There is a quiet assumption behind most website decisions: that what you see is what everyone sees.
You open the site on your phone. The images fade in, the menu slides out, the service cards populate, the reviews carousel starts turning. It looks professional. It works.
Then a machine visits, and gets something else entirely.
In our companion article, we covered the measured evidence that AI answers are absorbing a growing share of customer discovery. This one addresses the mechanical question underneath it: when those systems read your website, what do they actually receive?
For a large number of business websites, the answer is: far less than you would expect.
The Finding Most Business Owners Have Never Heard
Engineers at Vercel, working with the analytics firm MERJ, studied how AI crawlers behave by analyzing real traffic across their network — hundreds of millions of requests from the major AI systems.
The headline finding is blunt: the major dedicated AI crawlers do not execute JavaScript.
OpenAI’s crawlers, Anthropic’s ClaudeBot, Perplexity’s crawler, Meta’s, and ByteDance’s all retrieve the raw source of a page and read it as-is. They do not run the interactive code that assembles the page you see in a browser.
They will download JavaScript files without executing them — the study measured this at 11.5% of ChatGPT’s fetches and 23.8% of Claude’s — but downloading is not running.
There is one significant exception. Google’s Gemini inherits Googlebot’s rendering infrastructure and does process JavaScript fully. Apple’s crawler renders as well. So the picture is mixed, not uniform.
But the practical implication is direct: if your content only appears after JavaScript runs, several of the systems now recommending businesses cannot see that content at all.
Why This Produces a Website That Is Invisible in a Specific Way
The problem is that this failure is silent, and it does not look like a failure.
Consider what happens with a typical modern business website built on a heavy platform with interactive components:
| What a customer sees in a browser | What a non-rendering AI crawler receives |
|---|---|
| Six service cards, populated from a database | An empty container element |
| “Serving Tacoma, Lakewood, Puyallup and Gig Harbor” in a slider | Nothing — the slider had not run yet |
| A reviews carousel with 40 five-star ratings | An empty div |
| Pricing revealed after clicking a tab | Absent |
| Phone number rendered by a click-to-call widget | Frequently absent |
Everything that matters commercially — what you do, where you do it, who vouches for you, how to reach you — can be exactly the content that never arrives.
The site is not broken. It is not penalized. It does not throw an error anyone will report to you. It simply arrives as a mostly empty shell, and the AI describes your business from whatever else it can find: a directory listing from 2019, a competitor’s comparison page, or a review site with your old address.
The same study found another detail worth noting. AI crawlers are considerably less efficient than Google’s: roughly 34% of ChatGPT’s and Claude’s fetches in that dataset landed on pages that returned “not found,” versus about 8% for Googlebot. They are working with a rougher map of your site than Google is. That is all the more reason not to make the pages they do reach harder to read than necessary.
One honest caveat on this research: the Vercel/MERJ study was published in December 2024. It remains the most rigorous public analysis with disclosed methodology that we have found, and later reporting consistently points the same direction, but the underlying measurements are not fresh. Crawler behavior can change. We would rather tell you the age of the evidence than present it as current fact.
A Detour Worth Taking: What Does Not Work
There is a great deal of confident advice circulating about AI search, and some of it does not survive contact with evidence.
The clearest example is a file called llms.txt — a proposed standard where a site publishes a plain-text summary at its root for AI systems to read. It has been promoted across the SEO industry as an AI-visibility lever, ours included.
The evidence does not support it as a visibility tactic:
- Google has stated on the record that it does not use llms.txt and has no plans to. Google’s own search team has compared it to the long-discredited keywords meta tag.
- No major AI provider has publicly committed to reading it as a ranking or citation signal.
- Server-log analyses consistently find the file is almost never requested by the crawlers that actually drive citations.
- Adoption studies across large domain samples have found no measurable correlation between publishing the file and being cited more often by AI systems.
The file is harmless and cheap to publish. It has legitimate uses in developer documentation and agent-to-agent workflows — Anthropic and OpenAI both maintain such files for their own developer tooling. But as a lever to make an AI assistant recommend your plumbing company, the data does not support it.
We would rather revise our own position than repeat a claim we cannot defend. Structured data and schema markup sit in a similar category — plausible, widely recommended, but with genuinely mixed evidence about whether they influence AI citation. We treat both as reasonable hygiene worth doing, not as guaranteed levers worth promising.
If a vendor sells you a checklist of files as the path to AI visibility, ask them for the measurement. The honest answer is that the fundamentals do far more work than the novelties.
What Actually Holds Up
Strip out the tactics that cannot be evidenced and what remains is unglamorous and durable.
Content that exists in the page itself. If the words are present in the source when the page is requested — rather than assembled afterward by code — every system can read them, including the ones that do not render. This is the structural decision that matters most, and it is why we build on architectures that deliver finished pages rather than instructions for assembling them.
Saying the specific thing plainly. AI systems summarize. They do well with clear, declarative statements of fact — what you do, where you work, what you charge, what you are licensed for, how quickly you respond. They do poorly with vague brand language. “Excellence in customer service since 1998” tells a machine nothing. “Licensed and bonded electrical contractor serving Pierce County, with same-day emergency service” is something a machine can repeat accurately.
Answering real questions in real words. The queries reaching these systems are full sentences, not keywords. Pages that directly answer the actual questions customers ask — in the customer’s phrasing — give the systems something quotable.
Consistency everywhere your name appears. AI systems assemble a picture from many sources. When your address, hours, and service area disagree across your website, your map listing, and the directories, the machine resolves that conflict on its own, and not always in your favor.
Speed, because it is not free. Crawlers operate under time and resource limits. A page that takes too long to respond may simply not be waited for.
None of this is a trick. It is the same principle that has always governed being found: be clear, be consistent, be fast, and make it easy to understand what you do. The audience simply expanded to include machines that will not fill in gaps politely.
How to Check Your Own Site in Five Minutes
You do not need a vendor to get a rough read on this.
- Open your website in a browser. Right-click and choose “View Page Source.” That is roughly what a non-rendering crawler receives.
- Search that source for your phone number, your city names, your main services, and a line from your customer reviews.
- Note what is missing. Anything you cannot find in the source is content that several AI systems likely never saw.
If the source view is mostly script tags and empty containers, that is the finding. It is common, it is fixable, and it is not a reflection of anyone’s competence — it is a consequence of how most website platforms were designed in an era when only human eyes mattered.
The Underlying Shift
For thirty years, websites were built for people, and search engines adapted themselves to read what people read.
That relationship has partially inverted. A meaningful share of your prospective customers now encounter your business through a machine’s summary before they ever see a page you designed. The machine is not hostile and it is not clever. It reads what it is given, and it describes what it found.
The businesses that will do well here are not the ones chasing every new acronym. They are the ones whose websites plainly state true, specific, consistent things in a form any reader can parse — human or otherwise.
If you would like to know what these systems currently see and say about your business, we will take a look and tell you plainly — including if the answer is that nothing needs changing.
