A Resident Asks ChatGPT for Your Permit Page. It Links the Worst One.

1What's Actually Happening

A resident used to type a question into a city's own search bar, or a Google query, and land on whatever page the agency built and tested. Increasingly, they ask ChatGPT, Perplexity, or an AI Overview instead, and those tools pick the page for them based on how their models read the web, not based on your site map or your QA checklist.

Our own read on why lines up with the complexity data we've covered elsewhere: pages dense with scannable text, heavy semantic markup, and clear structure are exactly the pages AI models are built to extract information from efficiently. Those same traits don't reliably correlate with the WCAG criteria that make a page usable by a screen reader. A page can be excellent training-data material and still fail basic accessibility checks.

2Why This Is Worse for Government Than for a Retailer

A retailer that loses a sale to a bad AI-surfaced product page loses revenue. A resident who asks an AI assistant how to apply for property tax relief and gets routed to an outdated PDF, or a page with no labeled form fields, doesn't just bounce. They may not find the real answer at all, especially if they're using a screen reader and the page the AI picked can't be navigated with one.

And here's the part that makes this genuinely hard to catch: your own analytics dashboard doesn't show it. A visit that started with a question to ChatGPT, which then linked your site, shows up in your logs looking like any other referral, or sometimes like direct traffic. You can't audit a path you can't see.

3The Compounding Problem

This stacks directly on top of everything else going wrong in 2026. The WebAIM Million found accessibility errors up 10.1% year over year, driven partly by page complexity nobody's tracking against accessibility impact. AI search adoption is accelerating at the same time, for the same complex pages. A page can rank well with an AI assistant and fail a Title II compliance review in the same week, for reasons that have nothing to do with each other and everything to do with the same underlying markup.

4A Concrete Walkthrough

Picture a resident asking an AI assistant how to apply for a building permit in a mid-size city. The model doesn't know the city's site map or which page the webmaster considers the official one. It picks whichever page parses most cleanly: clear headings, well-structured paragraphs, specific detail instead of vague marketing copy. That might be the real permit application page. It might just as easily be a three-year-old news release that mentions permits in passing, structured well enough to rank but wrong enough to send someone down the wrong path entirely.

Now add a screen reader into that same scenario. The resident gets routed to a page that happens to read well to an AI model but was never built with real labels or heading structure for assistive technology. They've been sent to exactly the wrong page, through a channel the city's own analytics can't see, with no way to flag that the AI's answer was wrong unless they happen to report it somewhere the city will actually notice.

5How AI Crawlers Read Your Site Differently Than Google Does

Traditional search engines built decades of ranking signals around links, keywords, and user behavior. AI crawlers care more about whether a page can be cleanly extracted into a self-contained answer: a clear question-and-answer structure, specific facts stated directly rather than implied, minimal reliance on JavaScript to render the actual content. Some sites have started publishing an llms.txt file, a plain-text summary aimed specifically at AI crawlers, as a way to point models toward the pages an organization actually wants surfaced. It's a new lever, and one most government web teams haven't touched yet, for better or worse.

6What to Actually Check

Related Articles