The Gap That Doesn't Respond.

Why the unsaid works on people and disappears for machines.

Erich Posselt recently published a piece here that I believe is right. His text distinguishes clarity from smoothness: clarity is a stance, it can be rough, uncomfortable, polarizing. Smoothness is optimization for consensus, the result that upsets no one and truly moves no one. And strong brands, he finds, live not on completeness but on what they leave open. Apple never explained who "different" is. Hermès justifies neither its prices nor its waiting lists; the brand lets customers come to it, on its terms. The incomplete, he concludes, is not a weakness, it is an invitation.

I share that view from my own experience: I worked for Red Bull for several years. There, mystique was simply part of how the brand described itself. Something that is not explained. Up close, though, you noticed that behind the unsaid lies no accident, but a fairly precise idea of what is never said.

A brand that commits to nothing isn't open; it is merely undecided. And text that falls out of a language model and gets published unedited as a brand's voice is exactly what Posselt describes: technically correct, often elegant and without character.

Still, the principle has a limit. And that limit is becoming important: a growing share of those who deal with a brand are no longer human. AI search summarizes pages before anyone visits them. Browser agents compare offers on behalf of their users. And tools produce content in the brand's name without anyone sitting next to them who carries the brand in their head. This audience doesn't just read differently. It cannot do anything with the unsaid.

What happens when you read your own site like a machine.

So a few weeks ago I tried to see my own website the way a machine sees it.

I built an apparatus for this: it opens the page, reads its source code and sets it against what a human sees in the browser. On the left, the designed; on the right, the processed. Hover a line on the right and the matching spot on the left lights up.

The interesting lines are not the ones that light up. They are the ones that never do.

For testing, I created a page built like a brand's campaign page: mood image on top, claim, image gallery, a decorative divider, a background image behind a block of text. Seven images, the kind found on every brand site. The result: One arrives with its meaning intact. One disappears deliberately and rightly, because it isn't meant to say anything. Five are lost.

One case stands out: the claim. As usual, the campaign's most important sentence is in the image. The image is labeled; formally, everything is in order. What the machine reads is one word: "campaign visual."

The message did not arrive distorted. It did not arrive.

The silent premise.

The unsaid works because someone fills it. That is the silent premise in the praise of the gap. It is already there in the word resonance: the feeling that something responds. Responding is work. It is done not by the piece but by the person standing in front of it. The projection happens in the receiver.

A machine does not do this work. It does not project. It reads what is there. What is missing is simply nothing to it.

So the gap flips its function depending on who stands before it. For a human, it invites. For a system, it stays empty.

Hermès' waiting list is staged scarcity. It only works because someone interprets it: a human reads the waiting list as proof of desirability. An agent comparing offers on a customer's behalf reads it literally: not available. And "Think Different" remains one of the best invitations in brand history, but ask a language model what the brand stands for and it answers from what third parties have written, not from what the brand left open.

This does not refute the principle. It marks out its range. It holds exactly as far as the audience is human. And that share is beginning to shrink.

Two audiences, two rules.

It helps to think of a brand in two layers.

One is the execution layer: how the brand looks and sounds. In classic vocabulary, the corporate design with its base elements logo, color, typography, imagery and grid, plus the tonality of the language. This is where the industry has been good for decades. Part of it has even become machine-capable: design systems and component libraries carry the design through a company's own digital products. But even this is not solved. A standard format in which a brand hands its design to outside tools does not exist, nor are many guidelines precisely written.

The other is the layer of meaning: who the brand is and what holds for it, in classic vocabulary the corporate identity. What it stands for, what it would never do, which claim it can back and which it cannot. Almost everywhere, this layer lives in a PDF, in sentences a human has to interpret.

And it has an old weakness that is now getting expensive: these sentences usually describe the brand, they do not differentiate it. "Premium, innovative, close to the customer" appears in every second brand book. As long as humans were reading, the design made up the difference. An agent comparing offers filters for differences; where none is named, it finds none.

A human can infer one layer from the other. They see the imagery and sense the attitude. They read three sentences and feel the character. That ability is real.

A machine cannot. It doesn't guess; it knows only what is marked up. Which is why this layer needs a sentence that would be unnecessary for a human audience: What is meant to apply must also exist as text.

That is explicitly not the same as: explain everything.

Commitment is not explanation.

Here lies the confusion that creates the apparent contradiction in the first place.

A brand book says: "We communicate at eye level, approachable and clear." Three readers, three interpretations. The sentence isn't wrong, it just isn't applicable.

A specification says: "We use contractions. No more than one exclamation mark per text. Technical terms are explained at first use." Three readers, one result. And a system can apply the same text without guessing.

What gets committed is the rule. Not the meaning.

A brand can commit to staying silent about its prices, to never justifying its waiting times, to never explaining whom it is made for. The secret remains fully intact. It just stops being an accident that depends on who happens to be in the room.

The machine is not the cause.

That leads to the second half of the problem, which at first has nothing to do with machines.

A rule that exists only as interpretation in the head of an experienced person is already lost the moment that person leaves the room. The process is well known; it just runs so slowly that no one perceives it as an event.

It usually starts with good brand work. A team has wrestled for months, there is a result and there are three or four people who know exactly what is meant. They know because they were there. They have seen the discarded variants; they know the discussion behind every sentence. For them, the brand book is not a rulebook but a keepsake. It doesn't need to be precise, because the precision lives in their heads.

Then four years pass.

The lead agency changes because the account comes up for review. The new agency receives the document and interprets it as well as it can. It also interprets it differently, because it doesn't know the history. A social media agency joins and reads the same sentences differently again, because its channel makes different demands. Inside the company, the marketing lead changes, then the person who held the process together. A fourth market translates "approachable" into its language and its conventions of politeness; no one can say whether this is still the same brand, because there is no sentence against which to check it.

None of this is a mistake. Every single person worked carefully. Yet at the end there stands a brand that is five slightly different things in five channels. The discussion about it becomes a matter of taste, because no authority is left except memory. Whoever has been there longest wins. That is not brand management, that is seniority management.

This process is as old as brands. The industry has learned to administer it: with guidelines, with approval loops, with people who look at the result at the end of the process. This administration works because it has time. An approval cycle takes days. Within those days, someone can say: that is not what we meant.

Exactly this time is now disappearing. A presentation, a text, an ad, a reply to a customer now take minutes, in parts of the company that never produced anything before. Nobody decided this; the tool simply sits on every computer. Add to that a shift that was very much decided: more and more brands are bringing content production in-house. What used to pass through an agency now happens internally, faster and cheaper, but without the outside eye that asks whether this is still the brand.

That changes the math. An imprecise rule used to cost the occasional correction loop. Today it is applied a thousand times before anyone looks. And the interpreting is now done by a system that doesn't know the discarded variants, never heard the discussion and when in doubt picks whatever is most likely. This is exactly where smoothness arises, not from bad intent but from a gap nobody closed.

So the machine is not the cause of the problem. It is the amplifier. It makes visible and fast what used to be invisible and slow.

This has a pleasant consequence: the work ahead is not technical. A brand whose rules are precise enough for a system to apply them is also a brand whose rules a new employee understands on day one and an outside agency interprets correctly without a briefing. The effort is spent once and pays in both directions. Do it for the machines and the human half comes free.

Readable, applicable, transferable.

Where this is heading can be told in three stages.

The apparatus shows the first: machines read brands. That is where we are today. And much is already lost here.

The second stage is here as well: machines apply brands. Design tools now carry brand profiles from which they take colors, fonts and tonality; assistants write in the company's tone. Readable means: the machine registers what is there. Applicable means more: it can produce something new that follows the rules. A PDF is readable. It is not applicable.

The third stage is emerging: agents negotiating with agents. A customer asks their assistant, the assistant asks the vendor's agent and that agent hands over a package: product data, prices, terms and the elements of the brand meant to travel along. No human in this chain. The big platforms are already building the protocols. And at that point, the format question becomes a question of authority: what an agent may promise in a brand's name and who stands behind it. That is its own topic, bigger than this text.

For the brand, this means something simple and hard: with every handover, only what is explicit travels along. When people pass a message along, it shifts a little at every step, like in the game of telephone. Between agents, nothing shifts. What is in the package arrives exactly. What is not in the package is missing from the first handover on and then it is missing everywhere.

What of this is proven.

Up to this point, this is an argument and arguments deserve distrust, mine included. So now the evidence, including the inconvenient kind.

Proven: structure is the interface. Google says it in its own developer documentation: agents orient themselves by the accessibility tree, the structure the browser computes from every page and that screen readers rely on as well: every element with role and name, no layout. Since spring, Google's testing tool Lighthouse has been checking this in a category of its own. An element without a name does not exist for an agent.

Proven, too: how widespread the defect is. A survey of one million home pages found unlabeled buttons on 30.6 percent of them; on half, form fields lack labels. The average website reads to a system like a half-labeled form.

Both findings concern the first stage, the website, for a simple reason: it is the arena that can be measured best. For applying and handing over, such measurements barely exist yet. That is not a free pass; it calls for caution.

Because a whole series of things currently being sold is not proven.

Do accessible pages measurably help agents? It sounds plausible and is sold that way. It isn't cleanly proven. The figure usually cited for it in fact measures something else. And the real comparison, the same page once accessible and once not, exists only as an early prototype: 89 versus 49 percent success, across just five tasks.

Structured data, the machine-readable statements a page makes about itself, is considered the standard answer to the visibility question. The cleanest test so far compared about 1,900 pages that had added such data with about 4,000 matched control pages. The result: no measurable effect in ChatGPT or Google's AI Mode and in the AI Overviews even a slightly negative one. I still maintain the markup on my own pages, because it keeps the foundation clean. I don't expect an effect.

And the file currently touted as the ticket into AI search, llms.txt, receives not a single request in 97 percent of cases. Server logs across 137,000 domains show as much. The format was invented for tool documentation, not for visibility.

So whoever expects clean structure to create visibility will be disappointed. It is the ticket in, not the result.

What does correlate has been studied broadly by now. An analysis across 75,000 brands compared which variables go along with visibility in AI answers. Far in front: mentions by third parties, in articles, videos, forums and trade publications, with rank correlations between 0.66 and 0.74. The classic variables of a brand's own website trail far behind: the number of its pages at 0.19, domain authority between 0.27 and 0.33. For context: these are correlations, not causes, and only established brands were examined. The direction is clear all the same.

This is where classic brand thinking is proven right and technology meets its limit. A brand gets recommended because others have written about it, not because it marks itself up well. Structure makes sure that what was written arrives correctly. It cannot replace it.

What this means for brand leaders.

The useful question is not: how do we become machine-readable? Ask that and you end up buying tools that promise exactly what is not proven above.

The useful question is: Which of our rules exist only as interpretation in someone's head?

That question can be answered without technology. It takes a brand book, one honest hour and someone willing to mark the sentences that sound good and decide nothing. Everything else follows. What is decided can be written down. What is written down can be applied, by people and by systems. What is not decided will be decided by someone else anyway and increasingly by something else.

The second step is more uncomfortable and cheaper: look, once, at what actually arrives from your own site. That is not a project, that is an afternoon.

Thinking from the machine's point of view makes strategy work not smaller but more binding. The machine is the first reader that fills in nothing. In front of it, a strategy shows whether it contains differentiation or only description. And for design, the same principle holds one level down: it may keep moving people. But the decisions behind it, what it should show and what it should mean, have to exist as text somewhere.

Back to the gap.

A brand that explains everything has nothing left to believe in. For the layer of meaning, the praise of the gap remains right.

For the rules, the opposite holds. A rule that was never written down does not die of transparency. It dies of interpretation.

You only have to know whom you are leaving the gap to. To humans, it is an invitation and the strongest brands will keep using it. To machines, it is silence and silence is not interpreted. It is passed over.

Brand remains a stance. But it now needs a form that still holds when no one is left in the room who can explain it.

This text was written in dialogue with AI. Research and drafting were done with a language model; every number is verified and linked. The commitments are mine.

Sources

Studies and data this article draws on:

• Google, developer documentation "Accessibility for agents" (agents read the accessibility tree): https://developer.chrome.com/docs/lighthouse/agentic-browsing/accessibility-for-agents
• Lighthouse, "Agentic Browsing" category: https://developer.chrome.com/docs/lighthouse/agentic-browsing/scoring
• WebAIM Million 2026, survey of one million home pages: https://webaim.org/projects/million/
• A11y-CUA, Characterizing the Accessibility Gap in Computer Use Agents, CHI 2026: https://arxiv.org/abs/2602.09310 (the most-cited study; it measures how the agent operates, not the accessibility of websites)
• Designing Agent-Ready Websites for AI Web Agents, 2026: https://arxiv.org/abs/2607.12056 (prototype, 89 versus 49 percent)
• Ahrefs, controlled test of structured data's effect on AI citations, 2026: https://ahrefs.com/blog/schema-ai-citations/
• PPC Land on Ahrefs' server-log analysis (137,000 domains) of llms.txt usage, 2026: https://ppc.land/llms-txt-adoption-rises-8-8x-but-97-of-files-get-zero-ai-requests/
• Ahrefs, correlation analysis across 75,000 brands on visibility in AI answers (ChatGPT, AI Mode, AI Overviews): https://ahrefs.com/blog/ai-brand-visibility-correlations/
• Google, announcement of the AP2 agent payments protocol (agents buying from agents): https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol
• The comparison described in this text is public: https://robert-haase.de/en/machine-view.html