ChatGPT Often Answers Without Sources
Ryan KingsFounder & CTOPublished August 2026 · 11 min read
In July 2026 we asked ChatGPT, Gemini, and Perplexity 96 AEO-category questions and recorded whether each answer included a source list.
| Engine | Checks | Answers with ≥1 source URL | No source list |
|---|---|---|---|
| ChatGPT | 96 | 25 | 71 |
| Gemini | 95 | 95 | 0 |
| Perplexity | 176 | 176 | 0 |
ChatGPT’s sourced share is 25/96 (26%, Wilson 95% CI 18–36%). Gemini and Perplexity returned at least one source URL on every recorded check in this panel. Perplexity’s higher check count reflects same-day variance repeats. One Gemini check failed on the scam-study sweep (23 recorded instead of 24); that is a failed call, not an unsourced answer.
An answer without a source list cannot cite a URL. It can still name a brand. The rest of this paper separates those two outcomes and reports what three instruments measured.
Scope
The 96 questions come from four published engine studies — tool selection, agency trust, brand authority, and “is AEO a scam” — 24 questions each, asked to the same three engines. All four sets are AEO and AI-visibility topics. This is not a claim about health, finance, travel, or any other category.
Source papers and datasets:
- Answer engine optimization and AI search visibility
- Evaluating and trusting AEO and GEO platforms
- Building brand authority signals
- Is AEO a scam or a legitimate discipline
Two further instruments sit beside that panel:
- Citation board (aeoforged.com) — 58 tracked prompts, 1,610 completed checks, 6–13 August 2026. First-party dogfood, not a second category census. Detail: How AEOForged operates.
- Proof harness — 77 tracked URLs, one org, 5 June–13 August 2026, 18,535 direct engine observations with separate brand-named and URL-cited flags.
- llmstxtgen.com field note — named referrers and Search Console on the llms.txt generator, mid-June to 17 August 2026. A referrer click is not a source-list row. Detail: ChatGPT referral traffic on an llms.txt generator.
None of these instruments is a citation guarantee.
Definitions
| Term | Meaning in this paper |
|---|---|
| Source list | The engine returned at least one source URL with the answer |
| Mention | The brand name (or a recorded alias) appears in the answer text |
| Citation | A URL appears in the source list |
| Ghost citation | A citation where the brand name is absent from the answer |
| Generic prompt | Question contains no brand term |
| Branded prompt | Question contains a brand term |
Mention rate and citation rate are peer series. They are never averaged into a single “visibility %.”
When a source list exists, this category is editorial
Among citations that do exist in the July panel, 3,188 of 3,563 (89.5%) were editorial — blogs, publications, and vendor pages. Across the four studies the share ran 84.8–92.0%. Only youtube.com and linkedin.com appeared in every study’s top five cited domains. Reddit led two of the four papers and left the top five on the tool-evaluation set, where vendor comparison pages took over. Full rollup: What sources do AI engines actually cite?.
Same-day variance: each study repeated its first 10 questions three times the same day on the primary engine. The leading cited domain changed in 39 of 40 panels.
ChatGPT’s empty source lists and editorial dominance among citations describe different layers. When a list exists, these engines mostly cite editorials. ChatGPT often never opens that layer.
August board: empty lists concentrate on generic prompts
On the aeoforged.com board (6–13 August 2026), ChatGPT’s empty source lists were far more common on generic prompts than on branded ones.
| Prompt class | ChatGPT checks | No source list |
|---|---|---|
| Generic | 329 | 157 (48%, Wilson 95% CI 42–53%) |
| Branded | 193 | 7 |
Gemini empty lists: 7 of 316 generic, 27 of 185 branded. Perplexity: 0 of 587 empty in this window.
Page-tier citations on generic prompts. Of 999 generic checks, 36 cited an aeoforged.com page. All 36 were Perplexity. ChatGPT: 0 of 329. Gemini: 0 of 316. Of those 36, 7 named the brand, 11 cited the page without naming us, and 18 have no usable name flag (counted as name unknown, not as ghost). Branded prompts are a different signal: 560 of 611 checks cited us at some tier — plumbing, not category authority.
Who was cited instead on generic prompts (source-URL objects; an answer can list many URLs; own domain excluded for Perplexity): editorial hosts remained the majority on every engine — ChatGPT 1,096 of 1,121; Gemini 3,125 of 3,303; Perplexity 5,864 of 6,981. Top third-party hosts: reddit.com (386), youtube.com (379), linkedin.com (329).
Mentions outnumber URL citations in the proof harness
On 18,535 harness checks (77 URLs, one org, 5 June–13 August 2026), both the brand-named flag and the URL-cited flag are populated on every row.
| Outcome | Count | Share of checks |
|---|---|---|
| Brand named (mention) | 7,147 | 38.5% |
| Named, no URL | 5,985 | 32.3% |
| URL cited and named | 1,162 | 6.3% |
| URL cited, name absent (ghost) | 111 | 0.6% |
Mentions outnumber URL citations about 5.6 to 1. Ghosts are 111 of 1,273 citations (8.7%).
By engine:
| Engine | Named | Cited | Ghosts |
|---|---|---|---|
| ChatGPT | 3,017 | 337 | 0 |
| Gemini | 2,490 | 170 | 2 |
| Perplexity | 1,640 | 766 | 109 (14.2% of its citations) |
On ChatGPT in this harness, brand appearances are mostly names in the prose, not links. Perplexity is where URLs concentrate — and where unnamed citations concentrate. Google AI Overviews: 13 rows — too thin to quote.
This harness is our tracked assets, queried repeatedly. It is not a web sample and not a re-run of the July studies.
What the numbers imply
These implications follow from the measurements above. They are not a citation plan for every sector.
-
On ChatGPT, presence in the answer text matters more than presence in a source list. In the harness, ChatGPT named the brand 3,017 times and cited a URL 337 times. Optimising only for links targets the scarcer outcome on that engine.
-
Empty source lists and editorial citation mix are compatible. The July panel is ~90% editorial among citations that exist. ChatGPT often returns no list. A strategy that assumes “more blogs → more ChatGPT links” treats a scarce event as the main channel.
-
Generic and branded prompts are different instruments. ChatGPT attached sources far more often when the question already named a brand. Category questions — whether you exist when the buyer did not name you — are where empty lists concentrate. Branded citations prove plumbing; generic citations and generic mentions are the authority and brand signals. In the August window, generic ChatGPT and Gemini checks named aeoforged.com on 0 of 645.
-
Engines are not interchangeable. Perplexity produced all 36 generic page-tier hits for aeoforged.com in that window and also produced 109 of 111 harness ghosts. Gemini usually attaches sources and still cited none of our generic pages in those eight days. A Perplexity citation is not a ChatGPT mention.
-
YouTube and LinkedIn are not optional extras in this category. They were the only domains in every July top-five, and remained top third-party hosts on the August generic board alongside Reddit.
-
A same-day leaderboard is not a ranking. 39 of 40 variance panels flipped the leading cited domain. Report the date, the engine, and whether the answer named the brand.
This panel is AEO-category only. Other markets need their own freeze — same questions, named roster, misses kept, CIs shown. Method: Competitive AI citation benchmark for my sector. Steal the method, not the 89.5%.
Limits
- ChatGPT does cite. It returned sources on 25 of 96 July answers and on more than half of August generic checks for this board. The gap is relative to Gemini and Perplexity, and larger on generic prompts than branded ones.
- Editorials are most of the citations that exist. Necessary for the citation layer; not sufficient as a ChatGPT plan by themselves.
- Reddit and YouTube remain in the measured mix. Ignoring them contradicts the data.
- aeoforged.com’s generic zeros on ChatGPT and Gemini are eight days, 34 generic prompts, first-party — directional. The live board is the current record.
- Publishing a measured paper does not guarantee being named or cited. Misses are published with hits.
- A URL is useful when the name appears with it. The claim here is narrower: do not treat an unnamed citation as the brand win, and do not use citation rate as a proxy for whether the buyer heard the name.
How this was measured
July engine studies. Four published, non-retired Original Research Engine papers. Real ChatGPT, Gemini, and Perplexity calls. Source URLs stored per check. Variance panel: first 10 questions × k=3 same day. Narratives grounded only in each study’s frozen rollup. Combined citation-type counts (3,563 citations, 89.5% editorial) are taken from the published synthesis, not re-derived here.
August citation board. Visibility checks for the aeoforged.com workspace, 6–13 August 2026, joined to each prompt’s stored branded/generic class. Empty source lists are recorded as empty. Page-tier citation means an aeoforged.com page was the cited URL. A mention is a recorded brand-name hit. Where that flag is missing, the row is “name unknown,” not “unnamed.” Third-party host counts are source-URL objects classified with the same host rules used for citation-source intelligence. An answer can contribute many URLs.
Proof harness. Direct answer-engine observations on 77 tracked URLs for this org, 5 June–13 August 2026, 18,535 rows. Brand-named and URL-cited are separate flags. Mention rate, citation rate, and ghost share are never blended. One org — not a multi-client study.
Failed engine calls are not written as miss rows on the board. A day with no row means no completed check that day.
Observation pack for the live board (on request after review): dogfood data. Each July study publishes its methodology and dataset beside the article.