The Structural Extractability Gap Costing You AI Citations

You typed your category into ChatGPT to see what came back. It named a competitor, summarized their pricing accurately, and described who they serve. Your page covers the same ground, arguably better. It wasn't mentioned. Nobody on your team can explain it, because by every measure you have, your page is the stronger one.

Agent extractability is whether an AI agent can pull the key facts off a page cleanly: pricing, what the product does, who it's for, proof it works. When two comparable pages get cited unequally, the cause is usually a structural extractability gap, which is the distance between what a human sees on the page and what a machine can actually retrieve from it.

That gap has a precedent older than the web.

What library cataloguing figured out in 1876

Charles Ammi Cutter published Rules for a Printed Dictionary Catalogue in 1876, and one of his concerns was what cataloguers now call authority control: making sure that one author, one subject, one entity resolves to exactly one authorized form in the catalogue. Mark Twain and Samuel Clemens are the same person, so they need one record with cross-references, not two orphaned ones.

The consequence is the interesting part. Two equally good books, sitting on adjacent shelves, surface at wildly different rates depending on whether their catalogue records resolve. A book whose author record is fragmented across three name variants is findable only by accident. A book whose record resolves cleanly is findable by anyone who asks for that author. The quality of the writing has nothing to do with it. You can verify this yourself in about five minutes: open any library's OPAC, search an author whose name has variant spellings, and watch the holdings split.

Transfer that to your page. An AI assistant deciding what to cite is doing something structurally similar to a catalogue lookup. It needs to resolve your page to a defined entity, extract retrievable facts about that entity, and find corroboration elsewhere that those facts hold. When your page fails any of those steps, the assistant reaches for a page that doesn't fail, and the one that doesn't fail may well be worse writing. Nothing about that process is visible to a human reading your page in a browser.

Why AI agents skip my page when it reads perfectly well

Because "reads well" and "extracts cleanly" are separate tests, and most teams only run the first one. A visual review tells you whether a person scrolling the page can understand it. It tells you nothing about whether the substance survives being parsed.

Here are the patterns we see most often, all of which pass a human review without a blink:

Pricing rendered client-side with no server-rendered fallback. The numbers are right there on screen. A person sees them instantly. But if the pricing table is hydrated by JavaScript after initial load and there's no text equivalent in the served HTML, a meaningful share of agent read paths get an empty container. The page appears to have no pricing information at all.

Comparison tables coded as styled divs. Visually it's a table: rows, columns, aligned headers, the whole thing. Structurally it's a stack of <div> elements with CSS grid doing the work. An agent parsing for tabular relationships finds no <table>, no <th>, no row-cell association. The comparison you spent two weeks building is, to a machine, a pile of disconnected phrases.

Proof points locked inside an infographic. Customer count, uptime figure, average result, three logos, all rendered into one PNG with no adjacent text equivalent and alt text reading "hero graphic." The strongest evidence on the page is, functionally, not on the page.

Features behind an untriggered tab. Your feature matrix is complete and accurate, but it lives under the third tab of a component that only populates on click. An agent that doesn't simulate that click sees a third of your capability.

Any one of these can be enough. Combined, they produce a page that a person rates highly and a machine reads as thin.

Why an AI-readable page ranks worse than a competitor with weaker copy

Two pages can be editorially equal and structurally unequal, and the structural inequality is the only one the agent experiences. This is where the competitive damage happens, because the gap compounds silently.

Your competitor's page may have been built by a team that used semantic HTML out of habit, shipped server-rendered content because their framework defaulted to it, and put their pricing in plain text because nobody thought to make it fancy. They didn't optimize for agents. They just happened not to break anything. Meanwhile your page went through three design iterations, gained an interactive pricing calculator, and moved the proof points into a beautifully art-directed graphic. Every one of those decisions improved the human plane and degraded the agent plane.

And nothing in your analytics tells you. Session recordings show humans. Heatmaps show humans. Conversion tracking shows humans. The agent that read your page, found four of the nine facts it needed, and moved on to a competitor leaves no trace in any dashboard you own. You find out when somebody on the team types the category into an assistant out of curiosity and gets an unpleasant surprise.

I'll admit our own exposure here. On Nudgent's site, five pages canonicalized to the homepage because a root-layout canonical tag leaked into /privacy, /terms, /security, /early-access and /apply. We found it on 21 September 2026 while doing unrelated traffic work. Five pages were telling every crawler that reached them that they were actually a different page. No visual review would ever have surfaced that, because the pages looked exactly right. We were shipping a markup defect for months while reviewing the rendered output and pronouncing it fine.

That is the shape of the problem. It hides in the layer nobody looks at.

How page structure affects AI comprehension, dimension by dimension

Guessing at which specific element is broken is slow and usually wrong, because the symptom (not getting cited) doesn't point at the cause. A more useful approach is to check the page against the signals an agent is actually using. Nudgent's agent plane scores seven of them:

Extractability. Is the substance in clean, machine-readable text, or trapped in an image, a canvas element, or a JavaScript-only render?

Passage retrievability. Are there self-contained chunks an engine can lift verbatim and have them still make sense? A paragraph that begins "This means that..." is not retrievable, because the antecedent is three scrolls up.

Claim verifiability. Are claims specific, attributed, and independently checkable? "Trusted by leading teams" is unverifiable. "Used by 1,400 engineering teams, per our January 2026 customer count" is checkable, and an assistant treats those two sentences very differently.

Answerability. Does the page actually contain the facts an engine would need to answer a question about your category? If a user asks "does X integrate with Y," and the answer lives only in your docs subdomain, your marketing page is not answerable.

Freshness. Are dates present and current? A page with no published or updated date is harder for an engine to weight against a dated competitor.

Markup hygiene. Is the structured data and semantic markup clean enough to parse? This is where our canonical leak lived.

Entity and authority. Is the entity clearly defined, named consistently, and corroborated by third-party signals? This is Cutter's authority control, arriving 150 years later in a different medium.

The glossary entry on agent extractability goes deeper on the first of these, and the methodology page explains how the scoring works across both planes.

The reason to check all seven rather than one is that they interact. A page can be perfectly extractable and still fail on entity resolution, in which case an assistant reads every word and still can't confirm who you are. Fixing extractability alone would feel like progress and change nothing.

What to do about it this week

Pick your two closest competitors and audit their pages alongside yours on the agent plane specifically. Your score in isolation is not very informative. A 62 sounds bad until you learn the best page in your category scores 58, and it sounds fine until you learn two competitors score 81. The comparison is the signal.

Then work the list in order of how much substance each defect hides. A JavaScript-rendered pricing table is a bigger loss than a missing published date, because pricing is one of the facts assistants are most often asked to retrieve. Fix the thing that makes you unciteable before the thing that makes you slightly less citeable.

One more thing worth doing, and it takes two minutes. Open an assistant and ask it what your company does and who it's for. Read the answer carefully, then ask yourself where it came from. Did the assistant get that from your page, or did it get it from a directory listing, a review site, a three-year-old press release, or somebody else's comparison post about you? If the description is broadly right but not in your words, and the details are the ones you'd have chosen second rather than first, then something other than your own page is currently defining you to every person who asks a machine about your category. You don't control that description, and you probably don't know what it says.