TL;DR
AI doesn't see your site the way people do: GPTBot, ClaudeBot and PerplexityBot don't open a browser or run client-side JavaScript. They only read the raw code, the initial HTML.
Invisible trust signals: If your reviews, customer ratings, certifications or guarantees load through third-party JavaScript widgets or sit in images with no text, AI can't see them and leaves them out of its answers.
Three layers of AI visibility: Being mentioned isn't enough. The tone AI uses when it talks about you (sentiment) matters just as much, and so does the factual accuracy of what it says.
One investment, two returns: Making your site readable for AI also makes it faster and clearer for people, and that lifts conversion rates.
Being first in AI answers isn't enough if AI gets you wrong. Visibility, sentiment and claim accuracy are three different things.
Three different things hiding behind one word
When companies talk about AI visibility, most of them measure one thing: whether AI mentions them at all. You ask ChatGPT for the best online store in a category and you're either in the answer or you're not. Binary. Done.
But that's only the first layer. Underneath it are two more that most GEO strategies ignore.
Sentiment: what tone does AI take when it talks about you? Recommending? Neutral? Cautious?
Claim accuracy: what exactly does AI say about you? Does it match the reality of your company or product?
A company can rank first in AI answers while AI describes it coldly or inaccurately. The cause is simple: the trust signals that would change the tone and the content exist on its website, but AI can't read them.
We run into this problem again and again. It's also why our AI visibility and GEO service always measures all three layers, not just visibility.
Why can't AI see what you see?
AI crawlers work differently from a browser. GPTBot and ChatGPT-User (OpenAI), ClaudeBot (Anthropic) and PerplexityBot don't load a page the way Chrome does, and they don't run JavaScript. They download JavaScript files but never execute them (Vercel, The rise of the AI crawler, 2024). They load the initial HTML and read the content from it. Anything rendered by client-side JavaScript simply doesn't exist for them.
The exceptions are Google and Apple: Gemini and AI Overviews run on Googlebot's infrastructure, which executes JavaScript, and AppleBot renders JavaScript too. ChatGPT, Claude and Perplexity don't work this way.
And AI crawlers visit online stores in large numbers: almost half of all AI bot traffic on Akamai's network in the second half of 2025 went to e-commerce, and the most frequent bots came from OpenAI, ByteDance and Anthropic (Akamai, State of the Internet, 2026).
The result? Many modern websites look completely different to AI crawlers than to a real visitor in a browser:
- Customer ratings in a third-party JS widget → an empty space for AI
- Social proof (certifications, awards, badges) in images with no text description → invisible
- Reviews in a JavaScript-rendered carousel → AI can't see them
- Trustpilot scores, Google Shopping ratings → if they load through a third-party JS widget, AI won't read them
That doesn't mean the website is badly built or that the company is doing something wrong. It means the trust signals exist only for people. AI may find reviews from Trustpilot or Google directly on those platforms. On your own website, though, you decide whether it sees them in your context, next to your products.
That's also why we build on a headless CMS with full server-side rendering: the websites we deliver look the same to AI crawlers as they do to visitors.
When your strongest trust signals are hidden behind JavaScript, AI won't use them in its answers, because it can't see them.
A real-world paradox: top visibility, sentiment below potential
The client was among the most visible retailers in its category in AI answers and had the highest Share of Voice of all the brands we tracked. Even so, positive sentiment was lower than its actual reputation warranted, and some of what AI said about the store wasn't accurate.
We started with a technical audit. We looked at the initial HTML the way an AI crawler sees it, not the way a browser renders it. The difference was significant.
Customer ratings load through a JavaScript widget. They're one of the strongest trust signals in e-commerce, yet for an AI crawler they don't exist, while visitors see them clearly. Social proof in images has no text equivalent, and the alt text is just generic file names.
So the trust signals are on the website, just not in a form a machine can read. AI fills in the rest from other sources the client doesn't control. That's where the weaker tone and the inaccurate claims come from.
What the audit recommended: a text layer for ratings directly in the HTML, external reviews with a server-side text equivalent and AggregateRating schema, specific alt text instead of file names, and key content delivered through server-side rendering instead of client-side JavaScript. The client is currently switching e-commerce platforms, so these fixes are going into the brief for the new solution. We'll measure their effect against a baseline recorded before the switch.
The mechanism is straightforward: once AI can read a rating as text in the HTML, backed by the right schema, it can use it in its answers. Until then, it can't, because it doesn't see it.
We measure the result by how the wording of AI answers changes. The goal is for AI to describe the client accurately, with a sentiment that matches its real quality and reputation.
Speed and AI readability: pay once, benefit twice
Speed is measurable. In an A/B test, Vodafone made its main content render 31% faster and sales rose 8%. The two versions differed only in speed (web.dev, 2021).
An online store that loads in one second has a 2.5x higher conversion rate than one that loads in five (Portent, 2022, more than 100 million page views across 20 B2B and B2C websites).
Those numbers are about people and how quickly they lose patience. What's interesting is that the same changes that make a page faster also increase how much of it AI can read. The reason is the same in both cases: the less content is filled in after the page loads, the more of it both the visitor and the crawler get right away.
So website performance and AI visibility aren't two separate budget lines. They come down to one technical decision that pays off twice: once in your conversion rate, and once in what AI can say about you.
How to make trust signals machine-readable
Specific fixes that directly affect what AI reads about you.
Ratings as HTML text
If your reviews only appear through a third-party JavaScript widget, AI can't see them. Add an HTML text equivalent with your overall score, the number of reviews and AggregateRating schema. This isn't a technical detail in the background. It decides whether AI can quote your actual score and review count in its answers.
Schema markup for reviews and your organization
JSON-LD schema for Review, AggregateRating and Organization gives AI crawlers direct, readable data. But schema isn't a way to persuade AI to use a particular claim. It helps systems understand what the data on the page means. That's why it has to match what visitors see. If your JSON-LD says 4.9 and the page shows 4.7, you're sending a conflicting signal. We go into more detail in our case study on how AI started recommending us, which covers building AI authority from the ground up and how AI started mentioning us on the topics where we wanted to be visible.
Social proof: from images to text
Certifications, awards, partner badges: each one needs meaningful alt text and, ideally, a text equivalent in the HTML. A file called certificate.jpg tells AI nothing. Alt text that names who issued the certificate, what it covers and how long it's valid tells it everything.
Server-side rendering for key content
If your strongest selling point sits behind JavaScript, AI can't see it. Ratings, benefits, key facts about your company: render them on the server, not through client-side JS. JavaScript isn't the problem in itself. The problem is when the content that should drive a purchase depends on it.
llms.txt and consistent claims across your website
llms.txt is a proposed format that lets a website give language models a structured overview of itself. As of September 2026, none of the major providers has confirmed reading it when generating answers, so treat it as a cheap bet, not a measure with proven impact.
Consistency has a more verifiable effect: if one page says 15 years of experience and another says 10, AI will pick one number or leave the claim out.
Check it yourself: what AI can read on your website
Quickest rough check: view the page source (Ctrl+U on Windows and Linux, Cmd+Option+U on Mac). If most of the content is replaced by empty divs and JS calls, AI sees the same thing you do. Empty space.
Viewing the source in a browser can be misleading, though. Two other tests are more reliable:
- Fetch the page without JavaScript (for example with curl) and see what content is actually left in the response. That's exactly what an AI crawler has to work with.
- Enter your URL in Google Rich Results Test and check that your schema actually loads.
Beyond your website, check external sources too: do key facts (years in business, rating, prices) match on your Google Business Profile, Trustpilot and in business directories? When sources disagree, AI picks one figure or drops the claim.
These tests tell you what AI can read on your website. The other half of the picture is what it actually says about you in its answers, and that's harder to check by hand. That's why we built the free AI Visibility Test: 15 questions from your industry across ChatGPT, Perplexity and Google AI Overviews, with a report within two business days. You'll see who AI names instead of you.
How we measure what AI says about you
A GEO audit starts with two questions at once, not one.
The first is about visibility: does AI mention you? On which topics? How consistently across different tools?
The second is about content: when AI mentions you, what exactly does it say? Is it positive, neutral or cautious? Does it match what you want to say about your company?
To answer both, we combine manual testing with technical analysis. We ask ChatGPT, Perplexity and Google AI Overviews questions on relevant topics, compare the answers with what's in the initial HTML and pinpoint the gaps. The whole process is part of our AI visibility and GEO service.
Once the changes are live, we compare: has the wording of AI answers changed? Are there more accurate claims? Have the inaccuracies disappeared?
How AI describes you before and after the changes is a concrete, trackable metric. Sentiment and claim accuracy can be tracked systematically, and the results can be measured and documented.
What this means for your website
If your website drives revenue directly and more of your customers search through AI tools, three things are worth knowing:
- Are you visible in AI at all? This is the first question, but not the only one.
- What does AI say about you? Is it accurate? Does the tone match your reputation?
- Where are your trust signals? Are they in readable HTML text, or only in images and JavaScript widgets?
Our free AI Visibility Test answers the first question, shows who appears in AI answers instead of you and compares what AI says about you with what's on your website.
An AI visibility and GEO audit answers all three: it adds sentiment, a technical analysis of your website and a plan for what to fix and in what order.
FAQ
What is AI visibility?
A website's ability to appear in answers from AI tools like ChatGPT, Perplexity and Google AI Overviews, and to be described accurately when it does. It has three layers: whether AI mentions you, the tone it uses when it talks about you, and whether what it says about you is true.
Why can't AI see content I can see on my website?
Most AI crawlers, such as GPTBot, ClaudeBot and PerplexityBot, read the initial HTML and don't run client-side JavaScript. Anything rendered in the browser doesn't exist for them: reviews in a widget, a ratings carousel, data from third-party scripts. Google is the exception: its crawler does execute JavaScript.
How quickly can I find out what AI reads on my website?
In a few minutes: fetch the page without JavaScript (for example with curl) and see how much of the content is left. Use Rich Results Test to check your schema. For a systematic answer across tools and topics, you need a GEO audit.
Does llms.txt improve visibility in AI answers?
Not proven yet. The format exists and is cheap to set up, but none of the major providers has confirmed using it to generate answers. A safer bet is content available in the initial HTML and structured data that matches what's on the page.
How often should I measure AI visibility?
At least quarterly, and monthly in competitive categories. AI answers change faster than traditional search results, so a one-off test is a snapshot, not a trend.
