I run Cirrela, a small AI video studio built in France (text or photo in, short video with sound out). Last month we counted where our accounts came from, and one number stopped me: 34 of the 44 accounts with a known source arrived from chatgpt.com. Not Google, not Product Hunt, not Reddit — one AI assistant answering "best free AI video generator".

This post is what that taught us about building for AI search, and the mistakes we made that you can avoid.

Disclosure: it's our product. This post was written with AI help, then checked by me.

The funnel was one page wide

Here is what our own analytics showed in early November:

  • 44 accounts had a known acquisition source. 34 came from chatgpt.com.
  • 32 of those 34 landed on a single page of ours: /free-ai-video-generator.
  • Of the people who saw our offer there, 13 clicked through.

In other words, essentially our entire non-direct traffic was: ChatGPT decides to cite one page → that page's title and H1 happen to be shaped like a comparison ("Free AI video generators in 2026: what's really free") → visitors arrive already convinced they're shopping for a free tool → most bounce, but enough convert.

We didn't plan any of this. We built the page to answer a real question honestly, and AI search rewarded it.

Why AI assistants cite list pages, not homepages

This matches what the citation studies are finding: AI assistants overwhelmingly cite third-party comparisons, "best of" lists, and dated how-to pages — not company homepages. ChatGPT answers "best free AI video generator" by pointing at pages that compare tools, because that's what its sources look like.

So the SEO playbook changed for us. It's not "rank for our brand" anymore. It's "be a credible entry in the comparison the assistant is already reading" — which means the page has to be genuinely useful, dated, and honest about competitors. We list the real free options (Google Flow, Dreamina, PixVerse, open models like Wan 2.2) before ours, with sources and the date we checked them.

The part nobody warns you about: bots hit your infra differently

Two weeks ago we tested what ChatGPT's crawler (OAI-SearchBot) actually gets from our pages. Half the requests failed with 503s — not because of the content, but because our hosting plan caps CPU per request and the crawler's fetch pattern trips it. Googlebot and bingbot saw the same errors.

A search engine that gets an error page doesn't cite you. A person who gets one doesn't sign up. For about a week, our most important page was effectively invisible to the channel that fed us — and nothing on our analytics dashboard said so, because humans were still loading it fine.

The fix was boring: pay $5/month for the hosting tier without the CPU cap. The lesson wasn't "scale your infra" — it was "your free tier may already be lying to your crawlers".

The rule that saved us: never retitle a cited page

Early on we retitled that page for "better marketing". Signups from ChatGPT dropped to near zero within days. The page's title, H1 and URL are what the assistant matched on; changing them was like changing our storefront sign while the map still pointed at the old one.

Our rule now, written in the repo where every agent sees it: never retitle, rename or move a page an AI assistant cites — title, H1, URL — without treating it as a product change with measurement, not a copy tweak.

What we did after the scare

ChatGPT's answers drift. On Nov 14 we asked it the same questions again and Cirrela was cited nowhere — the assistants had moved to newer blog posts from competitors. So a single channel, however good, is a single point of failure. Since then:

  • We diversified: directories, an MCP server so Claude and ChatGPT users can use the product inside the assistant, review-site outreach, and this blog.
  • We made the cited page more citable, not less: dated sources, a "how we verified" section, honest "not for you" notes.
  • We watch citations like we watch uptime: a monthly checklist of the exact questions we want to be cited for.

Takeaways if you get AI-search traffic

  1. Find out which of your pages assistants actually cite — ask the questions your customers ask, in fresh chats, and note the sources.
  2. Keep that page's title, H1 and URL frozen. Improve the body all you want.
  3. Check what your crawlers see, not just what browsers see. A 503 to OAI-SearchBot is silent churn.
  4. Treat AI citations as a channel with its own rules: be listed, be dated, be honest about the alternatives.
  5. Never let one assistant be 75% of your signups. Channels decay; the fix is boring distribution work, not a clever hack.

We're small and the numbers are small — but the shape of the funnel is the story: one page, one assistant, almost everything. Build like that can change overnight, because it can.