The 80% Problem: Why Your Best Product Attributes Are Invisible to AI

There's a specific moment when the agent-readability problem becomes viscerally clear. You're looking at your product page. The structured data is there — schema.org Product type, name, price, description. The metadata is solid. The page loads fast. And then you ask ChatGPT to evaluate your product against a competitor and it gets your differentiators wrong. It attributes the wrong pricing tier. It misses the thing that makes you actually worth buying.

The problem isn't that the agent is bad. The problem is that the information it got wrong was in your marketing copy — a beautifully written paragraph about your founding story, a customer quote about the feature that changes everything, a case study embedded in a hero image with no alt text. Agents can parse structured attributes. They struggle with nuance buried in prose, and they can't read images.

The 20/80 split

Most product websites have approximately 20% of their product meaning in structured data — the fields in JSON-LD, the columns in a product feed, the attributes in an API response. The other 80% is what we might call tribal knowledge: it lives in marketing copy, in the heads of salespeople, in one-line callouts in a hero section, in a PDF that only a human would think to download. Pre-agent, this was fine. Humans are forgiving, generalist interpreters. We read between the lines. We infer. We follow up with a sales call when something is unclear. A lot of marketing exists specifically to paper over the gap between what a product actually is and what a customer needs to understand about it to make a purchase decision.

Why agents can't paper over the same gap

Agents are not forgiving. They are precise. When an agent evaluates 50 products against a specific set of constraints, it reads structured attributes, weighs them against the query, and returns a result. A product with incomplete structured data gets evaluated on what's available — which is often the wrong subset of what matters. Worse, if an agent reads your marketing copy and finds vague claims without verifiable structured backing ('industry-leading performance,' 'enterprise-grade security'), it treats those claims with lower confidence than a competitor who has the same claims backed by a structured attribute with a source.

Consider what happens when a buyer asks an agent to find a B2B SaaS tool that 'can scale to 10,000 customers.' If you have a blog post saying you support enterprise scale, the agent reads it with low confidence. If you have a structured attribute on your product schema with a verifiedByURL pointing to a technical case study, the agent can cite that claim with high confidence. These are not equivalent in agent-mediated recommendations — even though to a human reader both feel like 'evidence.'

Moving meaning from copy into data

The fix is not glamorous. It's a data project. You need to go through your product — its features, its differentiators, its proof points, its origin story if it's relevant — and ask: is this in a structured format that an agent can read and cite? Or is it in prose that an agent has to interpret? For every piece of meaning that lives only in copy, you have a decision: either move it into a structured attribute, or accept that agents won't reliably surface it.

For e-commerce, this might mean adding sourcing attributes to product schema (farm name, processing method, certifications). For SaaS, it might mean adding verifiable performance claims to your structured data with source links. For content businesses, it might mean ensuring that author credentials, publication dates, and topic categorizations are in JSON-LD, not just in the page layout. None of this is technically complex. All of it requires discipline and time.

The competitive advantage window

Most companies haven't done this work. The average site we scan scores below 50 on agent readability, and the structured data depth check — measuring how much of visible page content is represented in JSON-LD — is one of the most commonly failed. That means there's a window right now where doing this work creates a genuine competitive advantage: agents will more accurately evaluate and more confidently recommend companies that have clean, deep structured data over companies that have the same product in prose.

That window won't stay open indefinitely. As agent-mediated commerce grows, more companies will figure this out and close the gap. The question is whether you want to be the company that leads your category in agent readability when adoption reaches the mainstream — or the one that's scrambling to catch up after the recommendations have already been made.

Start with a scan

The first step is measuring where you are. Scan your site and get a structured data depth score. Then look at the check detail — which of your page content is and isn't represented in your JSON-LD. That gap is the 80% problem made visible. Work from highest-value differentiators first: the claims that, if an agent could cite them confidently, would most change whether you end up in the recommendation.