GraphRAG: The End of Keywords, the Beginning of Connected Entities
Summarize this article with AI
A phone call that changes everything
A client calls me on a Friday. He runs a parts site, 12,000 products online. He says: « Stéphane, I applied your thematic silos, that worked well for 2 years. But now my traffic is plateauing. Worse, on some long-tail queries, I’m losing positions for no obvious reason. »
I pull up his Search Console. 8,500 organic clicks per month. Respectable, but not for a catalog that size. What jumps out isn’t the keywords declining. It’s the clusters. Entire sections of his catalog are never traversed by Google’s AI. Product pages that are indexed fine, but isolated from one another. No semantic connection between a water pump and a cylinder head gasket, for example. Google AI Overviews cite them rarely. Perplexity ignores them.
The problem wasn’t the content.
It was the architecture.
GraphRAG: AI doesn’t read your pages, it maps entities
GraphRAG (Graph-augmented Retrieval Augmented Generation) is a term popularized by Microsoft Research in 2024. According to Search Engine Land, it describes the shift from text-based search to entity-centric search.
Instead of matching a query to a document, the AI decomposes the question into nodes—entities—and finds the most useful subgraph of knowledge.
An example: a user asks « What wheels are compatible with a Tesla Model Y in 19 inches with a pressure sensor? » The AI doesn’t hunt for an exact page. It maps multiple entities: wheel, Tesla Model Y, 19 inches, TPMS sensor. It crosses their relationships. And it generates an answer by drawing from the graph.
Your pages are no longer destinations. They become sources of relationships.
That’s the pivot.
From keyword to entity: the e-commerce silo trap
Classical e-commerce SEO rests on the category page, the product page, and the article page. Each page is optimized for one main keyword. We cross-link. For years, it worked.
But AI? It has no use for silos. It wants a graph.
I observe one simple rule at my clients: a product page never talks enough about connected entities. It lists the name, the price, compatibility. But it omits the why—why this product is a component of that system, its functional relationship with other parts.
Result: Google doesn’t create a link in its Knowledge Graph between entities. Generative AIs ignore them. Your catalog becomes disconnected islands.
For my parts client, we identified 3,200 core entities. We structured them in an internal graph, made explicit on each page. We avoided keyword stuffing. We created relationship paragraphs.
Example: at the bottom of a « Water Pump » sheet, a clean block: « This pump cools the engine. It works with the thermostat (model X) and the serpentine belt (Y). For a complete belt kit, see our distribution set. »
Three entities linked.
One connection.
One relationship.
Voici le processus que nous avons appliqué avec le client pièces détachées. Chaque étape a été clé pour passer de 8 500 sessions à +820 % de trafic organique.
Les 3 étapes de transformation vers un Knowledge Graph produit
Du silo thématique au graphe d’entités connectées
820% organic sessions in 14 months: the detailed case
I guided a parts client who deployed 12 thematic silos in 2021. He gained +142% traffic in 8 months. Then a ceiling.
We switched to an entity-centric architecture. Three steps:
- Audit of the internal Knowledge Graph. 3,200 product entities, 180 families, 47 main brands.
- Relational rewrite. Each product sheet got a « Technical Connections » block, linking the entity to 3 or 4 neighboring entities. In total, 6,200 relationships made explicit.
- Long-tail catalog expansion. We created bridge pages—like « Complete distribution kit for Renault Clio 4 1.5 dCi 90hp »—that aggregate entities and link them.
Quantified results:
14 months later, the site goes from 8,500 organic clicks per month to 78,000 clicks. A jump of +820%.
AI Overviews trigger on 37% of tracked queries. Before the redesign, that rate was 4%.
And crucially, visits from generative search (GPT, Perplexity) rose from anecdotal to 12% of total traffic.
No extra ad budget.
No massive link building.
Just an architecture legible to AIs.
The framework that transforms a catalog into a Knowledge Graph
Since 2016, I’ve built thematic silos. The DOSE framework, taught by Guillaume Attias at BMO Academy, taught me to systematize graph creation.
But in 2026, I added an « Entities » layer to that framework.
Here’s how I do it for an e-commerce site.
1. Exhaustive entity list. These are entities, not keywords. An entity is a unique concept: a product, a brand, a technical feature. I isolated 3,200 entities for the parts site.
2. Relationship mapping. Is each entity connected to others on the site? By what type of relationship (compatibility, complément, alternative, variant)? If the relationship doesn’t exist in text, it doesn’t exist for the AI.
3. Creation of node pages. One page per entity isn’t enough. I create pages that embody RDF triples: (Entity A) -[relationship]- (Entity B). These bridge pages are anchoring points for LLMs.
4. Sync with structured data. I use schema.org/Product, but especially schema.org/WebPage with « about » and « mentions » to declare entities. JSON-LD markup is your graph declaration.
5. Continuous interconnection. The graph is never finished. Each new reference enriches the network. With every addition, mesh density grows, and the AI perceives it.
Why your product sheets no longer suffice
Another client of mine sells measuring instruments online. 4,500 products. Very detailed sheets, technical specs, comparison tables. He thought he had « rich » content.
Problem: the entities « accuracy », « measurement range », « sensor type » were never linked cross-site. An infrared thermometer wasn’t related to a hygrometer, even though both often equip the same industrial environments.
We introduced 1,200 relationship connection blocks. In 6 months, organic traffic grew +320%. Citations in ChatGPT’s generative answers were multiplied by 5.
What changed?
The AI grasped that the site wasn’t presenting isolated objects, but a system. It began recommending product sets for complete use cases. No need to be the cheapest. You’re the most connected.
Competitive advantage turns on AI readability of your catalog
In 2026, search engines aren’t text indexers anymore. They’re graph explorers. GraphRAG is their native operating mode.
So keep optimizing your title tags, your speed, your backlinks. Those are fundamentals. But the differentiator is your ability to provide a clear map of your product universe.
Some of my competitors will tell you to « create more content ». I’m telling you to create more intelligible links.
One final number: at a European appliance client, 47 strategic queries moved from top 10 to top 3 after deploying this architecture. 47. Not 100. Not 1,000. But on those queries, average order value is 3x higher. Because they cover product associations no one had thought to link.
And you?
A 47-minute live audit of your entity architecture
I take your site, your catalog, and show you live where the breaks are in your entity graph. No PDF report. Just the map of what Google doesn’t see, and how to connect it.
Book a strategic call — 45 minFrequently Asked Questions
Is GraphRAG just for big e-commerce catalogs?
Not at all. Once you have more than 50 products, your entities deserve a graph. A small niche site can gain tremendously by becoming the reference on a very tight knowledge network, without competing with giants.
Do I need to put everything in schema.org?
JSON-LD markup of type <strong>WebPage</strong> with <em>about</em> and <em>mentions</em> is very useful. But what really matters is that relationships are written in plain text. The AI uses both signals. Schema without text does nothing.
Isn’t this just a fad?
No. Microsoft, Google, OpenAI—all are embedding graph search components. It’s structural. It doesn’t replace keyword research, it transcends it. Those who ignore this layer have amorphous traffic, or worse, declining traffic.
How long before I see first results?
At my clients, first movements arrive in 2 to 3 months on long-tail queries. For strong impact on AI Overviews and generative engines, I count 6 to 12 months. Progress is steady.
Can I do this myself or do I need an expert?
Entity mapping takes rigor. If you master the concepts and the DOSE framework, you can start the work. But most of my clients prefer a live audit to save time and avoid architecture mistakes.
