Back to HTML Pub

The Agent Referral Is Here: Designing Landing Pages for Visitors Who Arrive From ChatGPT, Not Google

Michael Sacca
AI-Native Publishing
Landing Pages
Publishing Strategy

A few months ago the question worth asking about AI and your landing page was a legibility question: can the agent read your page at all? We covered that one already — Your Landing Page's Newest Visitors Are AI Agents — Can They Even Read It? — and it still matters. But legibility is the entry fee, not the game. The newer question is what happens after the agent reads your page, decides you're a good fit, and hands the link to a human who clicks it.

That click is a different animal from anything search ever sent you. Google's click means "I'm curious." ChatGPT's click means "the model just told me this is probably what I want." The referral already happened before the visitor saw your domain.

What an AI referral actually looks like

The mechanics vary by platform, but the pattern is converging. ChatGPT Search pulls live web results into the conversation and cites them inline — the user reads a paragraph of synthesized advice with your domain cited as the source, then clicks through. Perplexity behaves like a research engine, citing your page in an answer box with follow-up questions that keep the user in the thread. And browsing-capable agents go further: they fetch your page on someone's behalf, extract the answer, and only surface the link when the user asks for it.

Three things follow from that, and each one should change how you build.

First, the visitor arrives pre-qualified but context-poor. They've read two sentences about you inside someone else's answer box. They don't have your headline, your framing, or the thirty seconds of context a search result snippet used to provide. Your page has to re-establish the offer in one screen or they bounce.

Second, the agent's summary is your new headline. Whatever the model said about you before the click is the copy your visitor actually read. If your page contradicts that summary, you've broken message match in a way no A/B test would have caught.

Third, the agent may never "visit" the way a browser does. When a browsing agent fetches your page, it doesn't render your hero image. It reads the DOM. The parts of your page the machine sees are the parts that get recommended.

Build the answer page, not just the sales page

The most practical shift: for every commercial page, build a companion page that exists to be cited. This isn't a blog post for its own sake — it's the page the model wants to quote.

ChatGPT Search and Perplexity both cite pages that directly answer a question in prose, with the answer near the top. A pricing page that buries the number three scrolls deep gets skipped; a short page titled "What does [product] cost?" that states the price in the first paragraph gets cited. Same facts, different packaging.

The structure that consistently works:

<article>
  <h1>Does [product] integrate with Zapier?</h1>
  <p><strong>Yes — [product] connects to Zapier with a native
  integration, and setup takes about five minutes.</strong></p>
  <!-- Then, and only then, the details, pricing of the
  integration tier, alternatives, caveats -->
</article>

The answer in the first fifty words, stated plainly, in a sentence a model could lift verbatim into a response. Everything else on the page exists to support that sentence. This is old SEO advice wearing new clothes — featured snippets and voice search wanted the same thing — but the stakes are higher now, because the citation is the click.

Your conversion path has to survive the summary

Here's the part nobody's optimizing yet. When an agent recommends you, it compresses you. "Sacca's team recommends X for teams under 50 people; it's $29/month and doesn't have a mobile app." Your visitor clicks with that compressed version in their head.

Your landing page's first screen now has three jobs at once: confirm the summary, add the one thing the summary left out, and give the visitor somewhere to go. If your hero says something the agent didn't mention — a different price, a different positioning, a claim about a feature you don't have — you've created a moment of doubt at the exact second trust was highest.

The fix is boring and effective: write your own one-sentence summary first. If ChatGPT described your product in twenty words, what would those words be? Then check that your hero, your pricing, and your CTA all agree with it. If your page has been accreting sections for two years and no longer has a one-sentence version of itself, that's the real problem, and the agent referral just exposed it.

Serve both audiences without building two sites

You don't need a separate "agent site." You need pages that degrade gracefully for machine readers the same way good pages degrade gracefully for screen readers.

A few concrete rules that serve both audiences:

  • Server-render the substance. Agents fetching your page see HTML, not your React hydration. If your pricing or your offer only appears client-side, part of your audience literally cannot see it.
  • Keep the answer in prose, not just a component. An interactive pricing calculator is great for humans and invisible to an agent. Put the base price in a paragraph next to it.
  • Let the headline do double duty. "Landing pages for paid campaigns, from $29/month" is a headline for humans and a citable fact for machines. "Build beautiful pages effortlessly" is neither.
  • Mark up what you can. Product, price, FAQ schema — the structured data work we recommended in the legibility post pays off doubly now, because it's how an agent knows your page answers the question before it recommends you.

The human version of the page can be as designed and animated as you like. Just make sure the skeleton underneath is legible, quotable, and consistent with whatever an agent would summarize.

Measure it, or you won't believe it

Referral traffic from ChatGPT and Perplexity shows up in analytics under a tangle of sources — chat.openai.com, chatgpt.com, perplexity domains, sometimes nothing at all if the referrer is stripped. Segment what you can and compare the behavior: AI-referred visitors typically arrive knowing more and convert faster, but in lower volume than search. That ratio is the whole story — it tells you whether to spend next quarter's effort on citation-worthy answer pages or on the volume channels you already have.

Watch your branded search volume too. A common pattern is the agent referral that doesn't click, then becomes a direct visit or a branded search a day later. If your answer pages are getting cited, the effect will show up there before it shows up in session counts.

The workflow is already yours

If you've been following the Claude-assisted publishing workflow on this blog, none of this requires new tooling. You already generate pages from prompts. Add one more artifact to every build: the companion answer page, written from the same brand rules file, with the one-sentence summary as the source of truth for both the sales page and the citable version.

The practical loop looks like this: draft the one-sentence summary of your offer, generate the answer page around it, generate the landing page around it, deploy both together. When the offer changes, change the summary first and regenerate both pages so they never drift apart. An agent that quotes the old price while your landing page shows the new one is worse than either error alone.

The agent referral isn't a future trend to file away. It's showing up in analytics right now, in small numbers that will decide large outcomes as the behavior compounds. The sites getting cited are the ones that made themselves quotable on purpose. That's a build decision, not a marketing one — which is exactly the kind of decision this blog exists for.

Keep Reading