Original insights by Olivier de Segonzac and the RESONEO research team
30-second rundown
Key learnings:
- Retrieval is not the same as being read or cited: RESONEO observed 58,000 URLs retrieved, 7,600 promoted as sources, 5,000 cited in answer text and only 760 pages actually opened by ChatGPT.
- Most pages are represented by a short cached snippet: ChatGPT often works from a title, URL and a passage roughly 200 characters long, so a weak summary can become the entire witness statement.
- Different ChatGPT modes consult different source pools: Free Think, the reasoning option available to users on ChatGPT’s free tier, leaned heavily on OpenAI’s Labrador index, while paid thinking, the reasoning mode tested on a paid ChatGPT account at medium effort, relied mainly on scraped Google results in the August tests.
This week: Ask ChatGPT one commercial question a potential customer would use before buying, such as “What is the best accounting software for a five-person agency?” Open that conversation in FanoutFox and compare the sources ChatGPT retrieved with the smaller set it cited in the answer. Note which sources may have influenced the response without receiving a visible citation, and whether your website appears in either group.
Coming 3 months: Use an AI Search monitoring workflow to repeat priority questions daily across free and paid modes, storing searched, retrieved, promoted, cited and opened events separately. Assign one owner to investigate changes and update pages whose facts disappear before citation.
Olivier de Segonzac and the RESONEO research team pulled apart ChatGPT’s web retrieval chain and found a pyramid full of disappearing bodies. Across more than 1,200 real answers and 27,000 examined pages, 58,000 retrieved URLs became only 5,000 visible citations. Just 760 pages were actually opened by ChatGPT and read.
The rest were represented by titles, addresses and snippets roughly 200 characters long. Your grand page can meet the answer engine as three lines scribbled on a card by somebody else.
1. Retrieval is a series of gates, not one victory
Marketers often treat a citation as the natural reward for appearing in search. RESONEO’s data shows a more brutal sequence: a URL can be retrieved, promoted into the source set, cited visibly, or opened for deeper reading. Each gate throws more pages into the drainage ditch.
Of 58,000 retrieved URLs, around 7,600 were promoted as sources and 5,000 were cited in the answer text. Only 760 were opened. A retrieved page can influence the answer without earning a citation, while a page that never enters the chosen corpus cannot compete at all.
Citation share ignores invisible inputs, while referral traffic misses influential sources that receive no link. Record whether a source was searched, retrieved, displayed, cited and capable of sending a visitor as separate events.
Imagine a running-shoe comparison. ChatGPT may retrieve a retailer’s durability test, use its snippet to shape the verdict, and cite a larger publication instead. The retailer influenced the sale but sees neither link nor session. The evidence passed through the room wearing another source’s coat.
2. Two hundred characters may carry your entire case
In most observed answers, ChatGPT did not open the page. It worked from a search result containing the title, URL and a short snippet. The research found that pages were opened in barely more than one case out of 80, mainly during thinking modes.
Your snippet can become your complete representation. If it contains generic introduction, old pricing or a sentence that depends on surrounding context, the model may never reach the clean explanation waiting farther down the page.
RESONEO also identified a read cache with a roughly 30-minute freshness window. Within that period, the page was not reopened. Its stored copy could persist for months, meaning “we updated the page” and “ChatGPT now sees the update” are not the same event.
Write the opening answer so it survives extraction. State the subject, claim, condition and evidence together. A hotel should not say “Our refreshed approach gives guests more.” It should say which resort changed, what changed, who benefits and when the information was last materially updated.
3. ChatGPT changes the road before choosing the sources
The August tests revealed different retrieval routes. Free Think pulled 74.7% of its results from Labrador, OpenAI’s own index, while paid thinking at medium effort pulled 75.3% from scraped Google results.
One click on Think also more than doubled the source volume for free users, from 15.1 to 35.3 URLs per conversation. Meanwhile, paid thinking’s search breadth fell sharply from July to August. There is no single stable “ChatGPT ranking.” The mode, account and retrieval path can change the corpus before your page gets a hearing.
That makes one-off monitoring dangerously theatrical. Ask the same commercial questions across relevant modes and repeat them over time. Track distributions, not one lucky screenshot pinned above the office kettle.
The research is a dated reverse-engineering snapshot, not an OpenAI specification. RESONEO carefully marks July findings that August changed. That volatility is the point: the machinery can move while the interface keeps the same polite face.
The citation pyramid is not a leaderboard. It is a sequence of trapdoors. Make every extracted passage understandable, monitor each retrieval stage, and never assume a page was read merely because its shadow appeared in the answer.
