A Third Of Fintech Is Invisible To AI Agents

A study of 274 fintech homepages from the CNBC World's Top Fintech Companies 2025 list found that 36% deliver less than 80% of their homepage content in raw HTML, meaning AI agents that do not render JavaScript cannot see critical information. This gap undermines visibility for AI crawlers like GPTBot and ClaudeBot, which default to raw HTTP fetches without JavaScript execution.

A third of the top fintech websites in the world deliver less than 80% of their homepage content in raw HTML. That is the version of the page an AI agent gets when it visits, before it decides whether to spend the compute on a full browser render. Most of them do not. The Structure pillar of Machine-First Architecture https://www.searchenginejournal.com/machine-first-architecture-how-to-build-websites-machines-can-identify-read-cite-use/574431/ says critical information must not depend on client-side JavaScript. Rendering independence. Until last month, this was a design principle. Now it is a number, and the number is uncomfortable. On May 25, I measured 274 fintech homepages from the CNBC World’s Top Fintech Companies 2025 list. I made two sequential measurements on each one: a raw HTTP fetch with no JavaScript execution, and a full browser render with Playwright. The gap between the two readings is the gap an AI agent has to close on its own. 36% of these websites force it to do that work for the most important page on the property. The full study is published on Web Performance Tools https://webperformancetools.com/research/state-of-agent-visibility-fintech-2026 . Most AI Visibility Coverage Skips The Rendering Step Most AI visibility coverage focuses on schema markup, structured content, brand authority signals https://www.searchenginejournal.com/what-search-engines-trust-now-authority-freshness-first-party-signals/570553/ , and optimization for AI Overviews, ChatGPT search, Perplexity citations, and Gemini’s grounding pipeline. The advice stacks up fast. All of that assumes the agent saw your content in the first place. Most AI crawlers do not render JavaScript. GPTBot, ClaudeBot, PerplexityBot, the AI user-agent landscape https://nohacks.co/blog/ai-user-agents-landscape-2026 feeding the models you are trying to be cited by, they make HTTP fetches and walk away. They are not browsers. Running a real Chromium https://www.searchenginejournal.com/ai-crawlers-account-for-28-of-googlebots-traffic-study-finds/535948/ instance per page costs compute that multiplies across the millions of pages these systems want to read. So they don’t, by default. They take what comes back in the raw HTTP response and move on. There are exceptions. Google’s crawler runs a deferred rendering pipeline for some pages. Some AI systems will render for high-value targets, or render selectively when the raw response looks empty. The pattern is not absolute. But the production default for the crawlers that feed the largest AI systems on the web today is raw HTTP fetch, no JavaScript, take what is there. This creates a gap that users do not see. A real visitor opens your website in a browser. JavaScript runs. The page assembles in the viewport. Content loads, layout settles, the hero image arrives. The visitor sees what you built. The AI agent fetches the response before any of that happens. Whatever does not show up in the first HTML response is, for that agent, not there. This is what the Structure pillar of Machine-First Architecture is about. Critical information must not depend on client-side JavaScript. The page must be parseable from the raw HTTP response, not from the browser-rendered view five seconds later. This is not a performance preference dressed up as architecture. It is a visibility requirement for the AI agents that now read the web on behalf of users https://www.searchenginejournal.com/how-ai-agents-see-your-website-and-how-to-build-for-them/570443/ . Until recently, the rendering-independence requirement was an argument. You could read the spec, look at the crawler behaviors, draw the conclusion, and still disagree about how much of a problem it is in practice. There was no number you could point to. The fintech data gives you the number. Two Reads On The Same Page: Raw HTML, Then Browser Render The test on each of the 274 fintech homepages was simple: two sequential measurements, run on May 25, 2026, from Portugal. The first was a raw HTTP fetch against the canonical homepage, no JavaScript executed, whatever bytes came back in the response. The second was a full browser render using Playwright 1.60.0 with Chromium 148.0.7778.96 in non-headless mode, capturing the page at five seconds post-TTFB and again at network idle. All measurements ran from Portugal on May 25, 2026, on residential broadband, viewport 1280 by 800, no network throttling. For each website, content was extracted from the <main , <article , or <body element and converted to Markdown to preserve structural elements. The raw-fetch text was measured as a percentage of the network-idle text. If the raw fetch returned 80% or more of what the browser eventually rendered, the website passed at full visibility. Between 60% and 79% was partial. Between 30% and 59% was low. Below 30% was near-zero. Three reads on the same page, in the same session, separated by what the browser had to do to make the page complete: raw fetch, five-second render, network idle. The interesting part of the curve is not the network-idle reading at the end. Almost every website in the sample resolves to full content by network idle. The interesting part is the raw-fetch reading at the start, because the raw fetch is the read most AI crawlers actually take. 36% Deliver Less Than 80% Of Their Content Without JavaScript Out of 274 fintech homepages measured, 99 returned less than 80% of their final content from the raw HTTP fetch. That is the headline number. Thirty-six percent. Inside that 99, the distribution is steep. Fifty-five websites 20% of the full sample returned less than 30% of their content without JavaScript. Forty-seven of those websites returned zero. The HTML response carried a shell, the layout scaffolding, some inline scripts, no readable content. Whatever the homepage was meant to communicate required a JavaScript runtime to communicate it. The 47 zero-content websites include major exchanges, well-known neobanks, large lending platforms, several public companies, and brands a person in finance would recognize without prompting. I am not going to call them out individually. Naming the websites would distract from the architectural observation underneath. Whether your homepage shows up to an AI agent at all is currently a function of decisions that nobody on the team was thinking about in those terms when they were made. The 24 websites in the 60-to-79 partial-visibility band have a different problem. They show up to the agent, but not all the way. The agent gets a hero headline, maybe the primary navigation, maybe a value proposition. It does not get the product descriptions, the trust signals, the calls to action, the third-party logos. Whatever was decided to render on the client side is the part the agent does not see, and that part tends to be the part that was made dynamic because someone wanted it to feel interactive. There is a recovery curve, and the recovery curve is where the story sharpens. Of the 274 websites, 273 reach 80%-plus visibility once a real browser renders the page for five seconds. Ninety-nine percent. The content exists. The websites are not broken. They are gated behind a runtime that the production AI crawlers do not pay for. The median website in the sample takes 21 times longer to reach network idle than to return its raw HTTP fetch. Thirty-four websites 12% do not reach network idle at all within the 30-second cap. That is a separate problem, but it points to the same root cause. The cost gap between fetching a website and reading it is widening, and the crawlers cannot keep absorbing the difference. Stripe, Adyen, And Plaid Prove The Stack Is Not The Problem One hundred and one websites in the sample returned 100% of their homepage content in the raw HTTP fetch. Full visibility before any JavaScript ran. The list includes Stripe, Plaid, Adyen, Marqeta, Remitly, Starling Bank, Neo Financial, Backbase, Thought Machine, and 92 others. Fiserv returned a complete homepage in 58 milliseconds. Acorns in 76. Trustly in 89. Ledger in 100. Look at what those websites actually are. Fiserv is a payments-and-banking infrastructure company at a $60 billion scale. Acorns runs a consumer app. Ledger is a hardware-wallet vendor with a product catalog. They are using modern stacks, content management systems, regional CDNs, the works. They have decided that the content the homepage exists to communicate will be there in the raw response, and they have not let the framework choice override that decision. This is the answer to the obvious pushback. The pushback is that a modern stack requires client-side rendering, that single-page applications are how the web is built now, that asking for server-rendered HTML is asking engineering to step back five years. The fintech sample disproves that on its own. The websites running the fastest raw responses took the rendering-independence requirement seriously when they made architectural choices, or rebuilt to take it seriously after the fact. There are exceptions to this read inside the sample. Three websites underperformed at the five-second render window, even though they eventually completed. All three are Asian companies measured from Portugal, and the latency penalty likely accounts for the curves rather than the architecture. The Web Performance Tools study https://webperformancetools.com/research/state-of-agent-visibility-fintech-2026 tested homepages only, from one geographic origin, on one day, with one measurement per website. It did not measure interior pages. It did not test multiple regions. It did not test content gated behind scroll or click. The picture this dataset gives you is the homepage of the biggest fintech companies in the world, on a single day in May, fetched from Western Europe. That is a slice. A useful slice for the load-bearing question this article is about, but a slice. The load-bearing question is whether the rendering-independence requirement holds at scale across a large, modern, well-resourced commercial cohort. The fintech sample answers it. Most of the cohort gets it right. A third gets it wrong, and the third that gets it wrong includes brands big enough that the architectural decisions almost certainly passed through several rounds of senior engineering review without anyone naming AI visibility as a constraint. Fintech Is Where The Homepage Is The Trust Signal Most categories can afford some of their homepage to be invisible. A consumer SaaS company can lose a hero subhead, and most of its visitors will not feel the difference. A media website can carry the masthead through schema and still rank for the topics the body covers. Fintech is not most categories. For a fintech, the homepage is where regulated disclosures live. The licensing footnote. The deposit insurance language. The bank-partner attribution. The security certifications. The country availability matrix. The risk warning under the rate quote. These are the elements that turn a brand from “an interesting product” into “a thing I would actually put money into.” A reader scanning the homepage is looking for them. So is an AI agent answering a question about which provider to trust for a specific use case. When 17% of the cohort returns zero content in the raw response, what disappears is the regulatory and trust layer of the brand. The agent does not see the bank partner. It does not see the deposit insurance. It does not see the security certifications. It sees a shell. The category compounds the problem in a second way. Fintech buying decisions are research-heavy. A person opening a savings account, picking a payment processor, deciding which broker to fund, evaluating a wallet, they go through several rounds of comparison before they act. That comparison loop is the part of the funnel that has migrated into AI surfaces the fastest. Eric van Buskirk’s clickstream study of 846,000 Google sessions https://www.searchenginejournal.com/google-search-sessions-show-how-users-pause-scroll-reconsider-before-clicking/575243/ showed AI Mode users close their loops inside the AI 64% of the time, never clicking through. The fintech research loop is increasingly happening inside an AI surface, and the agent doing the work for the user is choosing from a candidate set assembled out of the raw HTML it could fetch. If a fintech homepage returns zero content in the raw response, the brand never enters the candidate set the agent is choosing from. It is absent before the comparison begins. This is what the Structure pillar of Machine-First Architecture exists for. The pillar is the upstream requirement that makes every downstream AI visibility strategy possible. Schema markup does not help when the agent cannot read the page. Citation strategy https://www.searchenginejournal.com/lessons-learned-from-adobes-2026-q2-ai-traffic-report/574176/ does not help when the model never saw the content to cite. Brand authority signals do not help when the homepage that carries them returns empty bytes to GPTBot. Structure is the floor. Everything else stacks on it. The fintech sample shows the floor is broken for one in three of the biggest brands in the category. The sample is a snapshot. Tomorrow, some of those websites will be different, and some of the websites that scored well today will have drifted in the wrong direction. The 36% will move. What matters is that until last month, there was no number at all, and now the conversation about AI visibility for this category has a load-bearing measurement attached to it. Run The Audit On Your Own Homepage Open Chrome. Open DevTools. Hit Cmd+Shift+P on Mac or Ctrl+Shift+P on Windows. Type “Disable JavaScript” and hit enter. Reload your homepage. What you see is what the agent sees. If your hero, your value proposition, your product description, your trust signals, your CTAs, and your regulatory disclosures are all visible, your homepage is passing the Structure pillar. If the hero is there but the body is gone, you are in the partial-visibility band, somewhere between 60% and 79% of your final content. If the page is blank or close to it, you are in the same tier as the 47 zero-content websites in the fintech sample. This is the cheapest audit in the AI visibility category. It takes 30 seconds. No log-file analysis, no paid tools, no meeting with engineering. The result is binary enough that you do not need to argue about methodology. Either the agent can read your homepage, or it cannot. If the audit fails, the fix paths are framework-specific but well-known. Next.js has server-side rendering https://www.searchenginejournal.com/server-side-vs-client-side-rendering-what-google-recommends/545946/ and static generation, both of which return content in the raw HTTP response. Astro and SvelteKit ship server-rendered by default. React applications can be prerendered route by route using tools like Prerender.io or Cloudflare Pages’ prerendering layer https://www.searchenginejournal.com/all-you-need-to-know-about-cloudflares-agent-readiness-score/574226/ , which serve a snapshot of the rendered page to crawlers without changing the runtime architecture for users. Vue and Angular have equivalent patterns. The choice between these paths is an architecture conversation, not a content conversation. Most teams do not need to rebuild. They need to add a server-rendering layer for a specific set of routes: homepage, pricing, product pages, blog index, anything the brand depends on for first-impression or trust signals. The architectural change does not have to ripple through the entire application. Stripe, Adyen, Plaid, Marqeta, and the 97 other websites in the 100%-raw-visibility list of the fintech sample did not pick simpler stacks. They picked architectures that respected the rendering-independence requirement and shipped the content in the raw response. The pushback that the Structure pillar requires going back to server-rendered PHP from 2009 is the wrong shape of pushback. The pillar requires the content to be there at the HTTP response layer. How you get it there is up to the stack you already have. What This Sample Does Not Tell You The study measured 274 homepages on one day from one origin. Interior pages were out of scope, which means a website with a good homepage and weak product pages would pass the test and still have the visibility problem on the routes that drive conversion. Geographic variation was out of scope. Per-crawler behavior was inferred from the raw-versus-render gap rather than probed directly. Content gated behind scroll or click is its own category of agent visibility failure, and the study did not test it. A question this dataset cannot answer is whether AI crawlers will start rendering more pages as compute costs come down. They might. Some already do for high-value targets. If they all did, the entire framing of this piece changes. The 36% would still be invisible at the raw-fetch layer, but would be readable at the rendered layer, and the urgency of the rendering-independence requirement would soften. I do not think that is going to happen at scale soon, but I am watching for it. Rendering Independence Was Always Real. Now It Has A Number A third of the top fintech websites in the world are partially invisible to AI agents. Rendering independence is no longer a design principle to argue about. It has a number. The agent will be back tomorrow. The crawl will be the same crawl. The HTML will be the same HTML. If your homepage does not return its content in the raw response, the agent that fetched it will be working from a shell, and the answer it gives the user who asked about your category will be assembled from the websites that did return their content. Open DevTools. Disable JavaScript. Reload your homepage. The page that loads is the page the agent saw. More Resources: The Technical SEO Audit Needs A New Layer https://www.searchenginejournal.com/technical-seo-audit-new-layer/571583/ 846,000 Google Searches Reveal How AI Overviews Are Changing User Behavior https://www.searchenginejournal.com/google-search-sessions-show-how-users-pause-scroll-reconsider-before-clicking/575243/ What Is The Agentic Web? https://www.searchenginejournal.com/what-is-the-agentic-web/574224/ This post was originally published on No Hacks. Featured Image: Roman Samborskyi/Shutterstock