Most content strategy still treats the page as the atomic unit: write it, structure it, optimize it, ship it. Olaf Kopp's July 16 piece on Graph RAG and the Open Knowledge Format argues that assumption is already out of date for a growing share of retrieval. His claim: as engines move from matching whole pages by vector similarity to walking a knowledge graph of connected facts, the thing worth optimizing stops being the page and becomes the individual, decomposable claim inside it, what he calls an atomic knowledge unit, built to be picked up cleanly by an agent, not just read by a person scrolling a browser.
What Olaf Kopp actually argued
Kopp's framing, published on his site on July 16, 2026, is specific: Graph RAG and the Open Knowledge Format (OKF) push the unit of optimization down from the page to the atomic knowledge unit, a single fact, claim, or relationship that can stand on its own and connect to other facts in a graph. His argument is that a system retrieving by walking connections between concepts doesn't need, and often doesn't use, the whole page a fact came from. It needs the fact, cleanly separable from the sentence around it, with enough structure to know how that fact relates to other facts the system already holds.
That's a real shift in emphasis from most GEO advice circulating this year, which has focused heavily on page-level signals: is the page crawlable, does it have clear headings, does it lead with a direct answer. Kopp's point is one level below that. Even a well-structured page is still, to a graph-based retrieval system, a container of multiple potential atomic units, some of them extractable cleanly and some of them tangled up in narrative prose that makes them hard to lift without the surrounding sentence.
Graph RAG in plain terms
Retrieval Augmented Generation, RAG, is the general technique behind most AI answer engines: instead of relying purely on what a model memorized during training, the system retrieves relevant documents at answer time and grounds its response in them. The version most GEO advice implicitly assumes is vector RAG: content gets converted into embeddings, numerical representations of meaning, and retrieval works by finding documents whose embeddings are closest to the query's embedding. It's a similarity match, page to query.
Graph RAG adds a structural layer on top. Instead of, or alongside, similarity search, it builds and walks a knowledge graph, nodes representing entities and facts, edges representing the relationships between them. A query about your product's integration with a specific platform isn't just matched against pages that mention both terms; the system can traverse from the entity representing your product to the entity representing the platform to the specific relationship connecting them, pulling in facts from multiple pages, or multiple sources entirely, along the way. It's closer to how a human researcher would actually chase down an answer, following connections, than to a single similarity lookup.
| RETRIEVAL APPROACH | WHAT IT MATCHES ON | WHAT CONTENT NEEDS TO PROVIDE |
|---|---|---|
| Vector RAG | Semantic similarity between query and page text | Clear, well-written, topically relevant pages |
| Graph RAG | Traversable relationships between entities and facts | Discrete, sourced facts with explicit, machine-legible relationships |
Why Graph RAG makes the page stop being the unit
The practical consequence, if Kopp's framing holds, is that a page can be well-optimized by every classic GEO signal, extractable structure, demonstrated authority, machine access, and still underperform in a graph-based retrieval system if its individual facts aren't separable from each other. A comparison page that states three product differences in one dense paragraph is harder for a graph to parse into three distinct, connected facts than the same three differences presented as three standalone statements, each with a clear subject, relationship, and object.
This isn't a wholesale replacement for page-level optimization. It's an additional layer underneath it. A page still needs to be crawlable, still needs authority, still benefits from comparison content's outsized share of citations. What changes is the internal construction of the page: whether the facts inside it are written as atomic, standalone units a graph can lift and connect, or as prose where the facts are true but structurally welded to their neighbors.
A walkthrough: turning one page into atomic units
Take a typical enterprise comparison page: 'Our platform integrates natively with Salesforce and HubSpot, supports SOC 2 Type II compliance, and offers a 99.9% uptime SLA, unlike most competitors in this category who require middleware for CRM sync.' That sentence is accurate, readable, and reasonably well-written. It's also four separate facts welded into one clause, none of them independently extractable without dragging the other three along.
The output isn't a different page, it's the same page with an additional structured layer: a table, a definition list, or a schema block that restates the page's key facts as individually addressable statements. Something Inc.'s own llms.txt implementation already does a version of this at the site level, a Key facts block engines can lift verbatim. Kopp's argument extends that same instinct down to the page and paragraph level.
Illustrative only: relative ease of extracting a single fact cleanly, by content structure (not a measured study, a directional framing of the argument above)
Where this fits with what already works
None of this contradicts the practices that have consistently shown up in real citation data this year. Comparison content still wins the largest share of citations because it's already organized around discrete, comparable claims, which happens to be exactly the structure Graph RAG rewards most. The overlap isn't a coincidence: content that state clear verdicts and differences in short, sourced statements was always easier for a human skimmer to use, and it turns out to be easier for a graph-based retrieval system too. Kopp's contribution is naming why, and giving teams a more precise target than 'write clearly': structure your facts so each one survives being lifted out of its sentence.
It also raises the bar for content marketing work generally. Writing well and structuring a page with headings was enough to be a strong candidate for vector-based retrieval. Graph RAG systems reward an additional layer of discipline: does each individual claim on the page stand up as an independent, sourced, subject-relationship-object statement, or does it only make sense in the context of the sentence around it. That's a genuinely new editorial checklist item, not just a restatement of 'be clear.'
Do this next
Pick your three highest-value comparison or product pages and audit them at the sentence level, not the page level. For each dense, multi-fact sentence, ask whether a system could lift any single claim out cleanly without the rest of the sentence. Where the answer is no, add a structured layer, a table row, a definition, a labeled statement, that restates the same fact atomically. You don't need to rewrite the page's prose. You need to give it a second, structured skeleton underneath the one your readers see.
This is early-stage practice, not a settled standard the way schema or llms.txt are. Kopp's OKF proposal is one framing among several competing for how atomic, graph-friendly content should actually be marked up, and it's worth watching which format, if any, gets broad adoption before over-investing in one specific spec. What's not speculative is the underlying shift: retrieval keeps moving toward the fact, not the page, as its unit of work. Structuring content that way now is a reasonable bet regardless of which specific format wins, in the same way publishing an llms.txt file was a reasonable bet before any single machine-readability standard had fully won out.
The teams likely to benefit first are the ones with the most facts worth connecting: enterprise vendors with multiple integrations, compliance certifications, and pricing tiers, where a competitor's page might state the same information but bury it in a single dense paragraph nobody, human or machine, can cleanly parse on the first pass. If a graph-based system has to choose between two equally accurate sources and one of them hands over its facts pre-separated, that's the source doing the retrieval system's work for it, and that's very likely the source that gets used.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Josh leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.