The AI Read My Post, but You Landed on My Homepage

There’s a decent chance you didn’t find this post through Google. Maybe an AI summarized it for you and you never came at all. Or, and this is the odd part, maybe an AI cited this exact page in an answer, and the link you actually clicked took you to my homepage instead.

That second scenario is the core finding of a SimilarWeb article on optimizing your homepage for AI traffic. The citation and the click are two separate events. The AI builds its answer from whichever deep page best supports the claim — a comparison, a doc, a blog post. But the link it hands the human disproportionately points at the front door.

The numbers

The channel itself is no longer a rounding error. Generative AI platforms averaged 9.5 billion monthly visits worldwide between June 2025 and May 2026, up 70% year over year. And the visitors they refer are better than search traffic, not worse: ChatGPT-referred visitors spend roughly 15 minutes on site and view 12 pages, against 8 minutes and 9 pages for Google referrals, and they convert at 7% versus 5%. Fewer people than search sends, but better ones.

Then it lurched. On May 7, 2026, ChatGPT started rendering brand names inside its answers as clickable links instead of footnote citations. Total referrals jumped 157.7% week over week, and homepage referrals jumped 354.7%. Before the change, roughly a quarter to a third of ChatGPT referrals landed on a homepage. After it: about 60%, and it has held there since.

It’s not just SimilarWeb seeing this. Orbit Media analyzed crawl and referral data across 74 Cloudflare accounts and found homepages get roughly 15 times more AI attention than any other page type. They call the flip side the “Dark Library Effect”: your articles get read and summarized constantly, but the click, when it comes, goes to the homepage.

The uncomfortable conclusion: the visitor who arrives from an AI is already informed. The AI did the introduction. They don’t need your homepage to explain who you are — they need it to get them to the thing they were just told about. Most homepages, mine included, were built for the cold visitor of 2015, not the pre-briefed one of 2026. The article turns this into a seven-criterion audit, worth running even if, like me, you don’t sell anything:

1. A clear primary action
2. Copy that assumes context
3. Routes for different visitor types
4. Visible trust signals
5. Honest claims
6. A real search function
7. Fast loads

What this blog already does

modalthinking.blog is a personal blog, not a SaaS homepage, so my version of this problem is mostly about being legible to machines. Making it easy for AI systems to read, attribute, and cite the site correctly. A fair bit of that is already in place:

• llms.txt — a plain-markdown summary of the site for LLMs: who writes it, what it covers, where the content lives, how to reach me. It’s linked in the footer, right next to the one for humans.
• Structured data — JSON-LD on every page. The homepage declares WebSite and Person (with sameAs links to GitHub and LinkedIn, so machines can connect the identity dots), and every post ships a BlogPosting block with author, dates, and keywords. The homepage even carries SearchAction markup — the exact schema the SimilarWeb piece recommends for making search machine-readable.
• Full-text feeds — the RSS feed and JSON Feed both contain complete post content, not teasers. Any system that wants a clean corpus of this site can have one without scraping HTML.
• Open robots.txt — nothing is blocked. GPTBot, ClaudeBot, PerplexityBot and friends are all welcome. You can’t be cited by systems you lock out.
• IndexNow — every deploy pings search engines with changed URLs, so new posts get picked up in minutes rather than whenever a crawler wanders by.
• The boring hygiene — sitemap, canonical URLs, Open Graph metadata with auto-generated images per post. Old-fashioned SEO, which is most of GEO anyway.

What’s still missing

Honesty section. First, my SearchAction markup points at the tags page — browsing, not search. By the article’s own rubric that’s a soft failure: a visitor who lands here knowing exactly what they want has no search box to type it into. Second, my robots.txt welcomes AI crawlers only implicitly; explicit allow stanzas would make the policy legible instead of accidental. Third, llms.txt is hand-maintained — it will drift from reality the way all hand-maintained files do — and there’s no llms-full.txt or per-post markdown endpoint for agents that want the raw text.

And the biggest one: measurement. AI apps frequently strip referrer headers, so AI-referred visits mostly land in the “Direct” bucket looking like nothing. I genuinely don’t know how much of my traffic this is. If your Direct traffic has been quietly growing, some of it probably isn’t direct at all.

The takeaway

The homepage spent twenty years as the page that introduces you to strangers. AI added a second audience — visitors who already got the introduction somewhere else, plus the machines that gave it to them — and it’s now sending that audience to the front door in bulk. The fix isn’t a redesign or some exotic new discipline. It’s mostly web hygiene we already knew how to do, done properly, plus one honest question: if someone landed on your homepage already knowing what they came for, would they find it?

Mine’s not fully there yet. But at least the robots know where to look.