Why does the knowledge graph matter to AI models?
Search engines have worked with entities and facts about them for years now - and the Entity-Attribute-Value triple is the basic structure of every knowledge graph. “Mortgage” is the entity, “interest rate” is its attribute, “7.2%” is the value. Only the complete set forms a fact that can be stored, cross-checked against another source and quoted.
In GEO this translates directly: the model builds its answer from facts, not from paragraphs. A page that names entities but never gives their attributes and values is, to the model, a text without content - summarisable, but with nothing to quote. Conversely, a dense and consistent graph makes the model recognize the page as a source of knowledge in its field.
What will you see in the report?
You get a fact matrix: every attribute with your value next to the number of Top 10 competitors that describe it. Gaps are flagged and ordered by how common they are elsewhere, and your differentiators are highlighted separately.
Plus an entity graph with types and relations between them. Relations are additionally verified algorithmically - we check whether two entities genuinely co-occur in sentences of your text, whether they appear together in SERP questions, and how many competitors connect the same pair. The relation graph itself does not affect the dimension score; it exists to show where knowledge is coherent and where it thins out into single mentions.
Separately, as a warning with no effect on the score, we surface contradictions: the same entity and attribute described with different values. For a model, one contradictory number can disqualify the whole page as a source.
Sample recommendations
Fragments of a report from an audit of an electronics store category page.
Problem: The “weight” attribute is named without a value, and for esports players it is one of the parameters that settles the purchase.
Some mice also let you adjust the weight with removable weights.
Give the ranges: “from 58 g in ultralight models up to 120 g with the full set of weights”.
Problem: The product card describes the box contents instead of the entity — without the full model name the facts have nothing to attach to.
Product type: mouse. Signal transmission: wired. In the box: mouse, quick start guide.
Open the card with the full commercial model name and hang the parameters off it as attributes with values.
How do we measure the knowledge graph?
The model extracts EAV triples from your content, and before scoring we strip template noise - entities such as Menu, Newsletter, Cookies or the footer are not knowledge about the topic and must not inflate the result. Every entity also receives a type (topic, problem, brand, product or service, tool, concept, person, organisation, location), so the graph shows what the page is genuinely about.
What one fact is made of
| Element | The question it answers |
|---|---|
| Entity | What are you talking about? - “mortgage”, “solar panel” |
| Attribute | Which property are you describing? - “interest rate”, “efficiency” |
| Value | What exactly do you say about it? - “7.2%”, “21.3%” |
In benchmark mode every attribute is classified by how often it appears among competitors
| Attribute class | What it means and where it belongs |
|---|---|
| Baseline | Most competitors have it - a mandatory element of the topic, deserves its own H2 |
| Rare | Present at individual competitors - material for an H3 or the FAQ |
| Differentiator | Only you have it - should be visible in the H1 or lead, not at the end of the page |
What raises and what lowers the score?
Raises
- Every attribute carries a value, not just a name
- Values are concrete: a number, unit, range or date
- Coverage of attributes the Top 10 treats as mandatory
- A differentiator - an attribute nobody else has - placed high on the page
- Facts spread across topical sections rather than crammed into one table
Lowers
- Entities named without attributes (“we offer mortgages” and nothing more)
- Attributes without values (“competitive interest rate” instead of “7.2%”)
- Missing an attribute that is standard among competitors
- Internally contradictory facts - two different prices for the same thing
- A graph built from template elements instead of substantive content
Frequently asked questions
Do I need structured data for a dense graph?
No. The graph is built from the content itself, not from schema markup. Structured data is useful for other reasons and a separate tool checks it, but it does not affect this dimension.
What about an attribute my offer does not have?
State it explicitly rather than skipping it. A missing feature described in one sentence is still a quotable fact, while silence leaves a gap against competitors. If the attribute does not apply to your product at all, its absence costs far less than omitting one that is standard for the category.
Do contradictory values lower the score?
No, we show them as a warning with no effect on the result. They are still worth fixing: generative systems compute a contradiction risk per document and skip the page as a source once it crosses a threshold.