Explained! How your schema, rich results, and knowledge graph really work. In simple terms.

I've been doing entity work with a brand recently. A well-run premium streetwear, strong community, consistent drop model. A few months into the work I ran into something that I think illustrates the problem better than any abstract explanation I've managed to write.

Their schema was clean. Genuinely clean. The Rich Results Test returned zero errors and zero warnings. Organisation schema was present, Product schema was on every PDP with the correct fields populated, sameAs entries were there. By every technical measure the schema was correct.

And yet when I queried their brand in two separate AI shopping agents and asked about their product range and brand positioning, both returned a description that was noticeably off. The category was wrong. One of them described a founding story that was inaccurate by a few years. The language used to describe the aesthetic didn't match how the brand actually positions itself or how any credible editorial source had described them.

The schema said one thing. The agents believed something different.

This took me a while to fully understand, and I want to use this article to explain the mechanism. I think most brands doing schema work have an incomplete model of what schema actually does and doesn't do, and that gap is what causes this kind of problem.

Your schema is the accused’s witness statement

Schema is a self-declaration. It lives on your domain. You write it, you publish it, and machines read it. It tells Google and AI shopping agents what you say about yourself. Your brand name, your founding date, your product attributes, your legal identity, your website, your social profiles.

That's the extent of what it does. It's your witness statement to the court.

Here's the part that matters: witness statements are weighted according to the credibility of the witness. And Google assigns your own schema a specific and limited credibility precisely because you control it. You have an obvious incentive to describe yourself favourably and accurately. You might be doing exactly that. But Google doesn't have a way to verify that from your declaration alone. So your schema gets weighted as one input among several, and not the most trusted one.

This is the thing most brands don't fully understand when they do schema work.

Getting the schema correct is necessary. It is not sufficient on its own. The court has other witnesses, and if they say something different, the judge's notes will reflect that.

Google takes evidence from other parties

The Knowledge Graph is the judge's notes. It's the internal picture Google builds of your brand as an entity. Not a website. Not a keyword. A thing in the world with attributes, relationships, history, and a category. It's what Google believes about you, as opposed to what you've told it.

Google builds that picture by triangulating across multiple sources and weighting them differently. I want to be direct about something here: I don't know exactly how Google weights each source against the others, and I'd be sceptical of anyone claiming they do. What I can describe is the rough hierarchy that I observe across the entity audits I run, and what I've seen move the needle when something changes.

Wikipedia sits near the top of that hierarchy. Not because Wikipedia is always accurate. It isn’t. It represents independent editorial review that Google has been training on for a long time and has a reasonable basis for trusting. Many of the old sites were used as training data for LLMs.

The important nuance is that Wikipedia eligibility requires notability, which most fashion brands below about £50m in revenue don't meet.

Wikidata is different. Wikidata is the machine-readable equivalent of Wikipedia. A structured database of entity attributes that feeds directly into the Knowledge Graph. It doesn't require editorial notability. Any entity that can be defined with verifiable attributes can have a Wikidata entry. Brand name, founding date, country of operation, official website, industry, notable products. A brand with a Wikidata entry that's properly populated has the closest thing to a formal entity registration the open web currently offers, and most fashion brands haven't done it.

Third-party editorial coverage is next. When a credible publication describes your brand and places it in a category, attributes a founding story, describes what it's known for then that corroborates the entity signal. The graph is doing something close to what a careful journalist would do: cross-referencing sources and building confidence in the ones that agree with each other.

Your own structured data contributes, weighted lower than third-party sources for the reason I described. Your GS1 registration contributes. Your Google Business Profile, your social profiles, press database entries, third-party retail directories.

Each of these is a witness. Individually modest. Collectively significant if they're consistent.

Each witness is cross examined by Google to ensure the facts are correct.

That word, consistent, is the one I keep coming back to. When multiple independent sources agree on the same entity attributes, the Knowledge Graph's confidence in those attributes goes up. When they contradict each other, confidence goes down. The brand I mentioned at the start had contradictions across at least four different sources. The misclassification in the AI agents wasn't a schema problem. It was a Knowledge Graph confidence problem caused by conflicting testimony from witnesses Google trusted more than the brand's own declaration.

The Knowledge Panel is the judge’s verdict

Rich Results and the Knowledge Panel are the visible outputs of all this. They're the verdict, not the evidence.

Rich Results are the enhanced search appearances on individual product or page results. They include star ratings pulled from schema, FAQ drop downs, price and availability shown in the SERPs. These are generated when Google has decided to surface schema data visually on a specific search result. The Rich Results Test tells you whether your schema is eligible. It doesn't tell you whether Google is actually using it. They are the result of the Knowledge Panel.

The Knowledge Panel is different. It's the entity-level panel that appears when you search a brand name. The box on the right side of desktop search that summarises what Google believes it knows about the brand as an entity. It appears when Google has enough confidence in the entity to surface a public summary. If no panel appears, the entity signals aren't strong enough yet. If a panel appears with incorrect information, the Knowledge Graph has built its picture from sources that contradicted or incomplete.

The important distinction is that valid schema can generate zero Rich Results and still be doing useful work. Schema contributes to the Knowledge Graph even when it's not generating visible outputs. The absence of Rich Results isn't a sign the schema is failing. It's a sign Google hasn't decided to surface that data visually yet.

These are different things and conflating them leads to the wrong diagnostic.

Why the three layers don't always align

Here's the failure mode I've seen repeatedly that this framework explains cleanly.

A brand fixes their schema. They clean up sameAs, add a Wikidata reference, correct the founding date across their own structured data. The Rich Results Test returns clean. They check the Knowledge Panel a few days later and it still shows incorrect information.

This is the gap between submitting evidence and the verdict being updated. Google re-crawls and reconciles the Knowledge Graph on its own schedule. The new consistent signals you've introduced will eventually shift the picture. But third-party sources that contradicted your schema don't disappear because you fixed your schema. The Flannels listing that says "founded 2014" when the correct date is 2018. The press database entry with an outdated description. The LeadIQ record that was scraped from an old about page and has never been updated. These are witnesses that have already given statements. You can't retract them. You can only build a stronger body of consistent evidence and wait for the judge's notes to catch up.

It’s a slow process. But it’s a process.

In the entity work I mentioned at the start, we found a few things that illustrated this well.

A Google Maps short link in the sameAs array. A redirect to a physical store location, not an entity identifier. That's not a sameAs entry that helps the Knowledge Graph understand the brand. It's a location redirect that contributes nothing and may actively confuse the entity resolution. A founding date inconsistency appearing across at least three third-party sources. Two of them disagreed with each other and both disagreed with Companies House. The schema was correct. The external record wasn't. The Knowledge Graph had to pick a version and it didn't necessarily pick the right one.

An unclaimed Google Business Profile. The "Own this business?" link was still visible, which means Google hadn't received a verified signal from the brand that they were the authoritative source for that entity's commercial presence. A small thing. Not small in what it signals to the graph about whether the entity is actively maintained.

And separately, a counterfeit operation with its own About page ranking in brand-adjacent searches. That's entity confusion made visible. A competing entity with overlapping signals, and the legitimate brand hadn't done enough to establish clear separation in the graph.

This is why it’s so important to include sameAs data to link to social media profiles and link this back to your website. The same with a Wikidata page. It helps Google verify your witness statement with the independent witness statement. 

None of these are technically complex problems. They're co-ordination problems. The sameAs audit lives between technical SEO and brand. The Wikidata entry doesn't belong to anyone obvious. The Google Business Profile claim is something the marketing team vaguely knows about and the development team assumes is handled. The founding date discrepancy is in sources nobody on the current team has ever looked at.

How this pulls together and what to do now

A few specific things that I'd prioritise based on what I've seen matter most, roughly in the order I'd tackle them.

The Wikidata entry. If your brand doesn't have one, create it. It's free. The process is documented on Wikidata. You need a brand name, a founding date you can verify, a country, an official website, an industry classification, and any notable products or lines.

Populate the statements properly. A label with no statements is a thin entry that contributes less than a well-populated one. The sameAs field in your Organisation schema should then reference the Wikidata entity URL, which connects your self-declared entity to the machine-readable registry.

This is all about consistency building agentic trust.

The sameAs audit. Pull your Organisation schema and extract every URL in the sameAs array. Open each one. Does it resolve? Does it point to an active, brand-owned profile that a machine could read as the same entity as the one on your domain? An Instagram handle from two years ago that now resolves to a different account is a contradiction, not a corroboration. A Facebook page that redirects to a general platform page is noise. Fix the broken ones and remove the ones that don't add verifiable identity signal.

Make Google’s job as easy as possible. Just like traditional technical SEO. Do the heavy lifting so Google uses as little processing power as possible.

The Knowledge Panel claim. Search your brand name. If a panel appears and you haven't verified it, do that. Google provides a verification process that establishes you as the authoritative source. It gives you the ability to flag incorrect information and signals to the graph that the entity is actively maintained by the brand. If no panel appears, the entity signals aren't strong enough yet. The Wikidata and sameAs work will help over time.

The founding date and basic facts consistency check. Search your brand name and note what different sources say about founding year, location, and category. Check Companies House. Check any press database entries you can find. Check your own About page and your schema. Anywhere these disagree is a contradiction the Knowledge Graph has had to adjudicate. Getting them consistent doesn't require contacting every third-party source. It requires making your own signals loud and consistent enough that they outweigh the contradictions over time.

I want to close with something that I think is easy to lose in the technical explanation.

The court case resolves. Google does update the Knowledge Graph as new consistent signals accumulate. The brands that are being described incorrectly by AI shopping agents today aren't in a permanent state. The verdict can change.

It changes on Google's schedule, not yours, and the compounding dynamic runs in both directions. A brand that builds consistent entity signals over time gets an increasingly coherent, increasingly confident picture. A brand that hasn't built those signals is letting the graph form its own conclusions from whatever witnesses happened to give evidence. Press database scrapers, directory entries, whatever a counterfeit operation's About page says, the Wikidata stub that someone created and then abandoned.

You need to own the Knowledge base.

The work needed to have been started a year ago. The second best time is now, and it's not technically complex. It’s just consistent and slightly tedious and keeps getting deprioritised because none of it shows up in the dashboards anyone looks at on a Tuesday morning.

Which is, of course, exactly why it's still worth doing.

Next article will explain why this is so important for agentic ecommerce.