llms.txt Is Everywhere. The Bots Aren’t Reading It.

Your SEO platform flagged a missing llms.txt file and called it a site issue. It isn’t one. Somewhere in the last year, a small Markdown file at the root of your domain turned into the thing every audit tool nags you about and every other LinkedIn post swears you need. The pitch is tidy. Robots.txt tells crawlers where they can go, sitemaps tell them what exists, so llms.txt must be the file that gets your brand into ChatGPT. That logic is clean. The crawler data says it’s also wrong.

What llms.txt actually promises

llms.txt is a proposal from Jeremy Howard, published in September 2024. The idea is reasonable. Raw HTML is a mess of navigation, cookie banners, pop-ups, and scripts, and a language model has to strip all of that away before it can read your actual content. An llms.txt file hands the model a clean, curated Markdown overview of your site so it can skip the cleanup. That’s the whole design. It helps a model use your site at inference time. It was never built to change your ranking inside an AI answer box.

Somewhere between the proposal and the pitch deck, that got lost. Tools started treating llms.txt like the next robots.txt. It reads like progress. It’s mostly a category error.

The crawler data is in, and it isn’t close

OtterlyAI ran the experiment everyone kept theorizing about. They put a correctly built llms.txt at the root of a site and watched the server logs for 90 days. Out of more than 62,000 AI bot visits, 84 touched the llms.txt file. That’s about a tenth of a percent. An average content page on the same site pulled roughly 265 bot visits in the same window, so the file the whole industry keeps nagging you about performed three times worse than a normal page, and no better than a stray PDF. The logs tell the story the theorizing couldn’t.

It’s not one study, either. Search Engine Land tracked llms.txt across ten sites and found two with AI traffic gains, neither traceable to the file. Google has said plainly that Search doesn’t use llms.txt, and John Mueller compared it to the old keywords meta tag, which is about the most polite way an engineer can call something dead. Adoption keeps climbing, north of a hundred thousand files indexed this spring, and the bots keep not reading them. Publishing a file isn’t the same as being read. The gap between those two is the whole story.

Where llms.txt actually earns its place

This is where the honest answer splits from the hot take. llms.txt isn’t useless. It’s just aimed at a different target than the one marketers keep firing at.

If you run documentation, an API, or a product that other people’s AI tools plug into, llms.txt is a genuinely good idea. Stripe, Vercel, Cloudflare, and Anthropic all ship one, and none of them are chasing a ranking. They’re making it cheap for a third-party AI assistant to pull a clean version of their docs instead of scraping and de-noising a page on every request. When a tool has to hit a search API at ten to fourteen dollars per thousand queries, a pre-built Markdown snapshot saves real money and returns better answers. That’s an integration decision. It’s an engineering courtesy to the developers building on top of you.

So the test is simple. If people build AI products on top of your content, llms.txt does something real. If you just want Perplexity to mention your brand, it does nothing. Same file, two very different jobs.

What actually moves your AI visibility

The uncomfortable part is that the things that get you cited by an AI engine are the same things that have always worked, which is why nobody’s selling them as a shortcut. Write content with a real point of view. Structure it so a machine can tell what the page is about. Mark it up with clean, honest schema. Earn references from places that already carry authority. We’ve made this argument before, from the strategy side: if your team can’t agree on what you stand for, neither can ChatGPT.

None of that fits in a file you drop at your root and forget. It’s slower, it’s harder, and it’s the actual work. A tidy llms.txt on a thin, forgettable site is a clean label on an empty jar.

If you want a site that AI engines can actually read and want to quote, that’s a content and build problem, not a config file. That part we can help with. Add the llms.txt too, if you like. Just don’t file it under strategy.

The Agent-Ready Website Is Mostly the Web Done Right

An AI agent booked a hotel room last week without a human ever looking at the website. It read the page, found the form, filled in the dates, and moved on to the next task in its queue. That is not a preview of where the web is going. That is a Tuesday in 2026.

So here’s the question on the table for anyone who owns a site. When an agent shows up to read your content or complete a purchase on someone’s behalf, can it actually do the job? Most teams answer that by bracing for a rebuild. They have it backward. An agent-ready website is rarely a new website. It’s the fast, structured, semantic web you already owed your human visitors, plus two or three genuinely new hooks that are still early enough to be optional.

We build sites for a living, and we’ve watched every “the web is changing, tear it all down” cycle since 2000. Most of them were vendors selling a rebuild. This one is different, because the change is real and the fix is mostly boring. That’s a good thing. Boring is cheaper than panic.

What “agent-ready” actually means

For most of the last two years, browser agents worked like a person with bad eyesight and no patience. They took a screenshot, guessed which pixels were a button, clicked, read the screen back to a model, and hoped. It was slow, and it broke constantly, because your site was never talking to the agent. The agent was reverse-engineering your interface every single time, and a redesign could blind it overnight.

Agent-ready flips the direction of that conversation. Instead of the machine guessing at your interface, your site tells the machine what it can do and what each action needs. Two very different things are happening under that one label, and they get conflated constantly. Agents that read want to find you, understand you, and cite you when a person asks a question. Agents that act want to finish a task on someone’s behalf: compare three vendors, fill the quote form, add to cart, check out. Reading is a visibility problem. Acting is an infrastructure problem. Your site has to be good at both, and they are not the same work.

We’ve written before about why your brand is invisible to ChatGPT, and that piece was about the reading half: getting understood and quoted when an assistant answers a question. This is the acting half. The stakes are higher, because a purchase is worth more than a mention, and the plumbing is harder, because completing a transaction reliably is a lot to ask of a machine that’s poking at a page built for a human thumb. You can read about the visibility side in our take on AI search visibility. This article is about the side where the agent has a credit card.

The numbers stopped being hype

It’s fair to be skeptical of any trend that shows up with a fresh acronym every quarter. This one has receipts. Adobe Analytics measured AI-referred retail traffic growing close to 400% year over year in early 2026, and that traffic converted meaningfully better than ordinary search traffic once it landed. The IBM Institute for Business Value found that 45% of consumers already use AI for some part of the buying journey. McKinsey put a number on the destination: somewhere between $900 billion and $1 trillion of US retail influenced by agentic commerce by the end of the decade. Shopify has already exposed more than a million merchants to agent-driven checkout, and Amazon’s shopping assistant is serving hundreds of millions of people.

Now the part the trend roundups leave out. The infrastructure isn’t ready, and the data says so out loud. When Walmart let people buy inside a chat interface, those purchases converted about a third as well as sending the same shopper to Walmart’s own site. Read that twice. The demand is real. The plumbing is not. That gap between what shoppers want to do and what merchant sites can actually support is the entire opportunity, and it’s a build problem wearing a marketing costume.

Most of an agent-ready website is the web you already owe your users

Here’s the deflating truth for anyone hoping to buy a shiny new layer. The single best thing you can do to prepare for AI agents is the thing you were supposed to do for humans all along. An agent parses your page using the same signals an assistive screen reader and a search crawler have always relied on. Clean, semantic HTML. Headings that actually describe the content beneath them. Forms with real labels tied to real inputs. Buttons that are buttons, not a div wearing a click handler. Structured data that states, in a language a machine can read, what this page is and what lives on it.

We made the case a few weeks ago that the prettiest 2026 web design trends are also the slowest, and that argument holds double here. A three-second hero animation that hogs the main thread annoys a person. It defeats an agent outright, because the agent times out before your beautiful thing finishes painting. Core Web Vitals were never just an SEO scoreboard. They’re now the line between a machine finishing a task on your site and quitting halfway through. If you want the full version of that argument, we wrote it up here.

Accessibility tells the same story through a different door. The markup that lets a blind user move through your checkout with a keyboard is nearly identical to the markup that lets an agent move through it with no eyes at all. A labeled field, a logical tab order, a form that announces its errors: a screen reader needs those, and so does a machine. Do the fundamentals and you’re most of the way to agent-ready before you touch a single new standard. Skip them, and no protocol on earth bails you out. The agent is only as capable as the page is honest about itself.

Isn’t this just SEO with a new coat of paint?

Reasonable question, and the honest answer is: partly, and that’s the point. A lot of “optimize for AI agents” advice is the structured-data hygiene good developers have preached for a decade, repackaged with urgency. If someone is selling you an agentic web package that turns out to be schema markup and a sitemap, you’re not being lied to, exactly. You’re being charged a premium for fundamentals.

But the acting half is genuinely new, and it’s where the coat of paint stops explaining things. Search optimization was about being found. Agent readiness is about being operated. A crawler indexed your content and left. An agent wants to submit your form, apply your coupon, and complete your purchase, and it wants to do that reliably enough that a person trusts it with a card. That’s not a ranking problem. It’s an API problem, and it’s why the next section matters even though the last one told you to fix your HTML first.

The genuinely new hooks: WebMCP and llms.txt

Two things are actually new, and they’re worth understanding before you decide how much to bet on either.

WebMCP is the bigger one. It’s a proposed browser standard, backed by Microsoft, Shopify, and Booking.com among others, that lets your site register its own tools through a new browser API. Instead of an agent screenshotting your checkout and guessing at the fields, your site declares its capabilities out loud: here is a “checkout” tool, here is a “filter_results” tool, and here are the exact inputs each one expects. Google opened a WebMCP origin trial in Chrome in the spring of 2026, which means developers can switch it on for real users and test it in the wild. You can read Google’s own writeup on the Chrome for Developers WebMCP origin trial. There are two ways to expose those tools: annotate the HTML forms you already have, or define functions in JavaScript with a schema describing their inputs and outputs. If you’ve ever built an API, the mental model clicks instantly. You’re handing the agent a clean API to the thing your page already does, instead of making it infer that API from pixels.

The honest caveat matters more than the feature. An origin trial is not a finished standard. It can change shape, and it can be pulled. Building your entire storefront around it today would be a bet, not a plan, and we’d talk a client out of it. But if you run a real transactional flow, a booking, a quote request, a cart, it’s worth a pilot right now. The teams that understand WebMCP while it’s still experimental are the teams that ship it well the day it becomes table stakes. Early is cheap. Late is a scramble.

llms.txt is the low-effort one, and there’s almost no reason to skip it. It’s a plain markdown file at the root of your domain, the same basic idea as robots.txt or sitemap.xml, that tells language models what your site is about and which pages actually matter. Adoption sat around 10% of sites by the middle of 2026, so it’s early, but it costs an afternoon and it can’t hurt you. If you’d rather hand the models a curated map of your site than let them guess from your navigation, you write that map yourself. Most companies haven’t bothered. That’s a cheap edge sitting on the table. You can read the llms.txt proposal and have a draft live before lunch.

Agentic commerce commoditizes the middle. Brand is the defense.

Now the part that isn’t a developer problem at all, and the reason this belongs in front of whoever owns the brand and not just the codebase. When an agent does the buying, it optimizes on whatever it can measure. Price, specs, availability, star ratings. If the person only said “find me a good laptop under a thousand dollars,” the agent runs a spec comparison, and the product with the cleanest data and the lowest number tends to win. That’s a race to the bottom, and it’s a race you cannot win by being slightly cheaper than the next feed in the list.

There’s exactly one instruction that changes the whole game, and it comes from the human, not the code. It’s the person naming you. “Order me another one of those from Aerie.” “Book the Kimpton, not the cheapest room on the block.” The instant a human asks for you by name, the agent stops comparing and starts fetching. Distinctiveness isn’t an AI default. It’s the thing that survives the agent. We argued that taste is the only real moat in the AI era, and agentic commerce is where that argument gets tested for actual money. A brand strong enough to be requested by name is a brand the agent has to honor. A brand that only competes on the spec sheet is a line item the agent optimizes away without a second pass.

So the technical work and the brand work turn out to be the same project aimed at the same threat. Being readable gets you into the agent’s consideration set. Being memorable gets you requested by name and lifted out of the comparison entirely. You need both halves. Clean markup with a forgettable brand is a site that agents can use to shop you into a commodity. A beloved brand on a site an agent can’t operate is a name people ask for and then can’t get served. The failure modes are different. The fix is one coordinated push, not two.

Where to actually start

You don’t necessarily need a rebuild. You need a sequence, ordered by what pays off soonest.

Fix the fundamentals first, because they serve people and machines in the same stroke. Semantic HTML, labeled forms, honest structured data, and a site fast enough that an agent doesn’t give up waiting on it. If your Core Web Vitals are in the red, start there and close the trend-piece tabs. Then write your llms.txt, because it’s an afternoon of work almost nobody has done. Then, only if you run a genuine transactional flow, pilot WebMCP inside the origin trial and learn how it behaves before anyone makes it mandatory. Watch it closely. Don’t wager the business on an experiment.

Notice what’s not on that list. Ripping out your site. Buying an “AI-ready” template that looks like every other AI-ready template. Bolting on a widget that promises to make you agent-friendly overnight, the same way overlay widgets once promised to make you accessible overnight and mostly didn’t. The shortcut is the trap. It was the trap with accessibility, and it’s the trap here, wearing new branding.

The agentic web rewards the boring virtues. Structure. Speed. Clarity. A brand worth asking for by name. We’ve been building on those four for twenty-six years, which is a strange thing to feel vindicated by, but here we are. If you’d rather not sort the real signal from the acronym noise on your own, that’s the kind of build we do. Start with the fundamentals this week. The standards will keep. Your Core Web Vitals won’t fix themselves while you wait for them to.

Vibe Marketing Scales Your Brand Voice. Or Your Slop.

Vibe marketing is the rare trend that lives up to its own noise. One marketer, a stack of AI tools, and the output that used to take a department (yes, for those who don’t know, vibe marketing is essentially using generative AI and LLM’s to do your marketing for you VERY quickly). We’ve watched a single person spin up a campaign, a landing page, ten ad variations, and a week of social posts in an afternoon. The tooling works. That’s not the question. The question is what you fed it.

What vibe marketing actually changed

The bottleneck moved. For twenty years the slow part of marketing was production: the writing, the building, the resizing, the endless rounds. Vibe marketing collapsed that. Searches for the term jumped triple digits in a single year, and the reason isn’t a mystery. When one person can produce what used to need six, every budget-strapped team wants in.

So the marketer’s job changed shape. You’re no longer making most of the work. You set the direction and the voice, then orchestrate machines to execute the rest. The judgment moved up the stack. The typing moved out. That part is genuinely good.

Here’s what nobody priced in. The machine will execute whatever voice you give it. Including no voice at all.

Scale doesn’t create a voice necessarily.

Feed a generative tool a clear, decided voice and it will hold that voice across a thousand assets. Feed it three adjectives and good intentions and it hands you the same beige every other company’s AI hands them. That’s the slop problem, and it isn’t a tooling failure. It’s a foundation failure.

When you generate at scale with no documented voice, you don’t get your brand louder. You get the statistical average of the internet, wearing your logo. Marketers are already feeling it. The AI backlash building right now isn’t really about AI. It’s about everything starting to sound the same.

Think about the brands you can identify from one line of copy with the logo covered. Liquid Death. Progressive. Mailchimp in its prime. None of that is an accident, and none of it survives being handed to a tool that was told to sound engaging. A voice that distinct came from a person deciding, in advance, what the brand would and would not say.

We’ve made this case in a different shape before. Distinctive is what humans force into the work, not what a model defaults to. If your team can’t agree on how the brand sounds, neither can ChatGPT. The tool is a mirror. It reflects the clarity you bring or the lack of it, and at vibe-marketing speed it reflects it everywhere at once.

Decide the voice before you teach a machine to repeat it

A real brand voice foundation isn’t a mood board with three tone words on it. Friendly, bold, approachable describes half the companies in your category and tells a model almost nothing (I can’t tell you how many times we have had this conversation with our clients). A usable voice is a set of decisions. What do we sound like when we’re confident. What do we never say. Do we use contractions or write buttoned-up. Do we open with the problem or the promise. What’s the joke we’d make, and the one we wouldn’t.

Those are choices a person makes once, on purpose, so a machine can repeat them ten thousand times after. That’s the actual work, and it’s the same work whether you run a team of forty or a team of one with a tool stack. Skipping it doesn’t save time. It moves the cost to every asset you ship.

The brands winning at vibe marketing aren’t the ones with the slickest prompts. They’re the ones who knew exactly who they were before they pressed go.

The multiplier cuts both ways

Vibe marketing is a multiplier. That’s the whole point, and it’s also the whole risk. It multiplies whatever you put in front of it: voice or vacuum, clarity or confusion. Decide what you sound like before you scale it. If you don’t know yet, that’s not a vibe marketing problem. That’s a brand voice problem, and it’s the one worth solving first. If you’d rather not sort that out alone, that’s the part we’re good at.

Brand Differentiation in the AI Era: Taste Is the Only Moat

Generative AI didn’t kill brand differentiation. It exposed how few brands ever had any.

Strip away the vibe-coded landing pages, the prompt-driven logo generators, and the gradient-stack startup template, and you’re left with the question every agency has been quietly asking since 2000: what actually makes one brand impossible to confuse with another? Most of the trend pieces in 2026 will tell you the answer is taste. We agree. We just think most of the conversation about taste is using the wrong definition of the word.

Brand differentiation: the visual baseline just flatlined

The flood is here. Templates that look indistinguishable from competent agency work cost ten dollars a month and ship in an afternoon. Image generators output landing-page hero shots that fooled us when they first appeared and bore us now that they’re everywhere. The shared observation across every design-trend roundup we’ve read this spring is the same: the visual baseline has flatlined.

That’s not a complaint. It’s a diagnosis. AI didn’t make design worse. It made the floor much higher and the ceiling no different. The work that used to set a brand apart, the polished hero, the elegant grid, the moody photography, is now table stakes. A brand competing on production polish in 2026 is competing on the part of the work AI is best at.

We’ve watched this happen before. Every time a creative tool democratizes a craft, brand differentiation moves up the stack. Photoshop did it to retouchers. Squarespace and Wix did it to small-business sites. AI is doing it to everything below the strategy layer. The brands that win in the next decade won’t outproduce the AI. They’ll out-decide it.

Taste isn’t an eye. It’s a refusal.

The word taste gets thrown around in 2026 like it’s a personality trait. Taste is the only moat, the headline-writers say, and they’re right about the moat. They’re vague about the taste.

Here’s the version we’d defend after twenty-six years of building brands. Taste isn’t an eye. It’s a refusal.

A designer with taste isn’t someone who recognizes the good options. It’s someone who has spent years learning which good-looking options are wrong for this brand and saying so out loud. The discipline shows up as a long list of things the brand will not do. Fonts it won’t use. Words it won’t say. Categories it won’t enter. Discounts it won’t run. Trends it won’t follow even when its competitors are running toward them with their hair on fire.

That’s the moat. Not the choices on the page. The choices that didn’t make it.

 

What taste looks like when a brand actually has it

Look at the brands you can identify from a single object across the room. They share one trait. They have been ruthlessly selective about what they put into the world.

Liquid Death sells canned water. The brand could have leaned into wellness, hydration, mindfulness, eco-credentials, the same well-mapped territory every other beverage startup raced into a decade ago. It refused all of it. Heavy metal aesthetics, mock horror, a stripped-down can that reads more like a craft beer than a Gerolsteiner clone. The refusal is the brand. The product is incidental.

Apple has refused feature-comparison advertising for nearly thirty years. Every other consumer technology brand of the same era has, at some point, lined up specs on a chart and pointed at the bigger number. Apple’s competitors still do it in 2026. Apple doesn’t. The brand pays for that refusal in lost short-term clarity, and gets paid back in the part of the brand nobody else can copy: a customer who trusts that the company has already made the obvious decisions on their behalf.

Hermès refuses to scale. Patagonia refuses to grow recklessly. The New York Review of Books refuses to put a cover line on the cover. None of these refusals are aesthetic. They’re strategic. The aesthetic is what the refusal looks like once it’s been practiced for thirty years.

When we audit a brand for the first time, we don’t start with what it does. We start with what it has stopped itself from doing. If we can’t find a clear list of refusals, we know what we’re looking at. We’re looking at a brand that’s been improvising its identity, one tactic at a time.

Why AI cannot do this work, even in principle

This is the part the trend pieces tend to skip. Why can’t AI develop taste over time?

Generative models are statistical machines that produce work close to the average of their training data. That’s the technology, not a stage of development. A model can be tuned, prompted, fine-tuned, given style guides and reference images and tone descriptions, and it will get better at imitating a brand’s surface. It still can’t refuse on principle. It can only refuse because somebody told it to.

A brand’s principles, the things that make it impossible to confuse with anything else, are negative space. They’re the choices the brand has rejected so consistently that the absence becomes recognizable. AI is a yes-machine. Ask it for ten options and it gives you ten. Ask it which to keep and it picks the one that looks most like the rest. The model has no skin in the game and no reputation to protect, so it has nothing to lose by saying yes.

Humans with twenty-six years in the room have something at stake every time they say no. That’s where taste lives. Not in the skill of the hand. In the cost of the refusal.

Where most brand differentiation efforts go wrong

A lot of agency work in 2026 is going to sell taste as a deliverable. Most of it will be selling the wrong thing.

The most common mistake is treating taste as aesthetic preference. A brand hires a creative director with a strong portfolio, tells them to make it look great, and turns them loose on the homepage. The work gets prettier. The brand isn’t more distinctive. The CD is doing taste-as-eye, not taste-as-refusal, because nobody has given them the authority or the strategic frame to say no to the CMO’s pet feature, the founder’s favorite trend, or the board’s pressure to look like a competitor.

The second mistake is auditing for what’s there instead of what should not be. A brand audit that catalogs every touchpoint, every channel, every visual asset, and grades them against polished competitor work, will produce a tidy report and almost no useful direction. The useful audit asks the harder question. What is this brand doing that a competitor could do just as well? Cut all of that. What’s left is the brand.

The third mistake is the most common in fast-growing companies. The team is so afraid of leaving any segment unaddressed that the brand says yes to every audience, every channel, every category adjacency. The result is a brand that looks like every other brand at its growth stage. We’ve seen this play out for twenty-six years. The companies that broke through were not the ones who tried to be everything. They were the ones who picked what they were and refused the rest.

 

 

The no-list: how disciplined brands actually build it

If you’d like to start somewhere concrete, build a no-list before you build anything else.

A no-list is shorter than a brand book. It’s a written document, kept current, that names the things this brand will not do. Categories you won’t enter. Words you won’t say in your copy. Visual moves you won’t make. Discount mechanics you won’t run. Customer segments you’ll politely send to a competitor.

The no-list is the most underrated brand creative document in the agency’s toolkit, and it’s the one most brands don’t have. Not because it’s hard to write. Because writing it forces a fight nobody on the marketing team wants to have. Every “no” on the list is a position that somebody, somewhere in the organization, is going to want to violate the next time pressure is on.

That’s the point. The no-list is a contract with your future, more pressured self. We’ve seen brands keep one for years and treat it like the constitution. We’ve seen others keep one for a quarter and quietly let it go when the first big tactical compromise rolls in. The first kind of brand develops taste. The second kind develops a logo system.

Build the no-list. Update it once a year. Read it in every campaign meeting. The discipline of refusal isn’t glamorous. The brand it produces is.

What brand differentiation looks like once AI handles the rest

In a market where AI can produce competent creative for $100 dollars a month, the work that used to differentiate brands has been moved into the commodity column. The differentiating work has moved up. Strategic clarity. Editorial discipline. The judgment to refuse the obvious option even when it’s the option the AI most confidently recommends.

The next two years will sort brands into two groups. The first will use AI to produce more of what their competitors are already producing, faster, and they’ll discover that more of the same, faster, isn’t a position. The second will use AI to handle the work that no longer needs human judgment, and they’ll spend their human hours on the part of the work AI can’t touch: the refusal, the position, the standard nobody else is willing to hold.

The agencies that thrive will be the ones helping the second group. Not because we type prompts faster, but because we’ve spent two and a half decades in rooms saying no on a brand’s behalf. Distinctive isn’t an AI default. Distinctive is what humans force into the work, on purpose, by leaving most of the obvious options on the cutting-room floor.

This is what creative and strategy retainers are actually for. Not deliverables on a calendar. A standing relationship with a partner who knows your brand well enough to say no on your behalf, in the meeting where it matters, before the bad idea ships.

If you’d rather decide what your brand refuses than improvise it later, that’s where we come in.