In June 2025, we wrote about the zero-click era and the 30% drop in Google traffic that many websites were experiencing. It seemed clear that users are no longer visiting websites to find answers. AI answers them directly. But there is a deeper layer to this shift that we have not fully explored until now. It is not just whether users click. It is whether the AI ever "reads" your page at all. This is also the basis of the story of the many "AI visibility" tools that are being sold to you right now.

A study published in July 2026 by RESONEO, a French research group, put ChatGPT's web retrieval under a microscope. (I first go to know about this study from Lily Ray's post on X) They captured 1,249 real answers, traced 58,000 retrieved URLs, and mapped exactly what the model sees, what it cites, and what it actually opens. The findings are stark. They also validate much of what we have been saying about meaning design and AEO — and they add new urgency.

The Funnel: From Retrieved to Cited to Opened

A magnifying glass examining detailed charts, representing the scrutiny of AI retrieval and the narrowing funnel from search results to actual citations.

The RESONEO study traces a sharp funnel. For every answer ChatGPT gives using web retrieval:

  • 58,000 URLs are retrieved by the search backend.
  • 7,600 are promoted to the visible "Sources" panel.
  • 5,000 are cited in the answer text.
  • 760 are actually opened and read by the model.

That is only 1.3%. The vast majority of pages that ChatGPT "finds" are never opened. The model writes its answer from a title, a URL, and roughly 200 characters of snippet text. In most cases, that is your entire representation in front of the AI.

This changes how we think about visibility. Traditional SEO chases the top spot on a search results page. AEO chases the citation in an AI answer. But this study shows there is an even earlier gate: the retrieval snippet. If your page is retrieved but the snippet is weak, you are already out of the running before the model writes a single word.

Two Tiers, Two Completely Different Systems

The study reveals that free ChatGPT and paid ChatGPT use different search backends entirely.

Free tier (instant mode): - Uses OpenAI's in-house index, codenamed "labrador," for almost everything. - Pulls about 11 pages per answer. - Opens zero pages. Not one in the study's 493 free-tier answers went into thinking mode or opened a page. - The model sees a full title plus a 202-character snippet anchored on your H1 tag.

Paid tier (thinking mode): - Uses scraped Google results (via a pipeline called "bright") for about 75% of searches. - Pulls about 100 pages per answer. - Opens a page in roughly 1 in 6 conversations. - Sees truncated titles (median 48 characters) and shorter snippets (median 159 characters).

This is what it all means: Most people who use ChatGPT — the free tier — never trigger a live Google search. Their questions are answered from OpenAI's frozen in-house index. If your page is not in that index, or if its snippet is poorly constructed, you are invisible to the majority of ChatGPT users. (Estimates put paying subscribers at roughly 5% of weekly active users, with the free-to-paid conversion rate sitting between 5% and 6%.)

What the Model Actually Sees of Your Page

This is where the study intersects directly with our work on meaning design. The labrador snippet — the one served to free users — is built from three blocks in fixed order:

  1. Whatever sits before your H1 tag (about 7 characters on average).
  2. Your H1 tag itself (about 51 characters at the median).
  3. Whatever room is left after the H1 for actual content (about 146 characters).

Your meta description is ignored. Your structured data is stripped before the model sees it. The snippet is frozen at crawl time and does not update when the page changes. The researchers found pages whose snippets were more than a month behind the live content.

The things that steal space from your content are not what most people expect. Breadcrumbs and bylines are marginal. The real losses are section labels, publication dates, and — most surprisingly — the alt text of the first image after your title. That alt text alone can eat 50 characters. If your H1 is missing, the snippet starts at an unpredictable subheading, and you lose control entirely.

This is meaning design in practice. As Gunji Takeshi explained in our interview series, AI understands content through Query-Key-Value relationships. If the Query is the user's question and the Key is your page, the Value is what the AI extracts from that 200-character snippet. If your H1 and opening paragraph do not precisely define the Key-Value relationship, the AI assigns the wrong Value — or skips you entirely.

The Cache Problem: Your Page Is Frozen and Shared

The study uncovered another layer that should concern every publisher. OpenAI keeps two separate stores:

  • The index: a short snippet, frozen at crawl time, used to find pages.
  • The read cache: the whole page converted to Markdown, used when the model actually opens a page.

These two stores do not talk to each other. ChatGPT-User, the robot that refreshes the full-page cache, never updates the snippet index. A page can be re-read today and still serve a month-old snippet in search results.

Worse, the read cache is shared across all users worldwide. The researchers proved this with fingerprinted test pages. User A in one country requests a page at 2:00 PM. User B on a different account requests the same page 22 minutes later. User B receives the exact same copy — word for word — without User B's request ever hitting the origin server. The cache ignores Cache-Control: no-store headers. Copies were observed surviving for more than 90 days.

For SMEs, this means you cannot rely on freshness as a competitive advantage inside ChatGPT. If your page is cached, a competitor's update will not displace you — but your own updates will not reach users either, until the cache refreshes.

What Gets Opened Gets Cited

Here is the most actionable finding for anyone pursuing AEO. When ChatGPT actually opens a page, it cites that page 74% of the time. When a page is merely retrieved but not opened, the citation rate drops to 7%.

But openings are rare and specific. The most opened domains in the study were regulatory and primary sources: openai.com, anthropic.com, learn.microsoft.com, and digital-strategy.ec.europa.eu. The pattern is quite clear: The model opens pages when it needs authoritative, deep, or official material. It does not open blog posts, product pages, or marketing copy.

This aligns with what AJSA's Suzuki presented at the Tokyo International Forum: AI exhibits "academic elitism." It trusts government agencies, universities, and large established portals. For SMEs, the path to being opened is not to look like a marketer. It is to publish primary evidence, case studies with specific data, and firsthand activity reports that AI cannot generate on its own.

The Invisible Domains: arXiv, Reddit, and the Citation Gap

The study also exposes a trap in how we measure AI visibility. arXiv appeared 2,600 times in the corpus but was cited only 10 times and never opened. Reddit is similar. ChatGPT uses these sources to build background knowledge, then hides them from the user and cites ordinary web pages instead.

This is a warning to anyone tracking "AI visibility" through raw mention counts. Being pulled is not the same as being shown. Being shown is not the same as being cited. The funnel matters, and most tools on the market today do not distinguish these layers.

Five Verticals, Five Separate Systems

The study's final section is a reminder that web search is only one of five surfaces ChatGPT uses. News, local/maps, shopping, and images each have their own supply chain:

  • News: Articles are machine-rewritten into ~1,100-character summaries, not quoted verbatim. The original article is cited about 1 in 8 times.
  • Local: Feeds from Yelp, TripAdvisor, and Google Maps. The source label does not always match where each field came from.
  • Shopping: Product carousels run on merchant feeds, not web search rankings. Amazon is entirely absent from the offers.
  • Images: Re-hosted by OpenAI. They appear in answers without ever touching your server.

A visibility strategy that only handles web search leaves four doors shut. This is why Suzuki's four-channel framework — Google SEO, MEO, social media, and AEO — remains the right model. Each channel feeds a different surface.

What to Do Now

The RESONEO study gives us a concrete checklist:

  1. Mark up a clear H1. It anchors the snippet. Without it, you lose control.
  2. Clear the runway before the H1. Remove giant image alt text, section labels, and publication dates that eat your 200-character budget.
  3. Write the first 150 characters after the H1 as if they are your only chance to explain yourself. Because for most ChatGPT users, they are.
  4. Publish primary sources and case studies with specific data. These are the pages the model actually opens, and opened pages are cited 10 times more often.
  5. Do not rely on meta descriptions or structured data for ChatGPT's in-house index. They are ignored or stripped.
  6. Track the right metric. Count citations in AI answers, not raw mentions or retrieval events.

Conclusion

The zero-click era is not just about users skipping your website. It is about AI making decisions from a 200-character window that you may not even know exists. The RESONEO study confirms that the old pipeline — user searches, clicks, visits — is being replaced by something far more compressed. The brands that survive will be the ones that design their meaning precisely enough to fit inside that window, and authoritative enough to be opened when it matters.

This is not the end of SEO. It is the end of SEO as a keyword game. The future belongs to meaning design, primary evidence, and multi-channel presence. We have been saying this since our first dialogue with Gunji Takeshi. The data, at last, agrees.

Sources