Mark van Berkel is Co-Founder and CTO of Schema App, specializing in ontology and semantic data for search, AI and agentic experiences.
For two decades, search results hid our internal mess. A page could rank while the product name on it differed from the name in the help center or on the regional site. Ranking rewarded popularity, so incoherence went unnoticed.
AI systems don’t rank your page. They read your facts and reassemble them into an answer. Every contradiction becomes a public statement about your brand.
Despite these risks, organizations are rapidly funding AI initiatives before validating the data those initiatives depend on. A Gartner survey found that 63% of data-management leaders either lack, or are unsure if they have, the data-management practices required for AI. Gartner researchers also predict that organizations will abandon 60% of AI projects through 2026 when those projects lack AI-ready data. The Stanford 2026 AI Index likewise found that the share of respondents identifying AI inaccuracy as a “relevant risk rose from 60% to 74%” in one year.
Bad answers don’t throw errors. They simply get repeated.
From Ranking Pages To Synthesized Facts
You used to publish a page and a human read it. Now, a machine retrieves facts from your pages and other sources, synthesizes them and answers in a chat response that you’ll never see. That changes what controls brand representation: Increasingly, the synthesized answer, rather than the source page, shapes how users understand a brand. Pew Research Center found that users clicked a traditional result on 8% of visits when an AI summary appeared, compared with 15% without one. They clicked a source within the summary just 1% of the time.
Placement is still relevant, but it isn’t the whole game anymore. If a machine can’t form a consistent picture of who you are, what you offer and how those things relate, you don’t have only a ranking problem. You have an understanding problem.
The Contradictions Were Always There
This understanding problem rarely stems from negligence. It often comes from siloed teams responsible for various content, each with different owners, templates and definitions of the things they’re describing.
At a large financial-services firm, the same product carried different names across the core site, a regional site and a microsite on a platform marketing doesn’t control. The names were the least of it. The pages quoted different rates because a legacy property was still publishing terms the firm had retired. To a person, that’s one product described sloppily. To a machine reading prose, these are separate products because nothing the firm published anywhere said they were the same one. Assembled together, these became several products, at several rates, sold by several companies.
The model doesn’t have to hallucinate when two official versions of the same product already did that work. It’s just reconciling identities nobody else has. Then it answers with confidence. It won’t ask which version is true. It will pick.
I tend to see three common issues across large enterprises:
1. Coherence: The same fact is stated with several variations across their site. A machine has to pick one version, blend them or infer (sometimes incorrectly) to provide an answer.
2. Ownership: It’s unclear who on the team is accountable for the canonical version, so there’s no one to escalate to.
3. Interoperability: The relationship between product, business unit and parent company isn’t stated anywhere. A person infers it from a logo and a domain name. Inference is where brand risk lives.
What Machines Do With Your Brand’s Contradictions
Google’s Conflicts Benchmark shows that when retrieved sources disagree, models can fail to resolve the conflict the way a human would. Stale pages and errors all appear to be usable evidence for these systems to form a response. Google’s FACTS Search Evaluation found that leading models were factually accurate on roughly 60% to 84% of information-seeking queries. Being honest about limitations, it’s important to note that AI models also fail on their own. The AI Index reports hallucination rates from 22% to 94% across 26 models, and recent decoding research shows models can over-trust a retrieved passage and override a correct internal answer.
Cleaning up your own data won’t make an AI model omniscient. It can’t control every source the model retrieves or eliminate the model’s own failures. But it removes one class of error that you own: publishing multiple truths about a single entity on your site and hoping the machine picks the right one.
Four Questions To Ask Before Your Next AI Purchase
Data governance becomes operational when you break it down into four questions, each with an owner:
1. Who owns this fact about our brand?
2. Where does the canonical version live on our site?
3. Is it machine-readable or only human-readable?
4. How would we detect it drifting if the content or data doesn’t align?
Before buying another AI tool, define the entity once, name an owner, express it so a machine reads the same fact the same way twice, then monitor for content drift.
Key Takeaway
In an agent-mediated web, you don’t own your brand’s answer. You own its inputs. Ownership and interoperability are the only levers left because someone else owns the tool that provides the output. Stop treating AI search as a visibility campaign. Treat it like the governance audit you didn’t schedule.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?