Key takeaways
- Ranking well in Google says nothing about whether an AI engine can read, trust, or cite a site; they're different systems with different rules.
- A perfect robots.txt file means nothing if a CDN's bot settings are silently blocking GPTBot, ClaudeBot, or PerplexityBot at the edge.
- Schema makes a site eligible to be cited; the actual sentences on the page are what get it chosen.
- AI tooling has made writing code nearly free, but it hasn't touched the cost of understanding it, and skipping that cost quietly erodes a developer's debugging skill.
Most agencies are still reporting on rankings while clients quietly lose the conversation that decides who gets picked.
That's the disconnect Chandan Sharma, Founder of Doodle Web, set out to name on a recent episode of the Goodfirms Podcast. Chandan has spent years in digital strategy and web architecture, including WCAG 2.1 AA and ADA Title II accessibility work, and more recently auditing how prepared institutional websites actually are for AI crawlers. The short answer: almost none of them.
Who Is Chandan Sharma?Chandan Sharma is the Founder of DoodleWeb, a web design and development agency building on Drupal, WordPress, Webflow, and Shopify. On this episode of the Goodfirms Podcast: Conversations That Matter, he breaks down why sites built for 2018-era SEO are often unreadable to AI engines, what his firm found auditing 150+ university websites, and why he built SiteSonar to track what Google Search Console never could. |
Two Systems, Two Rulebooks
"Agencies think if a site ranks well in Google, it's automatically fine for AI. Those are two different systems with different rules," Chandan said. Traditional search indexes a site and sends people there; AI systems read it, extract an answer, and often send no one at all. What matters to an LLM is content structured so a machine can lift an answer: consistent entity facts, validated schema, a specific question answered specifically. "You're no longer just optimizing a destination, you're optimizing whether you exist in the answer," he said.
Most agencies are still building for 2018. The old playbook produces reports that look fine even as results decouple, traffic down 20 to 30%, rankings unmoved, leads flat or up, because AI is answering the top of the funnel. The real wake-up call is usually a client asking, "why does ChatGPT recommend our competitor and not us?" And when an AI engine answers directly, there's no click left to measure. "The click was always a proxy, what clients actually buy is being chosen," Chandan said, which is why his tool, SiteSonar, runs real prompts across ChatGPT, Claude, Perplexity, and Gemini and tracks visibility share over time instead.
What 150 University Websites Revealed
Chandan's team audited 150+ accredited university websites for the basics: crawler access, valid structured data, entity consistency across the web. Roughly 1% had the groundwork in place. The rest failed through accumulation: security vendors blocking GPTBot by default, plugin-generated schema never validated, key answers buried in PDFs, directories whose canonical tags quietly handed authority to a third-party domain. "These institutions spend millions on their web presence and are invisible in the channel where prospective students increasingly ask their first questions," he said.
Before any content fix matters, crawlers need to reach the site at all. Chandan checks robots.txt for GPTBot, ClaudeBot, and PerplexityBot, then the CDN, since Cloudflare's managed settings have been seen silently overriding an explicit crawler opt-in. "You can have a perfect robots.txt and still be blocking everything at the edge," he said.
Structure Gets You Eligible, Content Gets You Chosen
"Structure gets you eligible and content gets you chosen, and content is the bigger half," Chandan said. Schema helps a machine parse who a brand is, but sentences get cited, not schema. He's seen flawless, schema-rich sites never referenced because the copy is all adjectives, and mediocre pages cited constantly because one paragraph answers a question dead-on. His test for any page: could a machine reading it alone answer who this is, what they do, and how much it costs? "Markup describes the answer; it can't substitute for the answer."
The same logic touches his accessibility work: a screen reader and an LLM crawler both reconstruct meaning from structure, so WCAG remediation often makes a site more machine-readable too. But passing WCAG 2.1 AA doesn't guarantee AI visibility. "It's about a third of the job wearing the costume of the whole job," he said.
Unstable Citations Are Still Measurable
Ask an AI model the same question twice and the citation can change. Chandan treats that the way SEO once treated ranking fluctuation. "Rankings were never stable either, and the industry solved that with tracking tools, trend lines, and averages," he said. He sets three expectations with clients: the unit is a percentage of relevant answers, not a guaranteed citation; the timeline moves in months, not days; and the volatility is the opportunity, since no brand owns position one in an answer redrawn every time someone asks.
Overpromising says 'we'll get you cited.' Underpromising says 'it can't be measured.' The truth is in between: it's a probability you can track, benchmark against competitors, and demonstrably improve.
What AI Is Doing to Developers, and the New Service Line
Chandan tried building a marketing site with AI tooling himself; one animation took 50 prompt variations, and the fix broke two other components. "AI tooling collapses the cost of producing code but not the cost of owning it," he said. The sharper risk is to debugging itself, the loop of tracing a failure and understanding the system better, which AI short-circuits, leaving developers unblocked but learning nothing. His tell for agency owners: a shipped fix with a vague "why did that break," or the same components reopening in the bug tracker.
His advice for agencies building an AEO practice follows the same discipline: developers learn to audit machine access, writers add declarative writing to their range, and someone owns entity consistency across the web as a maintained asset. "You cannot sell what you cannot measure," he said. Firms exploring this can look at how established web design companies and SEO companies on Goodfirms are already positioning AI-readiness audits as a service of their own. He expects AI visibility to split off from SEO as its own budget line within 18 months. "The agencies that move in the next twelve months get to define what this service costs."
Rapid Fire With Chandan Sharma
- Bigger blocker to AI citation, the technical build, or the content itself? The content. The page gets you read, but the content gets you quoted.
- A website is invisible to AI in 2026 if it…? If it says how it feels instead of what it is. Adjectives don’t get cited; facts do.
For the rest of the rapid-fire answers, check out the full episode on YouTube.
Conclusion
AI visibility isn't a mystery metric, it's a measurable probability, and most agencies haven't started measuring it because the old scorecard still looks fine. Ask whether your robots.txt and CDN settings agree with each other, whether a machine reading one page could answer who you are and what you charge, and who gets recommended instead of you. Teams weighing a redesign with this in mind can also compare digital marketing agencies on verified reviews before committing.
Where to Find Chandan Sharma and Doodle Web
Website:Doodle Web
LinkedIn: Chandan Sharma
Conversations That Matter is the Goodfirms podcast featuring the practitioners, founders, and operators behind some of the most interesting companies in the world. Subscribe wherever you listen.
