# The True Cost of Accessibility | Sherwin Vishesh Jathanna

> Every website serves three users: the one you designed for, the one your design filtered out, and the machine reading it. Which one wins comes down to leverage.

Source: https://www.sherwinvishesh.com/blog/true-cost-of-accessibility  
Copyright 2026 Sherwin Vishesh Jathanna. Text may be quoted with attribution. The design and source code are not licensed for reuse.

By Sherwin Vishesh Jathanna. Published August 9, 2026. 10 min read.

On July 15, 2026, Indian Railways did something it had not seriously done in more than a decade: it let people see a different IRCTC website. The new Beta booking site is faster and dramatically calmer. CAPTCHAs and pop-ups are down, every travel class shows its availability on a single screen instead of demanding a separate click for Sleeper, AC 3-Tier, and AC 2-Tier, and the whole thing wears a glassy, modern skin that would not look out of place on a fintech landing page [1]. The backend work is real too: IRCTC reports completed bookings in the brutal three-minute Tatkal window rose about five percent, and online booking now accounts for 89.84 percent of all reserved tickets [1].

But the detail worth studying is smaller than any of that. In the corner of the new experience sits a link that says "Switch to Old Version" [2].

That link is this entire essay compressed into one UI element. Every website serves three users at once: the person the designers pictured, the person the design quietly filtered out, and the machines that now read the page alongside both of them. What we call accessibility is really the story of which of those three has enough leverage to make anyone care. IRCTC just declined to pick a winner, and that refusal is more honest than most of the industry manages.

## The ugliness was load-bearing

IRCTC's old site has been the internet's favorite punching bag for years. Design students critique it as coursework, and UX writers mine it for essays about broken heuristics: dense forms, cramped spacing, navigation that behaves differently on every page, feedback that arrives late or not at all [3][4]. None of that criticism is wrong. It is just incomplete.

The clutter was never really an accident. IRCTC sells train tickets to one of the largest and most diverse user bases of any website on earth: millions of bookings a day, in more than twenty languages, on everything from this year's flagship phones to the cheapest Android handsets running on 2G in places where 2G is the good news [5]. Its users include people booking their first ticket on their first phone in their first month on the internet, screen-reader users, and grandparents under Tatkal pressure. A private product gets to segment that audience and politely ignore the expensive parts. A railway monopoly does not.

Put IRCTC next to Amtrak's site and IRCTC looks like chaos. Then remember that it serves roughly a hundred times Amtrak's daily users, across far more languages and device tiers, with no option to simply not serve the hardest cases [5]. Design for that population and you learn to prioritize robustness and redundancy over restraint. Whitespace is a luxury enjoyed by designers whose users all have big screens and fast connections.

This is not an India-specific story. CBC, the Canadian public broadcaster, maintains CBC Lite, a deliberately text-only version of its news site built for dial-up and remote readers, citing regulator figures that basic broadband reaches about 87 percent of urban Canadian households but only about 46 percent of rural ones [6]. When you are obligated to reach everyone, you build for the connection your worst-connected reader actually has.

## The third user never had eyes

For most of web history, the excuse ended there. You accommodated the users you could see, plus whichever invisible users the law or your conscience forced into view. Then, over the last three years, a third audience showed up in the server logs.

ChatGPT reportedly reached about 900 million weekly active users by February 2026, roughly double the year before [7]. Gartner's much-quoted prediction that search volume would drop 25 percent by 2026 was a projection, and an imperfect one; search proved stickier than that [8]. The sharper evidence is what happens when Google answers first: analyses of queries with an AI Overview report organic click-through declines of roughly 34 to 61 percent, depending on methodology [7]. The numbers are fuzzy. The trend is not.

So an entire discipline, generative engine optimization or answer engine optimization depending on who is selling it, has appeared to make content legible to machines that answer questions. The mechanics matter: these systems break a question into sub-queries, retrieve specific passages rather than whole pages, and synthesize an answer from whatever comes back [9]. A section that states its point in a self-contained way, under a heading that says what it covers, gets pulled cleanly. A point smeared across nine paragraphs of vibes does not.

Here is the part that should sound familiar. When Vercel measured the major AI crawlers on its network, none of them rendered JavaScript: OpenAI's and Anthropic's bots fetch script files but never execute them, and only Google's Gemini escapes the limitation, by borrowing Googlebot's rendering pipeline [10]. A client-rendered single-page app, the default output of modern frontend culture, is close to invisible to every answer engine except Google's. And since Cloudflare began blocking AI crawlers by default for new domains in July 2025, a site can exclude this audience without anyone having decided to [11].

**In practice:** if your content only exists after JavaScript runs, most AI crawlers never see it. Server-render the words you want found and cited.

The gold rush has its snake oil. llms.txt, a 2024 proposal for a robots.txt-style index of your site written for language models [12], sits at roughly ten percent adoption among top sites [13], much of it because Shopify enabled it for every store without asking [14]. A 300,000-domain analysis found no measurable correlation between having the file and being cited by AI systems, and Google's John Mueller has noted that no AI crawler has confirmed using it [15][16]. Where it demonstrably works today is coding agents, which fetch a library's llms.txt when pointed at its documentation [17]. Machine legibility is real. Not every product sold in its name is.

## Three users, one interface

So name them properly.

User one is the user you can see. Modern phone, decent bandwidth, full vision and motor control, patience for animation and brand personality. Your analytics describe this person, your personas are headshots of this person, and roughly fifteen years of modern web design, the whole JavaScript-framework, custom-type, whitespace-and-delight era, was built almost exclusively for this person, even by teams whose mission statements say "for everyone."

User two is the user your design filtered out. Screen-reader users, motor-impaired users, people on 2G or a five-year-old Android, first-time and low-literacy internet users, older users. Someone whose connection never finishes loading your JavaScript bundle bounces before a single analytics event fires.

*"The user you excluded does not show up in your analytics as a bad number. They show up as no number at all, which is why nobody ever gets fired for losing them."*

User three was never human. Search crawlers first, and now the retrieval systems feeding answer engines. The twist inside the twist: user three's technical wishlist overlaps, to a suspicious degree, with user two's. Semantic headings instead of styled divs. Content in the HTML instead of behind a script. Short passages that stand alone. The two audiences are not the same: a modern screen reader runs on top of a full browser and copes with a JavaScript-built page far better than a crawler that never renders one, and passage-level writing is a retrieval preference, not a WCAG criterion. Their needs differ. Their incentives just landed in the same place: structure, semantics, and content that exists before the scripts arrive. So the machines are now dragging web design back toward what disability advocates were asking for two decades ago, for reasons that owe nothing to conscience.

## GOV.UK ruins the excuse

At this point the argument seems to write itself: accessibility gets you IRCTC, design gets you exclusion, pick your poison. One website has spent well over a decade quietly proving that framing wrong.

GOV.UK launched in 2012 as the single front door to the British government, and in 2013 the Design Museum named it Design of the Year, an award a website consisting mostly of black text on a white background took home over buildings and furniture [18]. It is fast on bad connections, works without JavaScript, reads cleanly through a screen reader, and is written in plain language that a 2026 roundup of government sites credits with beating almost every comparable public and private sector site for readability [19]. Behind it sits a design system of reusable components, built and tested for accessibility before anyone ships them and reused across a thousand-plus departments and services [20].

GOV.UK is what IRCTC's constraints look like when they are treated as a design brief instead of an excuse. Serve everyone, on every device, in plain language, under legal obligation. Same brief, radically different output. Which tells you the clutter never actually came from the constraints.

It comes from institutions. Too many stakeholders bolting one more banner onto a homepage nobody owns. Features that accrete because removing anything requires a meeting nobody wants to call. No team empowered to say no. "We have to support old phones and screen readers" becomes a sympathetic shield for unmanaged information architecture. Supporting a screen reader has never once required a blinking marquee.

A side note, because the design world keeps making the same mistake in the opposite direction: brutalist-looking sites are not automatically accessible either. Craigslist's plainness serves it well, but clashing colors and chaotic layouts punish low-vision users as hard as any glassmorphism, while merely looking authentic. Ugly is not a feature. It is also not the point.

Four sites, four answers to the same question:

- **IRCTC, pre-Beta:** Built for everyone in India by obligation; constrained by extreme device, language, and literacy diversity; dense, redundant, form-heavy; weak machine legibility, with tangled markup and CAPTCHAs everywhere
- **IRCTC Beta:** Built for the modern-phone majority; constrained by Tatkal-window throughput; glassmorphism and consolidated screens; machine legibility improving, still iterating in public
- **GOV.UK:** Built for every UK resident by design; constrained by accessibility as a component contract; minimal, typographic, almost austere; excellent machine legibility, semantic HTML with no JS required
- **CBC Lite:** Built for rural and low-bandwidth readers; constrained by dial-up era connections; text-only and nearly unstyled; excellent machine legibility, mostly by accident

The pattern across all four is the same. The visual style is downstream of who the site decided it had to serve, and machine legibility lands wherever structure lands.

## Follow the leverage

"How accessible can we afford to make this?" is the question most teams actually ask, and it is the wrong question, because it assumes accessibility and quality trade off. GOV.UK is the existence proof that they do not. The honest question, the one that actually predicts what institutions build, is: which of our users can hurt us if we ignore them?

Look at when accessibility budgets actually appear. The European Accessibility Act became enforceable on June 28, 2025, requiring accessibility from e-commerce, banking, and transport ticketing sold to EU consumers, including by companies outside the EU [21], and the first lawsuits under it were filed in France that November [22]. In the United States, the Department of Justice's ADA Title II rule made WCAG 2.1 AA the legal standard for state and local government websites; in April 2026, days before the first compliance date arrived, the DOJ pushed the deadlines back to 2027 and 2028, over the public objection of disability advocacy groups like AAPD [23][24]. Notice the shape of that story. Decades of moral argument moved roadmaps almost nowhere, then enforcement dates and court filings moved them immediately, and when the deadline slipped, so did the work.

Now watch the same industry respond to the machine user. No lawsuit compels anyone to care about AI crawlers. Ignoring them simply costs traffic, which costs revenue, starting next quarter. So the pitch has flipped: server-side rendering, semantic markup, and structured data, sold for years as accessibility virtue and deprioritized accordingly, are now sold as revenue protection, and suddenly the audits get commissioned [9][10]. User two waited twenty years for leverage. User three walked in holding it.

IRCTC fits the pattern with uncomfortable precision. The old site was not left unrebuilt for two decades out of malice. It was left because a monopoly faces almost no consequence for making people suffer through it; every ticket gets bought anyway. Then the pain reached someone with power: the Railway Minister reportedly heard directly from students in Rajasthan about how bad booking had become, and a rebuild that had been technically possible the whole time shipped within months, inside a 52-item reform package [25]. What changed was not capability. What changed was consequence.

And the constructive answer was sitting inside India's train-booking ecosystem the whole time. IRCTC authorizes a market of third-party apps, ConfirmTkt, ixigo, Paytm, MakeMyTrip, that layer polished, specialized interfaces over its universal backend and charge a small convenience fee for the comfort [26]. That is the actual resolution of the three-user fight. Nobody forces one interface to serve every audience at once. The maximally inclusive version becomes the substrate, the thing that works for everyone and everything, and specialized layers serve the users who want more and will pay for it. The substrate is not the degraded experience. The substrate is the product. Everything else is skin.

## The receipt

Which brings us back to that little link. "Switch to Old Version" is what it looks like when an institution declines to pretend there is only one user. The Beta courts the users whose frustration had finally become politically expensive: the young, the impatient, the ones with a minister's ear. The old version keeps faith with the users who had finally learned where every confusing button lived, on devices the new CSS may never love. Both interfaces, one backend. For a while at least, nobody loses quietly.

**The short version:** accessible and well-designed were never opposites; GOV.UK settled that in 2013. What actually shapes an interface is which users hold leverage over its owners. Lawsuits finally gave disabled users some. Revenue hands AI crawlers plenty. Design for the user with no leverage at all, and the other two turn out to be covered.

Accessibility was never a tax you pay on top of good design. It is the receipt: an itemized record of exactly who a product's owners decided they could afford to ignore, and who they could not. IRCTC's receipt just got shorter. Yours is being read by more parties than ever. The right question was never how accessible you can afford to be. It is who can make you feel it when you get this wrong, and if you answer that honestly, you already know what your homepage looks like.

## References

1. NE India Broadcast, ["IRCTC Beta Website: A Modern, User-Centric Digital Transformation"](https://neindiabroadcast.com/2026/08/03/irctc-beta-website-a-modern-user-centric-digital-transformation/), August 2026.
2. NativePlanet, ["IRCTC Website Revamp 2026: New Features, Faster Tatkal Bookings"](https://www.nativeplanet.com/news/irctc-website-revamp-2026-new-features-faster-tatkal-bookings-011-22883.html), 2026.
3. Pratt Institute IXD, ["Design Critique: IRCTC Website"](https://ixd.prattsi.org/2025/02/design-critique-irctc-website/), February 2025.
4. A. Dennis, ["What's Wrong with IRCTC's UX? A Deep Dive into 6 Broken Design Principles"](https://medium.com/@alvindennis80/whats-wrong-with-irctc-s-ux-a-deep-dive-into-6-broken-design-principles-9fee563d09fa), Medium.
5. Design Bootcamp, ["Why Indian Websites Look Ugly, and Why That Might Be Great UX"](https://medium.com/design-bootcamp/why-indian-websites-look-ugly-and-why-that-might-be-great-ux-6b847d5907c9), Medium.
6. CBC News, ["CBC rolls out low-bandwidth version of its news site"](https://amp.cbc.ca/news/canada/north/cbc-rolls-out-low-bandwidth-site-1.5950603), 2021.
7. The Publive, ["Generative AI Content Optimization in 2026"](https://www.thepublive.com/feeds/blog/generative-ai-content-optimization-2026), 2026.
8. Gartner, ["Gartner Predicts Search Engine Volume Will Drop 25% by 2026, Due to AI Chatbots and Other Virtual Agents"](https://www.gartner.com/en/newsroom/press-releases/2024-02-19-gartner-predicts-search-engine-volume-will-drop-25-percent-by-2026-due-to-ai-chatbots-and-other-virtual-agents), February 2024.
9. LLM Refs, ["Generative Engine Optimization: The Complete Guide"](https://llmrefs.com/generative-engine-optimization), 2026.
10. Vercel, ["The Rise of the AI Crawler"](https://vercel.com/blog/the-rise-of-the-ai-crawler), December 2024.
11. Cloudflare, ["Cloudflare Just Changed How AI Crawlers Scrape the Internet-at-Large"](https://www.cloudflare.com/press/press-releases/2025/cloudflare-just-changed-how-ai-crawlers-scrape-the-internet-at-large/), July 2025.
12. J. Howard, ["The /llms.txt file"](https://llmstxt.org/), llmstxt.org, September 2024.
13. Rankability, ["llms.txt Adoption Data"](https://www.rankability.com/data/llms-txt-adoption/), 2026.
14. C. Roberts, ["The State of llms.txt Adoption"](https://caseyrb.com/blog/state-of-llms-txt-adoption/), 2026.
15. OrganikPI, ["llms.txt Adoption and Impact"](https://organikpi.com/blog/distribution/llms-txt-adoption-impact/), 2026.
16. LinkBuildingHQ, ["Should Websites Implement llms.txt in 2026?"](https://www.linkbuildinghq.com/blog/should-websites-implement-llms-txt-in-2026/), 2026.
17. Limy, ["llms.txt in 2026: The Full Guide"](https://limy.ai/blog/llms.txt-in-2026-the-full-guide), 2026.
18. Government Digital Service, ["GOV.UK wins Design of the Year 2013"](https://gds.blog.gov.uk/2013/04/17/gov-uk-wins-design-of-the-year-2013/), April 2013.
19. Shadow Digital, ["Best Government Website Design Examples"](https://www.shadowdigital.cc/resources/best-government-website-design-examples), 2026.
20. Digital Government Hub, ["GOV.UK Design System"](https://digitalgovernmenthub.org/examples/gov-uk-design-system/).
21. European Union, ["Directive (EU) 2019/882 on the accessibility requirements for products and services (European Accessibility Act)"](https://eur-lex.europa.eu/eli/dir/2019/882/oj), EUR-Lex, 2019.
22. Accessibility.Works, ["European Accessibility Act Compliance Guide"](https://www.accessibility.works/european-accessibility-act/), 2025.
23. U.S. Department of Justice, ["Fact Sheet: New Rule on the Accessibility of Web Content and Mobile Apps Provided by State and Local Governments"](https://www.ada.gov/resources/2024-03-08-web-rule/), ADA.gov, updated April 2026.
24. American Association of People with Disabilities, ["AAPD Statement on the DOJ Title II Web Rule Interim Final Rule"](https://www.aapd.com/aapd-statement-title-ii-doj-web-rule-ifr/), 2026.
25. Gizbot, ["IRCTC's New Website Features Could Change How You Book Train Tickets"](https://www.gizbot.com/internet/features/irctc-launch-new-website-features-could-change-how-you-book-train-tickets-127365.html), 2026.
26. TrainApps, ["ConfirmTkt vs ixigo: Which Train Booking App Is Better?"](https://trainapps.in/comparison/confirmtkt-vs-ixigo), 2026.