Agent-Plane Optimization: The Second Reader of Your Page

Malcom McLean's converted tanker, the Ideal-X, left Newark for Houston in April 1956 carrying fifty-eight truck bodies bolted to its deck. The cargo inside them was nothing new. What was new was the box: a standard steel container a crane could lift without anyone opening it or knowing what was in it. Within twenty years that box had rearranged which ports mattered, which ships existed, and what it cost to move almost anything.

A dual-plane audit scores a page twice, from two different readers: the human plane (can a person understand this page, trust it, and act on it) and the agent plane (can an AI agent extract the facts on this page, verify them, and act on them). Agent-plane optimization is the work of raising that second score.

Containerization changed none of the cargo and all of the handling, and the transfer here is uncomfortably direct. Your page's substance is the cargo, and it is identical on both planes: the same pricing, the same proof, the same description of what the product does. The packaging decides who can pick it up. A price rendered by a JavaScript calculator and a price sitting in a table are the same fact in two boxes, and only one of them can be lifted by a machine that does not click. Which is why a page can be genuinely excellent for the people reading it and close to invisible to the thing summarizing it for them.

Where AI agent readability falls between the tools you own

Nobody sold you a tool for this, which is most of the reason it goes unmeasured.

Look at what your stack actually checks. An SEO platform checks crawlability, index status, rankings and link equity. It answers whether a page can be found. It has nothing to say about whether the page, once found, contains anything an engine would be willing to quote. A CRO or analytics tool checks what humans did: clicks, scroll depth, funnel exits. It has never seen an agent and could not distinguish one from a bot it filters out. An accessibility scanner comes closest, because screen readers and agents both consume the DOM rather than the render, but it checks for compliance with a standard about human assistive technology, which overlaps with comprehension without being the same question.

So the gap is structural rather than an oversight. Three tool categories, each doing its job, and the question "can a machine read this page and act on it" falls between all three. Meanwhile agents are already answering that question about your pricing page on a buyer's behalf, and nobody on your team sees the answer.

I should admit that we were on the wrong side of this ourselves. Before Nudgent shipped any content at all, an internal review in September 2026 found that every report and showcase page the product generated was set to noindex and the sitemap carried four URLs. A year of audits had produced a large amount of substance that no agent could reach, on the site of a company that sells agent-plane scoring. That is the failure in its purest form: the cargo existed and the box was welded shut.

What does optimizing pages for AI agents actually measure?

It measures 7 things, each scored zero to ten, each answerable by looking at the page. They are the agent-plane half of the 14 dimensions in an audit, and they are separate dimensions rather than the human 7 asked again in a machine accent.

Extractability. Is the substance in clean, machine-readable text, rather than trapped in an image, a canvas element, or a render that only happens after a script runs? This is the one that fails most often and costs the most, and it has its own longer write-up in the agent extractability entry.

Passage Retrievability. This one is about whether an engine can lift a chunk verbatim and have it still make sense on its own. A paragraph that starts "As we mentioned above, this means you can" is unquotable, because the quote arrives somewhere else stripped of the sentence it depended on.

Claim Verifiability. Are the claims specific, attributed, and checkable by someone other than you? "Trusted by leading teams" is not a claim, it is a mood. "Used by 1,200 engineering teams, listed by name on our customers page" is a claim, and an engine can decide whether to repeat it.

Answerability. Does the page contain the facts an engine would actually need to cite it? If a buyer asks an assistant "does this integrate with Snowflake and what does the integration cost", and both answers live on a different page or in a sales conversation, your page is not in the running regardless of how well it is written.

Freshness. Dates, and whether they are still current. An undated page is not neutral. It reads as possibly stale, and an engine choosing between two sources will often take the one that says when it was last true.

Markup Hygiene. Is the structured data and semantic markup clean enough to parse? Schema.org that contradicts the visible page, headings used for styling rather than structure, and three H1s on one screen all degrade this.

Entity and Authority. Is the entity clearly defined, named consistently, and corroborated somewhere other than your own marketing? This is the one you cannot fix entirely on your own page, which makes it the slowest to move.

Those names are the current set. If you have read an older description of this framework anywhere, including from us, the previous dimension names were replaced wholesale rather than renamed, and the seven above are the ones an audit reports against today.

The pricing page an agent cannot read

Take a specific, very common shape. A pricing page presents three plans through an interactive calculator: you pick a seat count with a slider, toggle annual or monthly, and the numbers update. It is a good piece of design. A human visitor gets an answer tailored to their situation in about four seconds, and the page probably scores well on the human plane for exactly that reason.

Now ask an agent what the product costs.

The agent receives the document. The slider's default state may or may not have rendered a number into the markup. The plan names are probably there. The feature lists might be in a component that only mounts on interaction. The number a buyer asked about often exists nowhere in the text the agent can see, because it was always going to be computed, and computation requires a click nobody made.

What comes back to the person who asked is not an error. That is the part worth sitting with. The assistant answers the question using whatever it could find, which might be a competitor's comparison page, a two-year-old review, or a forum post with a number that was true in 2024. Your page did not lose an argument about price. It was never in the room, and no dashboard you own reports a page as absent from a conversation it was never part of.

The fix here is rarely a rebuild. It is usually a static fallback: the same plan names, prices and feature lists present in the markup, with the calculator sitting on top as the human-facing layer. Same cargo, packed so a crane can reach it. The structural extractability gap piece works through more of these patterns.

A dual-plane model B2B SaaS teams can actually use

The two scores stay separate, and that decision is worth defending, because a single blended number would be easier to put on a slide.

A blend hides the case you most need to see. A page at 80 on the human plane and 30 on the agent plane averages to something respectable and tells you nothing about the fact that half your readership cannot parse your pricing. The same average could equally be a page that is machine-perfect and incomprehensible to a person. Those two pages need opposite work, and the composite cannot distinguish them.

Keeping them separate also keeps the honesty in. The planes genuinely do fail independently: fixing conversion friction for humans does not reliably improve extractability, and adding structured data does nothing for a visitor who cannot tell which plan is for them. Any tool that promises one number for both is either measuring one plane and rounding, or averaging away the signal.

Where our own method is weakest is the Entity and Authority dimension. It depends on corroboration that lives off your page, on third-party sources we can see only partially, and a score there is a reading of available signals rather than a full picture. We say so in the report. Tools built specifically to track AI visibility across assistants, which is a different job from scoring a page, cover some of what we cannot, and we have written about that division of labor in the Otterly.AI comparison.

What to do next

Run a dual-plane audit on the page where money changes hands, and read the agent-plane score first, before you look at the human one. If the human score is already strong, that is the interesting case rather than the reassuring one, because a strong human score is exactly the condition under which nobody thinks to check the other plane. Start with Extractability and Answerability, which move fastest and are usually a markup change rather than a content project. More of how the scoring works sits in the methodology, across the blog, and in the glossary.

Then there is the question this whole piece exists to make unavoidable.

You can tell me your signup page's conversion rate to one decimal place, and you should be able to. What is its agent-plane score? And if the honest answer is that you do not have one, which tool currently in your stack was ever going to produce it?