John Koblinsky loaded 28 interview transcripts into a qualitative AI research platform—transcripts from a messaging study he'd run himself, interviews he'd sat in on, findings he'd written up by hand. He asked the AI a simple question: which messages performed best with buyers?
The AI gave him a confident answer. It was wrong.
Koblinsky, who directs research and insights at SAP Concur, had a problem that won't surprise anyone who's worked with AI in the past year. The tool processed a subset of his data, produced some high-level thematic feedback, and returned a verdict—without flagging what it had actually read. He only caught the error because he'd also done the laborious work manually. Most stakeholders would have taken the answer at face value.
This experience led Koblinsky to develop what he calls Evidence Assurance—a framework for describing how much trust to place in AI-generated research. It uses three levels:
- Directional: Loosely supported. Good for exploration, not for decisions.
- Grounded: Tied to specific sources with moderate traceability.
- Auditable: Fully traceable to primary evidence. The receipts are available.
The framework solves a real problem. LLMs present every answer in the same confident voice—whether built on 200 interviews or invented on the spot. A statistic from a proper study carries margin of error and confidence intervals; an AI-generated insight carries neither. As Koblinsky puts it, "A finding that can't tell you how confident to be is little more than a gut feeling or, with AI, potentially a hallucination."
The Mirror Image
Here's what product marketers should notice: the evidence problem cuts both ways.
If you're using AI to synthesize competitive intelligence, you need Koblinsky's framework to evaluate the output. Is this ChatGPT response directional, grounded, or auditable? Can I trace this claim back to something real?
But there's a flip side that matters just as much—maybe more. When buyers use AI to research your product, the same evidence problem applies in reverse. Is YOUR content auditable? When Claude or Perplexity or a Copilot agent answers a question about your category, can it trace its response back to specific, verifiable claims from your site?
The AI doing the answering faces the same challenge as Koblinsky's qual research platform. It needs evidence. It needs sources it can cite. And when your content is vague—"industry-leading solution" instead of "reduced procurement costs 23% at Unilever"—the AI has nothing auditable to work with. It either hedges, skips you entirely, or (worst case) invents something confident-sounding that may or may not be true.
What "Auditable" Means for Content Strategy
Let's apply Koblinsky's framework to your own content:
Directional content is the marketing language that's been standard practice for decades. "Comprehensive solution." "Seamless integration." "Best-in-class performance." It's not wrong, exactly—it points in a direction. But AI can't do anything with it. There's no fact to cite, no proof to reference. When an AI encounters directional content, it either paraphrases it weakly or moves on to a competitor who gave it something concrete.
Grounded content includes specifics that can be partially verified. "Over 5,000 customers globally." "Named a Leader in the 2026 Gartner Magic Quadrant." "Integrates with 200+ enterprise systems." These are citable facts—AI can reference them. But they're still a layer removed from primary evidence. The customer number is a claim; the customer outcome is proof.
Auditable content gives AI—and humans—the receipts. "Accenture reduced procurement cycle times by 40% using SAP Ariba's source-to-pay automation." "Siemens onboarded 10,000 suppliers in 90 days." Named customer, specific outcome, verifiable timeline. This is what AI can cite with confidence, because there's a traceable chain from claim to evidence.
The Practical Test
If an AI cited your content in response to a buyer query, would the citation be embarrassing or impressive? Would it make your product look credible, or would it surface the vague marketing language you've been meaning to replace?
The Confidence Problem Reversed
Koblinsky's core insight—that AI presents everything in the same confident voice—has a corollary for content creators: AI treats all content with the same skepticism.
An LLM doesn't know that your "industry-leading platform" claim is backed by serious R&D and customer success. It sees a phrase with no evidence attached. Meanwhile, your competitor's page says "reduced accounts payable processing time from 14 days to 3 days for [Named Customer]." The AI has something concrete to cite from one source and nothing but adjectives from the other.
This isn't about keyword optimization or even traditional SEO. It's about evidence architecture. The structure of your content either supports AI citation or it doesn't.
Implications for PMMs
Koblinsky built his Evidence Assurance framework to help researchers evaluate AI output. But the framework translates directly into a content audit:
- Inventory your claims. List the key assertions on your product pages, feature descriptions, and competitive content.
- Apply the framework. Is each claim directional (adjectives and positioning), grounded (citable facts), or auditable (named outcomes with evidence)?
- Close the gaps. For every directional claim that matters, find the grounded fact. For every grounded fact, find the auditable proof. "Fast implementation" becomes "average deployment in 90 days" becomes "Siemens deployed in 90 days and onboarded 10,000 suppliers."
- Make evidence indexable. Customer outcomes buried in PDF case studies don't help AI cite you. The evidence needs to be on pages that crawlers can read and language models can parse.
The goal isn't to replace marketing with a spec sheet. It's to ensure that when AI encounters your content—and it will, increasingly—there's something auditable underneath the positioning.
The Shared Standard
Koblinsky notes that the market research industry spent 80 years developing methodological discipline—sample sizes, margins of error, confidence intervals—so that findings could be trusted. AI bypasses that discipline. The output looks the same whether it's built on evidence or fabricated from patterns.
Content marketing is facing a similar reckoning. We spent 20 years optimizing for Google—keywords, metadata, backlinks—and now the ground is shifting. AI doesn't care about your keyword density. It cares about whether you gave it something it can cite without looking stupid.
The evidence problem really does cut both ways. Evaluate AI's claims with the Evidence Assurance framework. And make sure your own content passes the same test.