You turn over eight million, you've been in the district for thirty years, and yet when a buyer asks ChatGPT who the reliable suppliers in your category in Lombardy are, you're not there. The problem isn't that the AI doesn't know you exist: it's that it doesn't know your relationships — the groups you work with, the district you belong to, the trade associations. Without declared connections, you're an isolated node, and an isolated node doesn't get recommended. Declaring those relationships the right way is worth more than a hundred articles published in solitude.
Imagine running a Brescia-based company that makes precision tools. A buyer asks ChatGPT: “reliable suppliers of precision tooling in Lombardy for the automotive sector?”. The AI answers with five names. You’re not there, even though you turn over eight million and have been on the market for thirty years.
The problem isn’t necessarily that the AI doesn’t know you exist. The problem is that it hasn’t mapped your relationships: it doesn’t know which large groups you work with, it doesn’t know you’re part of the Brescia mechanical district, it doesn’t know that your co-founder sits on the board of a trade association. Without these connections, the AI has no reason to put you in that specific answer.
In this series on entities I’ve already explained how AI recognizes who you are. Today I’ll walk you through the next step: how it recognizes who you’re connected to and why this determines the contexts in which you get cited.
What mapping relationships between entities means, for an AI model
An AI model doesn’t learn entities as isolated words in a dictionary. It learns them as nodes in a graph, where each node is connected to others by labeled relationships: “founded by”, “supplier of”, “partner of”, “competitor of”, “based in”.
In the research world, this construction process is described explicitly. Zhu et al. (2023), in a paper on Large Language Models applied to Knowledge Graphs, write:
“Constructing KGs typically involves multiple tasks such as Named Entity Recognition (NER), Relation Extraction (RE), Event Extraction (EE), and Entity Linking (EL).”
Translated: building a knowledge graph requires four distinct steps. First recognizing the entities (NER), then extracting their relationships (RE), then linking events, and finally tying each entity to its unique identity in the world (Entity Linking).
The practical consequence is that your brand, to the AI, isn’t a point but a web. If the threads are missing, you exist but stay isolated: the AI knows you’re there, but it doesn’t know where to place you. And when a user asks “suppliers in sector X”, your isolated node doesn’t surface.
Why relationships come before everything else
In previous articles I explained how AI turns text into vectors in a semantic space and how it comes to recognize an author as an entity. Mapping relationships is the layer above: it’s not just about who you are, but about the position you occupy in the network of all the other entities in your sector.
The same authors say it clearly:
“Entity and Relation Extraction, along with Event Extraction, are pivotal for Knowledge Graph (KG) construction tasks [25–28].”
Translated: extracting entities together with the relationships that bind them is the core of building a knowledge graph. It’s not an optional detail, it’s the mechanism.
The result is that your visibility in AI answers depends not only on how many times you’re named, but on how many clear, repeated connections there are between you and the other relevant entities in your sector. A company cited 100 times but always in isolation performs worse than one cited 30 times, but always within a coherent network of relationships.
A company cited 100 times but always in isolation performs worse than one cited 30 times, but always within a coherent network of relationships.
The test you can run in 10 minutes
Here’s a simple check to figure out whether your relationships are readable to an AI. It’s not a professional audit, it’s a first honest look.
Open Google’s Rich Results Test and paste in your homepage URL. In the response, look for the “Organization” block. If it’s there, check whether it contains fields like `founder`, `parentOrganization`, `sameAs`, `brand`, `subOrganization`. If the block is missing or contains only name and logo, your brand’s relationships aren’t declared anywhere in machine-readable form.
Second step: go to Wikidata and search for your company name. If you don’t exist as an entry, no external system can tie your brand to a stable identifier. If you exist but the record is two lines long, your network of public relationships is minimal.
Third step: open ChatGPT or Perplexity and ask “who is the founder of [your brand]?” and “who are the main competitors of [your brand]?”. If the AI can’t answer, it can’t answer the other way around either when someone asks “who are the suppliers in sector X”.
Decision thresholds: an Organization schema present with at least founder + sameAs = an acceptable baseline. A Wikidata record with at least 5 populated properties = you’ve entered the public graph. A correct AI answer on founder and sector = the AI has internalized the key relationships.
Rewrite the About page as readable text, with the proper names of key people, the founding year written out in full, the city of the headquarters, and 3-5 exemplary client companies cited by name (with their permission)
The comparison I ran: two twin companies, opposite results
Instead of giving you average figures across a large sample, I’d rather tell you about a direct comparison that gets the idea across better than any percentage.
Two companies in the Brescia component industry, same size (around 8 million in revenue), same sector (precision tooling), same number of employees (35-40). One — let’s call it Company A — has a rich LinkedIn network: the two partners cited and connected to well-known partners/suppliers in the sector, case studies published on industry sites with the names of the main clients, mentions in B2B magazines with explicit roles and relationships. The other — Company B — has a clean website but zero explicit relationships published: no named partners, no cited clients, no linked trade associations.
I tried 20 B2B queries on ChatGPT and Perplexity such as “reliable suppliers of precision tools in Lombardy”, “who works with [a large client group common to both] in the tooling sector”, “Italian companies specialized in high-precision milling cutters for automotive”. Company A was cited in 11 answers out of 20 (55%), Company B in 0 (zero). Same revenue, same product quality according to B2B reviews, same seniority. The difference is entirely in the network of declared relationships.
The honest limit of the test: 20 queries are an indicative sample, not a study. I didn’t check whether ChatGPT pulled one of the two from a specific forum I didn’t investigate. Serious analysis requires professional AI search monitoring tools, not two hours of manual prompting.
The lesson: when the AI has to choose who to cite between two options equivalent in product and size, it picks the one that’s readable as a node in a network. The other, to the model, simply doesn’t exist in the context of the question.
The mistakes I see most often
Organization schema emptied of relationships. Many companies have the schema, but inside there’s only `name` and `url`. Without `founder`, `sameAs`, `parentOrganization` that block declares no relationship. It’s a signature with no content.
About page that tells stories, not entities. The “About us” page of Studio Associato Rossi says “we’re a close-knit team with twenty years of experience”. Zero names, zero dates, zero exemplary client companies cited by name. The AI extracts zero relationships.
A site with no links to social profiles and Wikidata. The `sameAs` field in the schema exists precisely to tell the AI “this entity is also the one on LinkedIn, on Wikidata, on Crunchbase”. Without this bridge, every mention of your brand on the web stays disconnected from the others.
Relationships declared only in PDFs or images. Org charts, partnerships, clients: they often end up in PDF brochures or JPG graphics. It follows that even the most advanced entity and relation extraction systems, by their nature, process text: if your org chart is an untagged image, those relationships don’t exist for the model.
What to do concretely
- Add to the homepage’s Organization schema the fields `founder`, `sameAs` (links to LinkedIn, Wikipedia, Wikidata if they exist), `parentOrganization` if you’re part of a group, `brand` if you have multiple product lines
- Open or enrich the Wikidata record for your brand and your founder: add at least country, industry, founded, headquarters, founder (if the record is for the brand)
- Rewrite the About page as readable text, with the proper names of key people, the founding year written out in full, the city of the headquarters, 3-5 exemplary client companies cited by name (with their permission)
- Create a “Partners and certifications” page with outbound links to official entities: trade associations, certifying bodies, Politecnico di Milano if you have university collaborations
- Compare yourself with the 3-5 competitors the AI cites in your sector: if they have a populated Wikidata and you don’t, you know where to catch up
What you take away
Relationships between entities are the layer where it’s decided when the AI recommends you. It’s not enough for it to know who you are: it has to know who you’re connected to. The comparison between Company A and Company B tells you better than any average how much this layer weighs: same numbers, same product, opposite visibility.
The thread of your visibility in AI answers runs through here: the more clear relationships you declare in the right formats, the more contexts can trigger a citation of you. It’s not a magic factor. A perfect network of relationships on a brand with no authoritative content produces no results: the principle of E-E-A-T applied to AI always holds. But without declared relationships, even the best content stays an isolated node.
In the next articles in this series we’ll look at how to build a Wikidata record that actually works, how to manage consistency between brand mentions across different sources, and how to measure over time the growth of your node within the public knowledge graph.
