There used to be a "Sources" button under a ChatGPT answer. Clicking it opened a panel listing every source: site name, article title, date, summary. If you had been cited, you were visible, and you had a shot at the click.
That panel is gone. In its place, each paragraph ends with a single small marker showing one site name and a count — "iTechGuides +1". The "+1" means there is another source you cannot see unless you tap.
We sampled October answers and found more than nine in ten source markers carry a "+N", and the number folded away is roughly equal to the number shown. In practice, about half of all sources are invisible by default.
Key takeaways
- Source display narrowed in stages: full panel → "Sources" moved into the overflow menu (13 August 2026, some accounts) → one site name per paragraph (October).
- Over 90% of markers we sampled carried "+N", with roughly as many sources hidden as shown.
- Being cited is no longer the finish line. The first-listed source takes most of the visible exposure.
- Citations and buy links are different systems — but being in the sources raises the chance the buy button points at your store from 6% to 34%.
- ChatGPT rewrites the shopper's question before retrieving. Optimise for the rewritten query, not the spoken one.
What the change looks like
On the left, OpenAI's own illustration from the ChatGPT search launch: a Sources panel enumerating every source with enough metadata to judge it. On the right, our 21 October capture: a paragraph ending in one site name and a "+1".
This did not start in October. The narrowing has been running for months:
When | How sources were shown | What it meant for a cited brand |
|---|---|---|
Before May 2026 | "Sources" button under the answer; opens the full list | Being cited gave you a real shot at the click |
From 13 Aug 2026 | "Sources" moved into the "…" overflow menu, on some accounts | Two clicks to see any source |
October 2026 (our test) | One site name + N per paragraph, the rest folded | The first-listed source takes most of the exposure |
The direction is consistent, and it is the same direction every other answer surface has moved: fewer outbound affordances, more of the answer resolved in place.
Two link systems, one answer
Before drawing conclusions, separate the two things in the answer that look like links.
The source marker at the end of a paragraph is an attribution. It says "this claim came from here". Clicking it opens a card; only clicking through the card opens the original page.
The "Shop at Best Buy" or "View at Apple" line inside a product entry is a buy link. It sends the shopper to somewhere that sells the thing. Since May 2026, brand names mentioned in the body also link to the brand homepage — also a buy link, not a citation.
These generate different kinds of traffic and should be reported separately. But they are connected, in one direction, and the connection is the most actionable finding in our whole October data set.
Why citation decides the sale
When a brand's own product page is among the sources cited in a shopping answer, the buy button points at the brand's store 34% of the time. When the brand page is not among the sources, 6%. Stricter definition: 39% versus 7%.
So the two systems meet here. Citation is an editorial outcome owned by content and PR. Buy-link ownership is a commercial outcome owned by ecommerce. In a ChatGPT shopping answer they are the same event, roughly five-fold apart.
Combine that with the display narrowing and the goal sharpens considerably: do not aim to be in the source list. Aim to be the first source. Everything after the first name is behind a tap, and taps are rare.
What ChatGPT is actually searching for
You cannot win a retrieval you cannot see. Signed out, ChatGPT shows only "Searching the web". Signed in, it shows the terms it generated.
The shopper asked: "My husband has tried several sleep earbuds but they're uncomfortable or fall out during the night. Which ones actually stay in?"
ChatGPT searched: "best sleep earbuds stay in ears overnight side sleepers 2026 reviews".
Three additions the shopper never made: a year, a content type ("reviews", "best"), and an audience qualifier ("side sleepers"). OpenAI's help documentation confirms the general behaviour — questions are typically rewritten into one or more more-specific queries.
The content implication is direct. The pages that get retrieved are the ones matching the rewritten query, which in our observation follows the shape year + category + audience or situation + "reviews". A beautifully written brand page that never uses the words a model generates will not be retrieved, however good it is.
We did not isolate whether the signed-in display of search terms is caused by the session state itself or by the model that signed-in sessions run, and we did not test the effect of account region. The method for finding the queries that actually ground answers is in how to find grounding queries; the underlying mechanism is in query fan-out.
What to do about it
Target first position, not inclusion. Audit your priority questions by running them, opening each source marker and logging every entry in order. Record your rank among sources, not just presence.
Write the content types that get retrieved. Reviews, comparisons, ranked lists, question-and-answer pages. Our buy-link data shows the same bias on the commercial side: the pages that get cited are the ones that evaluate, not the ones that advertise.
Cover the rewritten query shape. Year, category, audience or situation, plus "reviews" or "best". Mechanically unglamorous and consistently effective.
Do not expect citation traffic to scale with citation count. With roughly half of sources hidden, incremental citations below first position buy you influence on the answer's content, not clicks. Both are worth having; only one shows up in analytics.
Measure citation directly. If roughly half of your citations are invisible and most of the rest do not produce a click, your analytics cannot tell you whether your source strategy is working. You need to observe the answers themselves — which is what Share of Model tracking does across ChatGPT, Google AI Mode, Gemini, Google AI Overviews and Microsoft Copilot. GEOly's brand visibility tracking runs that sampling continuously.
Deeper mechanics: how AI search picks sources and how to get cited by ChatGPT.
FAQ
How does ChatGPT decide which sources to cite? It rewrites the user's question into more specific search queries, retrieves against those, and cites the pages it used. In our observation the rewrites added a year, a content type and an audience qualifier. The ranking stage is not publicly documented; how AI search picks sources covers what is known about the pipeline.
Why does ChatGPT only show one source now? The display narrowed over 2026 — the Sources panel moved into an overflow menu in August and became a per-paragraph marker by October. OpenAI has not published an explanation. Over 90% of the markers we sampled carried a "+N" with further sources folded behind it.
How do I see all the sources in a ChatGPT answer? Tap the site-name marker at the end of a paragraph. The card that opens lets you page through the remaining sources; clicking an entry opens the original page.
Does being cited bring traffic? Less than it used to. Roughly half of sources are hidden by default and the first-listed source takes most of the visible exposure. The bigger payoff is commercial: cited brand pages get the buy button 34% of the time versus 6% uncited.
Is the source marker the same as the buy link? No. Source markers attribute claims; buy links send shoppers to a merchant. They are separate systems with separate destinations — see how ChatGPT shopping results work after GPT-6.
Sources
The display-narrowing timeline draws on Search Engine Roundtable, "OpenAI Makes ChatGPT Sources Less Visible", 13 August 2026, and on OpenAI's "Introducing ChatGPT search". Buy-button percentages by citation status are from WISLR, published 14 October 2026 and updated 16 October: 4 brands, 250 prompts, 10–13 October — four days of data, directional.
The "+N" sampling, the source-marker behaviour and the query-rewriting observation are from GEOly Research: 28 shopping prompts in a signed-out Chrome window on 21 October 2026, with the signed-out and signed-in comparison run separately across several accounts.
