Your next prospect will probably meet your website secondhand. They ask an AI tool built on a large language model who does what you do, the tool reads a dozen sites in the time a person reads one headline, and yours either survives that reading or it never comes up. HubSpot published the numbers behind this shift when it retired the INBOUND name this year. Customer organic traffic fell 27% year over year while AI referral traffic roughly tripled. The front door moved.
We wrote about what this did to trust already. This is the companion piece, written from having rebuilt our own site around what comes next.
Machines read labels, not layouts
Schema markup is a labeling layer underneath your pages, invisible to people, that tells machines exactly what they are looking at. This is an article. It was written by this named person, who holds this role at this organization. It was published on this date and revised on that one. This page describes a service, and this is what the service is called. Without the labels, an answer engine is guessing at your site the way a stranger guesses at an unlabeled warehouse. With them, every claim has a who, a what, and a when attached, and one study measured a 37% visibility lift for content that gives machines a named expert to attach. The chain matters more than any single label. Article points to author, author points to organization, and everything resolves and agrees. A machine deciding what to cite behaves like a careful reader deciding what to trust, and it looks for the name.
Your page builder may be shredding your best facts
Old-school WYSIWYG builders compose pages out of stacked widgets, and widgets break meaning apart. Think of the stat tile every B2B site has, a big number in one block with its descriptor in a separate block underneath. A person reads “87%” over “of buyers start with a shortlist” as one fact, because human eyes join things that sit near each other. A machine reads the markup, and in builder markup those are two unrelated fragments living in different nests of code. The number arrives with no meaning attached, the sentence arrives with no number, and your most quotable claim dissolves on contact. The same shredding happens to icon-headline-blurb cards, to text trapped inside images, and to anything that only appears after a script runs. The fix is structural. Content lives in the page in reading order, a number and its descriptor share one semantic unit, headings are real headings instead of styled text, and the page still says everything it means with JavaScript turned off. This is much of why we build without page builders at all. What renders is what the machine reads, so the markup has to say what the design shows.
A byline is only as strong as the person behind it
Putting names on your content is step one. Step two is making the name resolvable, because an answer engine checking credibility does what a careful editor does. It looks the author up. That means every author gets a real bio page the byline links to, marked up so machines read it as a person with a role at an organization, and connected outward to where that person exists beyond your site, starting with LinkedIn. It means the bio states checkable facts, years, roles, and past employers, instead of adjectives. And it means letting your experts publish repeatedly under their own names on their own subjects, because authority accrues to a person the way it never accrues to a brand account. An author your site invented, who exists nowhere else and signs one post a year, reads as decoration. An author with a resolvable identity and a body of signed work reads as a source.
One page per thing you want to be known for
The architecture mistake AI punishes hardest is the single “what we do” page with your whole offering in six bullets. A machine can only recommend what it can resolve to a URL. If a buyer asks their AI tool who does the specific thing you sell, and that thing lives as a bullet point, you lose to whoever gave it a page. So the site gets built around the catalog. Every product, service, and solution gets its own page, named in the words a buyer actually uses rather than the words your deck uses, opening with a direct answer to what it is and who it is for, carrying its own FAQ and its own labels, and linked from every place the offering is mentioned. The same rule runs downward into the questions around each offering. Pricing questions, comparison questions, and how-it-works questions each earn their own page, because each is its own search. The point is legibility, your offering made addressable one thing at a time.
Answer the question where the question is asked
Answer engines quote pages that answer directly. That changes how pages get written. The question your buyer actually types becomes the title, the answer leads and the argument follows, and the page carries its own description so nothing ships unlabeled. It also changes what you publish. An llms.txt file gives machine readers a maintained guide to the building. Ungated content gets read and cited, while gated content is invisible to the reader that matters most now. If your best thinking sits behind a form, the machines will cite whoever published theirs.
Speed stopped being a nicety
Google’s health read on your site comes from real visitors’ devices now. It measures how fast the main content paints, whether the layout jumps, and whether interactions respond. Every plugin, tracking pixel, and page-builder layer works against you twice, once in the ranking and once in the reading. Our rebuild treats bytes as a budget the build enforces. The whole stylesheet crosses the wire smaller than most sites’ logo file, and the scores follow, 100 out of 100 on desktop with a layout shift of exactly zero. The numbers matter because a slow site tells both of your readers the same thing about how you work.
The human reader still makes the decision
The machine filters; the person chooses. And a person who arrives pre-informed by AI needs something the summary could not give them, evidence of taste and a reason to come back. That is experience, and it is why the useful site is beating the brochure site. Free tools that do real work earn return visits and the links that machines weigh. Distinctive design earns memory. A site a visitor enjoys is doing strategic work, because the reader it wins is the one who signs the contract.
A website is maintained, not launched
Both readers punish neglect. Machines read staleness directly, in dates that stopped moving and links that quietly died. People read it as what it is, a company that stopped tending its own front door. The fix is treating the site as an operated system. Content lives on a review rota instead of a memory, dates stay honest when work is revised, and the site checks itself, because broken links and missing metadata should be found by your own machinery before a buyer’s AI finds them. Ours walks its own pages weekly and files a report only when something changed.
If you put AI on your site, govern it
Site chat and AI tools are becoming table stakes, and most of them are a liability wearing a widget. Ours runs on hard limits, with rate caps per visitor, a daily budget, every answer validated before it renders, routes clamped to pages that exist, and a log that keeps questions but never people. And the policy that governs all of it is published where anyone can read it, because in the age of AI, how you use it is part of what buyers are evaluating.
All of it is plumbing, and that is good news because plumbing is checkable. None of it waits on a redesign. We rebuilt our own site as the proving ground, then wrote the whole approach down as a standard we now build from, so the details are covered before the design conversation even starts.
A website both readers can trust is the whole assignment now.
