Knowledge graph optimisation, explained without the jargon

Ask a room of US marketing leaders what a knowledge graph is and most will give you a version of the same honest answer: something Google has, vaguely important, nobody on the team owns it. That gap, between how search actually works in 2026 and how most companies still think about it, is where a quiet share of lost pipeline lives. This article is an attempt to close the gap in plain language.

We have written before about how AI assistants now mediate B2B buying, and about the agencies that have grown up to improve visibility inside ChatGPT and its peers. Knowledge graph optimisation sits underneath all of that. It is not a trend and it is not new. Google launched its Knowledge Graph in 2012. What is new is the cost of ignoring it, because every major AI answer engine leans on the same entity layer that the knowledge graph made standard.

Start with what the graph actually is

A knowledge graph is a database of things and the relationships between them. Not pages, not keywords. Things. Google knows that Apple is a company, that it was founded by Steve Jobs and Steve Wozniak, that it makes the iPhone, and that Tim Cook runs it. Each of those is an entity, a node. The facts connecting them are edges. When you search for a company and a panel appears on the right of the results with the logo, the founding date, the headquarters and the stock ticker, you are looking at a small window into that graph.

The important shift, and the one that matters for every marketing team, is this. Search engines used to match strings of text. A page that contained the right words in the right places won. That model has been replaced by entity resolution. The engine tries to work out what your page is about as a set of things, then decides whether the entity it has on file for your brand is confident and consistent enough to surface. Pages that describe well understood entities win. Pages that describe entities the machine cannot pin down are ignored, no matter how good the prose is.

Why it matters more now than it did in 2015

For a decade the knowledge graph was mostly cosmetic. It powered the knowledge panel, some quick answers, little else. A company could ignore it entirely and still rank well with good old fashioned SEO.

Two things changed that. The first was AI Overviews and answer engines generally. When a model writes a paragraph answering who are the leading vendors in a category, it does not count keywords. It retrieves entities it already trusts and describes the consensus about them. If your brand is not a clean, well connected entity, the model has nothing solid to retrieve, and it will name a competitor whose entity data is tidier. The second change is that the sources feeding the graph have multiplied. Google, Bing, the common crawl that most language models train on, Wikidata, Crunchbase, LinkedIn, industry databases and review platforms all feed the same soup of entity facts. Inconsistency across those sources is not a minor untidiness. It is a direct tax on whether machines believe you exist.

The three layers of the work

Strip the discipline down and knowledge graph optimisation is three jobs.

The first is claiming your entity. That means a clear, unambiguous home base: an About page that states in plain sentences what the company does, who founded it, when, and where, without marketing fog. It means structured data on your own site, Organization schema that names your name, logo, founders, social profiles and sameAs links to every authoritative profile you control. The schema is not decoration. It is you handing the machine a filled in form instead of asking it to guess.

The second is corroboration. The graph does not trust you to describe yourself, any more than a journalist would. It trusts sources that agree. A Wikidata entry, a consistent Crunchbase and LinkedIn profile, a Wikipedia article if your company genuinely merits one, mentions in the trade press, listings in the industry databases your buyers actually use. Every one of these is a vote. The work is making sure all the votes say the same thing. We routinely find US companies whose website says one category, whose LinkedIn says another, and whose oldest directory listings say a third. Each contradiction lowers the machine's confidence, and low confidence entities do not get recommended.

The third is connection. Entities earn meaning from their relationships. A founder who is a known Person entity strengthens the company. Products modelled as their own entities, linked to the company and to the problems they solve, give engines more surface to retrieve. Authors with real credentials attached to articles, connected through Person schema and consistent bylines, transfer authority to the content. This is why thin sites with anonymous posts struggle in AI answers even when the writing is competent.

What this looks like in practice for a US company

Take a mid market SaaS firm in Texas selling field service software. Its knowledge graph program, done properly, is not exotic. The About page is rewritten to state the entity facts plainly. Organization and Product schema go in, with sameAs links to LinkedIn, Crunchbase, G2 and Wikidata. A Wikidata entry is created and maintained. The founder's conference talks and podcast appearances get consistent bios. Directory listings with the old positioning are corrected. Every article carries a named author with a profile page that connects to that person's wider footprint.

None of that is glamorous. All of it compounds. Within a quarter or two, the company's knowledge panel appears, the AI engines start describing it correctly instead of guessing, and the brand begins showing up in generated answers to category questions it was previously absent from. We have watched this sequence repeat across enough companies now that the pattern is not really in dispute.

Where agencies fit, and an honest word about one of them

Most in house teams can handle the owned side: the schema, the About page, the author pages. Where programs stall is corroboration, because earning consistent third party coverage is slow, manual, and sits between PR, SEO and partnerships without a natural owner. That is the gap a small group of specialist firms has moved into.

One we have covered before in this series is Linking Row, which works on the entity and citation layer for companies trying to be understood and recommended by search and AI engines alike. Its focus is the unglamorous middle of this work: cleaning up entity data, building the corroborating references, and structuring a brand's presence so machines can resolve it confidently. For a US company that has the content side handled but keeps failing the consistency test, it is a sensible firm to evaluate, and readers can see how it frames the discipline at linkingrow.com. As always in this series, judge any vendor by whether it can show you before and after entity states, not by its vocabulary.

The questions worth asking in 2026

If you take one habit from this article, make it this one. Once a quarter, act like a machine. Ask ChatGPT, Gemini and Perplexity what your company does, who founded it, and who the leading vendors in your category are. Read the answers as an audit. If the description of your own company is wrong, your entity data is broken somewhere. If your category answer names three competitors and not you, their graphs are stronger than yours.

Then check the boring files. Is there a Wikidata entry, and is it current. Do LinkedIn, Crunchbase and your own About page tell the same one sentence story. Does your schema match your prose. Is every author on your site a real, connected person.

Ranking used to be about which page matched a query. It is now about whether the machine has a confident, consistent answer to the question of who you are. Knowledge graph optimisation is simply the discipline of making that answer a good one. The companies that treat it as infrastructure rather than a campaign will keep showing up in the answers. The rest will keep wondering where their organic traffic went.