Your site's first reader isn't a person

I've spent my whole career tuning interfaces for people. Now there's a machine reading the site before they do, summarizing it its own way, and deciding whether the person ever arrives. Pull up a chair, this one is more structural than it looks.

I've been noticing something in my own behavior lately. I need to know which library to use for a thing, or what happened to some brand, or how a concept I forgot actually works. I ask. I read the answer. I close the tab. I didn't click anywhere. Somebody wrote that content, carefully, and I never saw the page.

Yeah. And it isn't just me.

The Pew Research Center tracked the actual browsing behavior of 900 US adults in March 2025, close to 69 thousand Google searches, almost 13 thousand of which came back with an AI summary at the top. When the summary showed up, a click on a traditional result happened on 8% of visits. When it didn't, 15%. Roughly half. And the click on a link inside the summary itself, the one that would send the visit back to whoever wrote the thing, happened on 1% of visits.

One percent.

There's another number in the same study that bothers me more than that one: with an AI summary on screen, 26% of people ended their browsing right there. Without a summary, 16%. The ready-made answer takes the click and closes the session with it. The person left satisfied. Just not on your site.

If the click is gone, what's left?

Being read. And that changes the question entirely.

For about fifteen years we designed around whoever lands on the page: where the eye falls first, how much scrolling until the information that matters, which button earns weight. That still holds. But a reader showed up in the middle of the path who reads the page before any human does, pulls out whatever it manages to understand, builds a summary, and that summary is what the person will read. If it misunderstood, the wrong version is the one circulating. If it couldn't read anything at all, you simply don't exist in that conversation.

There's a name going around for this: Machine Experience, MX. I'll use the term here because it's handy, with an honest caveat attached: it has no owner and no settled definition. It's a lot of people writing about the same realization at the same time, somewhere between 2024 and now, all reaching for the same acronym. Treat it as a useful nickname, not as an established discipline.

What isn't vague is the technical part. That's what I want to pull on, because that's where our half of the job lives.

Why is this a design problem and not one for the marketing folks?

Short answer: because the levers sit in the code and in how the page is built, not in the content calendar.

In December 2024, Vercel published an analysis of AI crawler traffic together with MERJ, and there's a finding in there that changes what we hand over. Looking at what those bots went after on the Next.js site, GPTBot spent 11.50% of its requests downloading JavaScript files, and ClaudeBot 23.84%. And none of the major AI crawlers execute that JavaScript: not OpenAI's, not Anthropic's, not Perplexity's, not Meta's, not ByteDance's. They grab the file as text and move on.

Translated for those of us who build product: if your content only appears after React has mounted the screen in the browser, the bot read an empty page. The gorgeous layout, the hierarchy you fought to defend, the copy that went through three rounds of revision, none of it arrived. It read a <div id="root"></div> and left.

There are exceptions: per the same analysis, AppleBot renders through a browser-based crawler, and Gemini leans on Googlebot's infrastructure, so it renders too. But the two highest-volume dedicated AI crawlers in that analysis, OpenAI's and Anthropic's, do not. Designing for the worst case here is the cheap move.

Still in the same study, the number I find funniest and most revealing: 34.82% of GPTBot's fetches landed on 404 pages. ClaudeBot, 34.16%. Googlebot, over the same period, 8.22%. A third of these bots' effort went into broken links and stale URLs. They aren't clever. They're persistent, and they show up badly informed. The people who fix that are the ones handling routes, redirects and sitemaps. Which is to say, us.

Here's a useful way to hold it: accessibility and MX pull in the same direction. Real semantic markup, heading hierarchy that doesn't skip a level, alt text that actually describes the image, content that exists in the HTML without waiting on an interaction. Everything we should already be doing for screen readers works exactly the same for a reader that has no eyes for a different reason. The difference is there's a commercial incentive sitting on top of it now. Take it, because accessibility has always needed a business argument to get out of the backlog.

What doesn't work, and why is that exactly what gets sold to you?

I have to ruin the party for a few of the trend lists you're going to read out there.

llms.txt isn't delivering anything. It's a proposed file that tells models what's on your site. Sounds great. In April 2025, Google's John Mueller compared llms.txt to the keywords meta tag, the one that became a joke in the 2000s, and his argument was very practical: why trust a file where the site owner declares what the site is about, when you can just read the site? In July 2025, Gary Illyes said at a Google event that they don't support the format and don't plan to.

Here's a correction I need to make to my own argument before somebody makes it for me: around the same time, Ray Martinez posted his own server logs showing OpenAI fetching llms.txt on his sites every fifteen minutes. That's one practitioner's report on a handful of sites, not a replicated study, but it's enough to knock down "nobody even looks for it". Fetching still isn't the same as using, and evidence of use is what's still missing. In January 2026, Ana Fernández published a follow-up in Search Engine Land tracking 10 sites, 90 days before and 90 days after implementing the file, watching AI crawl and traffic from ChatGPT, Claude, Perplexity and Gemini: 8 out of 10 showed no change at all, and on the ones that grew she traced the cause to other things, like content restructuring and technical fixes. It costs little, if you want to experiment. Just don't put it in the report as a deliverable.

There's no magic schema for showing up in AI answers. Google's own documentation is blunt about it: for your page to be eligible as a supporting link in AI Overviews or AI Mode, it needs to be indexed and eligible to appear with a snippet, and there's no special schema.org structured data you need to add. Structured data still matters a lot, for rich results and for giving explicit context about entities. It just isn't the gateway it's being sold as.

Ranking well on Google no longer guarantees being cited. In March 2026, Ahrefs analyzed 863 thousand keywords and 4 million URLs cited in AI Overviews and found 38% of citations coming from pages that ranked in the organic top 10 for the same query. In their previous study, from July 2025, that number was 76%. Before you run off with that stat under your arm: Ahrefs itself notes that it improved its parsing method between the two studies and now sees more citations than before. They compare the two numbers anyway and attribute the drop to a change in Google's behavior. Here I'm going past what they claim, but measurement that changed midway makes me wary of holding both sides up as the same ruler. Part of the drop may just be seeing better. Even so, the direction is the same: these are two similar games with different rules.

And on "let's post on Reddit because that's where the AI drinks from": Semrush tracked 230 thousand prompts between July and October 2025 and saw ChatGPT cite Reddit in nearly 60% of responses in early August, and around 10% by mid-September. Wikipedia fell from about 55% to under 20% in the same window. Both remained ChatGPT's most-cited sources, which tells you how high the starting point was. The point is the lurch: this changes when a vendor decides to change it, with no notice and no changelog. Building strategy on that number is building on sand.

Where would I start?

This is the part to take to your team. I put it in order of return on effort, because a small project won't do all of it.

1. Server-render whatever needs to be read. If you're on Next.js, Astro, Remix, anything with SSR or static generation, your main content has to be in the HTML of the first response. Test it the dumbest and most reliable way there is: curl your own URL, or "view page source" in the browser, and look for the text of your most important paragraph. If it isn't there, the bot didn't see it either.

2. Answer the question in the first two paragraphs. The model extracts a passage, it doesn't read your whole page lovingly. Suspense is great for an essay and terrible for being cited. Say the thing, then develop it.

3. Turn subheadings into questions. People ask in natural language. An <h2> that restates the question with a paragraph right below that answers it is the most extractable structure there is. Notice this post is doing exactly that.

4. Fix broken links and redirects. A third of these bots' effort goes down the drain on 404s. Run a simple crawler over your site, fix what's broken, make sure old URLs 301 properly, and keep sitemap.xml with an honest lastmod.

5. Semantic hierarchy without decoration. One h1 per page, no skipping from h2 to h4 because the h3 looked wrong at that size. If the problem is the size, fix it in the CSS, which is where size lives.

6. A visible, truthful updated date. There's a lot of talk that recency is a strong signal for the models, and I couldn't find a primary source backing that with a number, so it stays my hunch and not a fact. What I'll defend firmly is the human side: an honest date is the difference between trusting and doubting a technical piece. Revising four good posts a year beats publishing twelve half-finished ones.

7. Structured data on the main types. Person, Article, FAQPage, Organization, validated in the Rich Results Test. Do it for the right reason, which is giving explicit entity context and earning rich results, without expecting a miracle.

8. Measure differently. Keyword position stopped telling the whole story. Once a quarter, ask the AIs themselves the five questions that would lead someone to you, and write down three things: whether you appeared, via which source, and whether it spoke well or badly of you. It's manual, it's incredibly tedious, and it's already more than almost anyone is doing.

One warning about item 8: every model answers differently and changes behavior without telling anyone. Treat any number from this world, including every single one I quoted here, as a snapshot of a moment, with the date stuck to it.

What makes me uncomfortable about all this?

I'll be honest that this part doesn't sit well with me.

We're being pushed to write in a way that's more legible to machines, and the machine's ruler rewards predictable text: question, answer, short paragraph, clean structure. That's excellent for documentation and it's the death of any writing with a personality. If everybody optimizes all the way down, the web becomes one giant FAQ, and the models will train on that, and round it goes.

My personal bet, and here I'm well past what any study supports, is that writing with a voice survives better in the medium term precisely because it's the material you can't compress without losing the part that mattered. The summary of a data point is the data point. The summary of a point of view is a lukewarm version of it. I keep writing with clear structure because it's a kindness to whoever is reading, silicon or flesh. And I keep writing crooked wherever the subject calls for it.

If you only do one thing from this list, do item 1. Open your site's source today and look for your own words in there. Finding out they aren't is the kind of scare that's worth a whole afternoon of work.

A kiss and see you next time!

References

Pew Research Center, Google users are less likely to click on links when an AI summary appears in the results (07/22/2025): https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/
Vercel, The rise of the AI crawler (12/17/2024): https://vercel.com/blog/the-rise-of-the-ai-crawler
Google Search Central, AI Features and Your Website: https://developers.google.com/search/docs/appearance/ai-features
Google Search Central, Guide to optimizing for generative AI features on Google Search: https://developers.google.com/search/docs/fundamentals/ai-optimization-guide
Ahrefs, Louise Linehan, study on AI Overview citations vs. organic ranking (03/2026): https://ahrefs.com/blog/ai-overview-citations-top-10/
Semrush, The Most-Cited Domains in AI: A 3-Month Study (11/10/2025, data from 07/14 to 10/12/2025): https://www.semrush.com/blog/most-cited-domains-ai/
John Mueller on llms.txt, via Search Engine Journal (04/2025): https://www.searchenginejournal.com/google-says-llms-txt-comparable-to-keywords-meta-tag/544804/
Gary Illyes on llms.txt, and the logs showing OpenAI fetching the file, via Search Engine Roundtable (07/2025): https://www.seroundtable.com/openai-crawling-llms-txt-files-39811.html
Ana Fernández, Does llms.txt matter? We tracked 10 sites to find out, Search Engine Land (01/20/2026): https://searchengineland.com/does-llms-txt-matter-467740