semantic understanding, citation, ranking, and AI generation across modern search and language systems. In plain terms, it is the work of building a brand's whole body of content so AI systems can find it, understand it, and use it. The work covers 5 practices: designing, structuring, expanding, maintaining, and optimizing the corpus. 6 components set the scope: corpus accessibility, semantic structure, information gain, corpus expansion, retrieval optimization, and corpus maintenance. The whole corpus is the thing you work on, not 1 page. The work is engineering, not content production. And upkeep is part of the job from day 1. This article coins the term and defines it.
Search and AI systems no longer judge 1 page at a time, the way Google judged pages in 2010. They break content into passages, turn it into numbers called embeddings, and track the entities in it. AI Overviews, ChatGPT, Claude, Perplexity, and Gemini all pull from many sources at once to build an answer. The words SEO uses to describe the work have not caught up. This article names the discipline that does.
TL;DR
- Corpus Engineering is a rigorous way to organize SEO for AI. That fits Google's position that optimizing for AI answers is still SEO. The new part is the name and the frame around the parts. The parts themselves (embeddings, drift, information gain, entities, passage retrieval) come from years of search research. Nobody had put them into 1 named discipline before.
- 6 components. Corpus accessibility, semantic structure, information gain, corpus expansion, retrieval optimization, and corpus maintenance. Each one matches work people already do. Nobody had named them as 1 discipline before.
- Relevance Engineering, introduced by Michael King at iPullRank, is 1 of those 6 components. It is the retrieval optimization component. Credit to King for naming it and building it out. We differ on scope, not substance. To me, access, lifecycle, infrastructure, expansion, and upkeep are equal parts of the work. They are not steps that come before relevance work.
- 3 layers of corpus. Owned (the brand's website and content), extended (what other sites say about the brand), and reference (everything else AI systems pull from). Corpus Engineering works most directly on the owned layer, and its rules reach all 3.
- 2 paths, 1 corpus. A brand's corpus feeds the retrieval index AI systems search today and the training data the next model learns from. Corpus Engineering covers both. Graphite calls the training-side work parametric knowledge optimization.
- Sits beneath the MERIT Framework as the hands-on discipline. MERIT says what visibility takes. Corpus Engineering builds the corpus to meet it. Different unit of work. Different depth. Same goal.
- Does not replace SEO. Traditional SEO still matters. Roughly 60 to 70% of what earns organic visibility in classic search also earns visibility in AI answers. Google says the same: it states its generative AI features run on its core Search ranking and quality systems. Corpus Engineering covers the parts traditional SEO does not.
- Lifecycle comes first, not last. Content ages. Entities change. Models change how they read content with every upgrade. A corpus that AI systems favored in January can slip by October without a single edit. Upkeep is part of the discipline, not a phase that starts after the work is done.
Which Corpus?
When this article says "corpus," which one does it mean?
"Corpus" means 3 different things in AI search, and Searchbloom's own articles use the word more than one way. I'd keep them apart:
- The training corpus. What a model learned from during training, now stored in its weights as parametric knowledge.
- The retrieval index. What AI systems search at the moment someone asks a question.
- A brand's own corpus. The body of content a brand controls and influences, which is what Corpus Engineering works on.
A brand's corpus feeds both of the others: today's retrieval index serves it, and the next training run reads it. Beyond the Corpus uses "corpus" for a model's training data; this article uses it for the brand's body of content. In the 3-layer model below, the brand's corpus is the owned and extended layers, and the training corpus and retrieval index both sit inside the reference layer.
Why This Needs a Name Now
AI systems don't read the web 1 page at a time. They pull passages from many sources, match them by meaning and by the entities in them, and favor what is fresh. That is a different job than ranking a single page.
The words practitioners use lag behind the systems. Much of what people sell as "AI SEO," "AEO," or "GEO" today is keyword-and-page work with a new label. Google has made the same point, mythbusting common AEO/GEO claims and confirming its spam policies apply to generative AI responses. The shift is real. The names have not caught up.
A small number of practitioners have begun naming the work. Michael King at iPullRank introduced Relevance Engineering as the discipline for matching content to a query by meaning. That work is foundational. The AI search conversation would not be where it is without it.
But relevance work, as scoped, does not cover everything. Accessibility, lifecycle, the structure under the pages, expansion, and upkeep sit outside it. A team can do excellent relevance work and still publish into a broken corpus. AI crawlers might not be able to reach it. It might leave out the topics and entities it should cover. Or it might be slipping out of favor as models change. The relevance work can be excellent and the brand can still go unseen.
This article names the broader discipline: Corpus Engineering.
Credit Where Credit Is Due
Michael King saw the shift first in this industry and named 1 part of it. His work and the iPullRank team made embeddings, query fan-out, retrieval precision, and matching by meaning everyday SEO ideas. That work is real and foundational. Without it, this industry would still be arguing about keyword density and on-page checklists.
This article does not try to top Relevance Engineering. The terms and the hands-on methods King built stand on their own. Corpus Engineering builds on top of them.
The disagreement is on scope. Relevance Engineering centers on matching content to queries by meaning. Lifecycle, infrastructure, and expansion show up inside it, but not as equal, standing parts. My argument is that they are not steps that come first. They are equal parts of the same system. Treating them as side issues leaves blind spots.
Corpus Engineering covers the full scope. Relevance Engineering sits inside it as the retrieval optimization component. The 2 do not compete.
The Conceptual Chain
Corpus Engineering works across the full chain, from source content to AI answer:
Corpus > Embeddings > Vector Space > Retrieval > Generation
Corpus. The source content. A webpage, a website, a knowledge base, a slice of the public web, a training dataset, or any collection of material that AI systems pull from.
Embeddings. Numbers that capture what a passage means. An embedding model turns the corpus into them.
Vector Space. The map where those numbers live. Ideas, entities, and passages sit close together when their meanings are close.
Retrieval. Finding and picking the right content from the corpus for a question or task.
Generation. The answer an AI system builds from what it retrieved. AI Overviews, chatbot replies, and other AI answers all sit here.
Corpus Engineering improves the conditions along this whole chain. A second path runs through training, and the next section covers it. The corpus goes in. Visibility comes out. The discipline builds the path between them.
The Parametric Side of Corpus Engineering
Does Corpus Engineering only matter when a model searches?
No. The chain above is the retrieval path, and when I first defined Corpus Engineering, I framed the discipline mostly around it. A model reaches a brand by 2 paths. The librarian framing is the simplest way I know to explain them: memory is what the model absorbed in training, and the catalog is what it retrieves when someone asks.
The memory side now has a name, and the credit goes to Graphite. Gregory Druck, Ethan Smith, and Bevin Benson call it parametric knowledge optimization. It means shaping what lands in the training set, so a model describes and recommends a brand better. Their October 2026 data shows memory matters even when a model searches. A brand's visibility with search turned off predicted its visibility with search turned on. Across roughly 957,000 ChatGPT fan-out queries from prompts that named no brand, 59.6% contained a brand, so memory shapes which brands the model looks up.
So Corpus Engineering has a parametric side as well as a retrieval side, and the same brand corpus feeds both. What the next model remembers depends on training crawler access, a consistent description of the brand, and mentions on many distinct third-party sites. The retrieval work shapes what today's model finds. The difference is the clock: retrieval can move in weeks, but memory moves only when a lab releases a new model. What Is Parametric Knowledge Optimization? covers the training side in full.
Why Corpus, Not Document
What you treat as the unit of work matters. SEO grew up working on 1 page at a time: a URL, a piece of content. That habit gave us on-page optimization, keyword targeting, and link building.
AI retrieval does not treat the page as the unit. It sees passages, entities, embeddings, and how they connect. How a brand shows up across the whole web matters. Topic clusters, topical authority, and who cites whom shape what AI systems pull.
For any brand or topic, the corpus has 3 layers.
Owned Corpus
The brand's website, blog, knowledge base, and owned media. The brand controls this layer directly, and most SEO and content teams already work on it.
Extended Corpus
What other sites say about the brand: reviews on G2, Clutch, Capterra, and Trustpilot; media coverage; community posts on Reddit, Quora, and LinkedIn; syndication; content written with partners. Brands can influence this layer but cannot control it.
Reference Corpus
Everything else AI systems pull from when they answer questions about the brand or topic: training data, search indexes, pages fetched live, and the public web at large. No brand controls this layer. The owned and extended layers shape it over time.
Corpus Engineering works most directly on the owned corpus. Its rules reach the extended and reference layers through corpus expansion, information gain across the web, and upkeep over time. The discipline treats all three layers as interrelated parts of the same surface that AI systems retrieve from and train on.
The Six Components of Corpus Engineering
1. Corpus Accessibility
Making sure retrieval systems and training crawlers can reach, read, and process the corpus. That covers server-side rendering, crawlability, clean HTML, indexing, and content that splits cleanly into passages. It goes past classic technical SEO. It checks access for systems beyond Google, such as GPTBot, ClaudeBot, PerplexityBot, Google-Extended, and the tools that feed AI answers.
That list mixes 2 jobs. GPTBot, ClaudeBot, and CCBot collect training data. Google-Extended controls whether Google may use pages for Gemini training. Retrieval crawlers such as OAI-SearchBot, Claude-SearchBot, and PerplexityBot fetch pages when an AI system answers a question. Accessibility covers both. Blocking training crawlers drops a corpus out of the next model's memory, and blocking retrieval crawlers drops it out of today's answers. Allowing them in robots.txt is not enough if a firewall or bot-management rule blocks them anyway; crawler access covers each bot.
If AI crawlers cannot see a corpus, no retrieval system can pull it today and the next model cannot learn from it. Accessibility is the foundation everything else depends on.
2. Semantic Structure
Organizing content so the links between ideas are clear. That covers how entities relate, topic clusters, internal links, related topics placed near each other, hierarchy, and clean passage breaks. Corpus Engineering treats this as architecture, not editing. The structure decides whether retrieval can find a passage in the first place.
3. Information Gain
Making the corpus more original and more worth citing. Think original research, first-hand experience, synthesis, your own data, expert takes, and a fresh angle. Corpus Engineering measures information gain across the whole corpus, not page by page. The question is whether the corpus as a whole adds new, citation-worthy information that others don't have.
AI systems tend to surface and cite content that adds something new over content that repeats what exists, and the studies in my Information Gain in SEO article show it. Information gain is not optional. It is fast becoming the main thing retrieval looks for. That article goes deep; this section shows where it fits in the larger discipline.
4. Corpus Expansion
Widening the topics and coverage of the corpus across all 3 layers. Think supporting content, related entities, nearby topics, evidence, mentions on other sites, and outside confirmation. Corpus Engineering treats expansion as deliberately growing the brand's footprint across everything AI systems retrieve from and train on.
Most visibility engagements treat outside mentions and third-party content as marketing or PR. Corpus Engineering treats them as building the extended corpus. A separate article covers Corpus Coverage, the measure this component moves.
5. Retrieval Optimization
Improving how retrieval systems find, score, and pick content from the corpus. That covers passage structure, clear passages, answer formatting, precise meaning, and wording that retrieval picks up. This component overlaps most with Relevance Engineering as Michael King scoped it. Corpus Engineering treats it as 1 of 6 components, not the whole discipline.
6. Corpus Maintenance
Managing drift in meaning, facts, and entities over time. That covers freshness, fact updates, entities that change, shifting terms, corpus drift, vector drift, and semantic-relationship drift. Corpus Engineering treats upkeep as a full component from the start.
Content ages. Entities change. Models change how they read content with every upgrade. A corpus that AI systems favored in January can slip by October without a single edit, just because the embedding models changed. Managing the lifecycle is part of the discipline, not a phase that starts after the work is done.
What Is New, and What Is Not
The parts inside Corpus Engineering are not new. Each one comes from earlier work: search research, language processing, corpus linguistics, technical SEO, semantic SEO, or the AI search debate of the last 2 years.
- Corpus accessibility draws on technical SEO, crawler work, and making content machine-readable.
- Semantic structure draws on entity SEO, topical authority, and the work of practitioners like Koray Tugberk Gubur and Bill Slawski's patent-driven analysis.
- Information gain traces to Google's own patents, and newer industry studies back it up.
- Corpus expansion draws on link building, digital PR, and the AEO emphasis on third-party validation.
- Retrieval optimization draws on Michael King's Relevance Engineering and on research into dense retrieval, query fan-out, and passage scoring.
- Corpus maintenance draws on academic work on embedding drift, model evolution, and the old SEO habit of content audits.
What is new is the synthesis. Corpus Engineering names the whole discipline. It puts these 6 components in 1 named scope. It makes lifecycle a core part of the work. And it gives practitioners 1 frame for work that used to sit across half a dozen separate fields.
The synthesis-as-a-discipline is the coinage. The component mechanics are not.
That is the honest claim. Corpus Engineering rests on decades of search research and years of hands-on SEO work. What it adds is 1 named, structured discipline with clear deliverables, which did not exist as a single thing before.
Relationship to the MERIT Framework
The MERIT Framework is Searchbloom's strategic framework for visibility in AI search and AI answers. It has 5 pillars: Mentions, Evidence, Relevance, Inclusion, and Transformation.
MERIT says what visibility takes. Corpus Engineering is the hands-on discipline that builds the corpus to meet it.
Each Corpus Engineering component fits a MERIT pillar, and each pillar gets Corpus Engineering work. The 2 frameworks work at different levels. MERIT is the strategic playbook. Corpus Engineering is the engineering underneath it.
A MERIT engagement without Corpus Engineering gives you strategy with no engineering. The reverse is also true: Corpus Engineering without MERIT gives you engineering with no direction. Together you get both.
Relationship to Traditional SEO
Corpus Engineering does not replace SEO. It expands it.
Traditional SEO works on pages, keywords, backlinks, and rankings. That work still matters. Roughly 60 to 70% of what earns organic visibility in classic search also earns visibility in AI answers. Google says the same: it states its generative AI features run on its core Search ranking and quality systems. Strong traditional SEO tends to surface brands in LLMs over time.
Corpus Engineering covers what traditional SEO does not. It treats the corpus as a designed system and works on the structure under the pages. It plans for change across model upgrades, and it builds the extended corpus on other sites. Content has to be ready for retrieval, not just aimed at a keyword.
The simplest way to put it: traditional SEO makes a brand eligible for retrieval. Corpus Engineering decides whether AI systems retrieve it, cite it, remember it, and keep doing so as those systems become how people find information.
What Corpus Engineering Is Not
Corpus Engineering is not:
- A direct replacement for SEO.
- Just an "AI SEO" buzzword for repackaged keyword-and-page work.
- Proof that rankings, backlinks, or keywords no longer matter.
- Limited to AI search; the discipline applies to classic search too.
- Another name for Relevance Engineering or Semantic SEO.
- A claim to have invented retrieval, embeddings, or search itself.
Corpus Engineering is a named, systems-level discipline. It pulls 6 components of corpus work into 1 frame, with deliverables, a lifecycle, and a defined scope.
Questions and Answers
What is Corpus Engineering?
Corpus Engineering is the systems-level discipline of engineering a corpus for retrieval, semantic understanding, citation, ranking, and AI generation across modern search and language systems. The work covers 5 practices: designing, structuring, expanding, maintaining, and optimizing the corpus. 6 components set the scope: corpus accessibility, semantic structure, information gain, corpus expansion, retrieval optimization, and corpus maintenance.
Who coined Corpus Engineering?
Cody C. Jensen, CEO & Founder of Searchbloom, coined and defined Corpus Engineering as a named systems-level discipline. The new part is the name and the frame. The parts underneath are older. Embeddings, drift, information gain, entities, and passage retrieval come from search research, language processing, corpus linguistics, and nearby SEO fields. What it adds is 1 named, structured discipline with clear deliverables, which did not exist as a single thing before.
How is Corpus Engineering different from Relevance Engineering?
Relevance Engineering, introduced by Michael King at iPullRank, is a discipline for matching content to a query by meaning. Corpus Engineering is the broader discipline, and Relevance Engineering is 1 of its 6 components. Relevance Engineering improves how content matches a query. Corpus Engineering improves what makes that match possible in the first place: access, structure, new information, wider coverage, retrieval, and upkeep.
Is Corpus Engineering replacing SEO?
No. Traditional SEO still matters. Roughly 60 to 70% of what earns organic visibility in classic search also earns visibility in AI answers. Corpus Engineering covers the parts SEO does not. It treats the corpus as a designed system, works on the structure under the pages, plans for model upgrades, and builds the extended corpus on other sites.
What is a corpus in this context?
A corpus is the source content AI systems pull from. For any brand it has 3 layers. The owned corpus is the brand's website and content. Next comes the extended corpus: what review sites, media, and communities say about the brand. Everything else AI systems pull from is the reference corpus. Corpus Engineering works most directly on the owned corpus, and its rules reach all 3 layers. The training corpus a model learned from and the retrieval index it searches both sit in the reference layer; the brand's own corpus feeds both.
Does Corpus Engineering cover training data or only retrieval?
Both. A brand's corpus feeds the retrieval index that AI systems search today and the training data the next model learns from. Graphite calls the training-side work parametric knowledge optimization. What a model remembers depends on training crawler access, a consistent brand description, and mentions on many distinct sites. Retrieval work shapes what it finds when it searches.
What are the six components of Corpus Engineering?
6 components define the discipline. Corpus accessibility makes sure AI systems can reach and read the corpus. Semantic structure organizes content by entity and topic. Information gain makes it more original and worth citing. Corpus expansion widens its topics and coverage. Retrieval optimization improves how systems find and pick content. It overlaps with Relevance Engineering. Corpus maintenance manages drift over time.
How does Corpus Engineering relate to the MERIT Framework?
The MERIT Framework is the strategic framework for AI visibility, with 5 pillars: Mentions, Evidence, Relevance, Inclusion, and Transformation. Corpus Engineering is the hands-on discipline beneath it. MERIT says what visibility takes, and Corpus Engineering builds the corpus to meet it. Each Corpus Engineering component fits a MERIT pillar.
Why does the corpus need lifecycle management?
Content ages. Entities change. Models change how they read content with every upgrade. A corpus that AI systems favored in January can slip by October without a single edit, just because the embedding models changed. Managing the lifecycle is part of the discipline from the start. Upkeep, drift checks, and regular health scoring are core parts of the work.
What is corpus drift, and what is vector drift?
Corpus drift is the change in the content around you over time: new pages appear, entities change, terms shift, and coverage grows or shrinks. Vector drift is the change in embeddings when a new model reads the same content and places it somewhere else on the map. Both matter, and corpus maintenance covers both. A corpus that has not changed at all can still slip in retrieval as embedding models update.
Where can I read more about the discipline?
This article coins and defines Corpus Engineering. Later articles in the series go deeper. 1 compares it to Relevance Engineering. Another covers what a Corpus Engineering engagement delivers. A third shows how it fits with the MERIT Framework. A fourth argues why the industry needs it now.
The Bottom Line
Corpus Engineering is the systems-level discipline of engineering a corpus for retrieval, semantic understanding, citation, ranking, and AI generation across modern search and language systems. The work covers 5 practices: designing, structuring, expanding, maintaining, and optimizing the corpus.
6 components set the scope: corpus accessibility, semantic structure, information gain, corpus expansion, retrieval optimization, and corpus maintenance.
The whole corpus is the unit of work. Engineering is the practice. Lifecycle is part of the job from day 1.
The brand's corpus feeds 2 paths: what a model retrieves today and what the next model remembers.
Michael King's Relevance Engineering sits inside it as the retrieval optimization component, with the other 5 around it.
Beneath the MERIT Framework, Corpus Engineering is the hands-on discipline that builds what MERIT defines.
It doesn't replace traditional SEO. It expands it.
Corpus Engineering is the discipline. This article names it.
