How We Built an AEO-Optimized Website With Claude AI and Vercel

We built the Omni Online Strategies site as a static, crawler readable Next.js project on Vercel, with Claude doing much of the building and deployment through git, and we are publishing what worked and what we got wrong. The bottom line: the technical foundation, complete HTML and clear crawler access, was straightforward, and the hard part was content quality and accuracy, where an audit of our own articles found most of them too thin to be useful.

For a full definition, read our answer engine optimization guide.

The stack, briefly

Piece	What we use	Why
Pages	Static HTML files in a Next.js project on Vercel	Every page is complete HTML, so no crawler depends on rendering
Deploys	Claude writes and edits pages, commits, and pushes to GitHub, and Vercel deploys on push	One path from change to live, with a record of every edit
Demo library	Supabase tables filled by a GitHub Actions workflow that reads a JSON block from each demo page	New demos appear in the library, the related links, and llms.txt without code changes
Discovery	Generated sitemap, dynamic llms.txt, robots.txt naming the AI crawlers, IndexNow at build	Give search engines and AI crawlers every route in
Verification	Headless browser renders, deployment checks, and live page reads	Catch what source code review misses

What we got wrong, and how we found it

These are real findings from auditing our own site, listed because they are the useful part.

Most of our AI visibility articles were thin. Of 81 articles on AI visibility, AEO, and GEO in our own library, 17 were under 150 words, 50 more were under 500, and 5 sat between 600 and 825 words. Only the 9 we had just rebuilt with sourced research had real depth.
Our most visible AEO article had a broken meta description. It was cut off mid sentence in four places, and the body was five short paragraphs, on a page that already had 193 impressions for the query it targeted.
New hero images rendered stretched. The page CSS set the image width but not the height, so the browser used the raw height attribute and distorted the picture. It affected four articles until we measured the computed styles on the live page and found it.
We published two wrong facts and corrected them. We described Ahrefs' free checker as a technical crawl scanner when it is a brand mention checker, and we quoted a 46 percent share of ChatGPT queries that trigger a search from late 2024 when Semrush's February 2026 data puts it at 34.5 percent.
Our sitemap stamps today's date on every URL. Both generators do, which makes lastmod meaningless to a search engine.
Our build resubmits every URL to IndexNow on every deploy. Roughly 1,950 URLs each time, instead of only the pages that changed.
Our scanner's bot list is out of date. It still checks Claude-Web and anthropic-ai, which Anthropic no longer uses, while its documented bots are ClaudeBot, Claude-SearchBot, and Claude-User.

What we do now

Check every claim against a primary or clearly attributed source before it goes in an article.
Render pages in a headless browser instead of trusting the source, and read the live deployment after it ships.
Measure content depth across the whole library, not just the newest pages.
Say plainly when evidence is thin, as with schema markup and citations.

Early results

The site is early. In Google Search Console's generative AI features report for the three months to September 19, 2026, 334 pages on the site earned at least one AI Overview impression, 922 in total. Search traffic is still small relative to impressions because most pages rank below the first page. We are publishing the numbers as they are because a result you cannot check is not worth much.

What to copy, and what not to

Copy the foundation: complete HTML, explicit robots.txt, a real sitemap, structured data in the server output.
Copy the habit of auditing your own content against a depth standard.
Do not copy our early mistakes. Set accurate sitemap dates, submit only changed URLs, and keep your crawler list current.

Questions people ask

What tech stack does Omni use for its website?

Static HTML pages in a Next.js project on Vercel, deployed on push from GitHub, with Supabase for the demo library and generated sitemap, llms.txt, and IndexNow submission.

Did Claude build the site?

Claude does much of the building: writing and editing pages, running scripts, committing, and pushing to GitHub, with Vercel deploying on push. Decisions and review stay with us.

What was the biggest mistake?

Content depth. Most of our AI visibility articles were far too short until an audit of the library showed how many were under 500 words.

Is the site getting AI Overview visibility?

In Search Console's generative AI report for the three months to September 19, 2026, 334 pages earned at least one AI Overview impression, 922 in total. Clicks are still small.