BuildNot yet confirmed elsewhere1 publisher3 min readPublished
Gemini's repeat runs share only 0.319 of cited domains in the Kingy AI Citation Index
Kingy AI Citation Index found the same provider's two runs overlapped 0.664 on cited domains for OpenAI, 0.515 for Claude and 0.319 for Gemini. One check cannot separate a lost citation from a different draw, so a change needs logged repeat runs before anyone calls it real.
The Engineer · Build desk
Cited-domain overlap in the Kingy AI Citation Index: Gemini's two runs overlapped 0.319, OpenAI's 0.664 and Claude's 0.515. Different providers answering the same question overlapped about 0.06 for every pair.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Gemini, same provider, two runs | 0.319 domain overlap | 3 | |
| OpenAI, two runs | 0.664 domain overlap | 3 | |
| Claude, two runs | 0.515 domain overlap | 3 | |
| Different providers, same question | 0.06 overlap, about | for every provider pair | 4 |
What happened
- The Kingy study put the same 1,200 designed English questions to OpenAI, Claude and Gemini twice through their official APIs, giving 7,200 completed answers.
- Across different providers asked the same question, cited-domain overlap was about 0.06 for every pair.
- Share of answers with any native citation at all was 71.6% for OpenAI, 53.7% for Claude and 25.2% for Gemini.
- On cloro's monitoring corpus, ChatGPT's rate of attaching a source was 99.6% for July, then 79.6% over the 30 days ending 19 August and 97.6% over the 30 days ending 7 October.
Why it matters
- constraint A missing citation has three possible causes: no citations that run, a different handful from a shifting pool, or a real loss. A single check cannot tell them apart, so one absence is not evidence of a drop.
- cost The minimum protocol is 30 logged rows per session for five prompts across three providers, before the second session that the change rule requires.
- contradiction Gemini's baseline differs by about 2.7 times between sources: 25.2% of answers with a native citation in Kingy's API panel, 68.2% attaching a source on cloro's production corpus. Panel design and method set the denominator.
- decision Teams have to set a threshold for calling a move real. The write-up's author uses two fresh runs in two sessions, same direction. With Gemini overlap at 0.319, we would run three, which the author also rates as better.
Kingy scored each pair of runs with Jaccard overlap: shared domains divided by all domains in either run [2]. Subtract that from one and you get the share of the combined domain list that showed up in only one run. That is 0.336 for OpenAI, 0.485 for Claude and 0.681 for Gemini [22]. A repeat check on Gemini changes about two thirds of the list [22].
The study's author lists the limits. The panel was designed for the study and not sampled from real user demand. Results through the APIs may not match what the consumer apps return. And two runs in a single window cannot tell you how the engines will behave later [6]. The work was commissioned and paid for by Kingy.ai [7]. For the numbers to transfer to your tracker, your buyer questions would have to behave like a designed panel, and your tracker would have to read the same API output.
Minel Gunesoglu's round two orders the engines differently. Working from the re-cite counts, the rates are 63% for AI Overviews, 29% for ChatGPT and 59% for Gemini [23]. ChatGPT is the least stable there. In Kingy, OpenAI is the most stable [3]. The units differ: Jaccard on domains in one study, the share of first-run pages re-cited in the other. The repeat set was ten questions, in English, from a New York location, and the author sells an AI visibility product [12]. We would not carry an engine ranking from either study into a plan.
Single answers swing hard. In one case, a ChatGPT answer on local moving prices had 5 of its 6 figures appearing on the pages it cited, and then 0 of 6 [10]. As the study puts it, "One answer is one draw." [11]
Kingy's 25.2% means 74.8% of Gemini answers carried no native citation [25]. The denominator moves too. cloro sells a search data API and says its own two methods disagree, with Gemini at 32.9% against 49.5% in the same July fortnight [16]. Its conclusion is that no single figure for these engines "survives long enough to plan around" [15].
The dev.to write-up logs one row per run. The cells are date, engine, prompt id and run number, then whether the answer had any citation, whether the brand was named, whether the URL was cited, and whether stated figures were correct [18]. Its author says the any-citation cell is the one teams most often leave out. When Gemini returns no sources at all, there was no citation to lose [21]. Results are reported as rates, and engines get separate columns because a blended score averages things that barely share sources [20].
What to watch
- A later wave of the Kingy index that repeats the test on consumer apps or on prompts sampled from real demand, the two limits its author states.
- A repeat set larger than ten questions from the Is My Brand in AI study, to see whether ChatGPT stays the least stable engine there.
- Any reconciliation by cloro of its two Gemini methods, which gave 32.9% and 49.5% in the same July fortnight.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence50
- Adoption
- Insufficient
- Hype gap−10
- Incentives70
- Confidence55
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
The Kingy AI Citation Index, Wave One (Curtis Pyke, published 5 October 2026) asked OpenAI, Claude and Gemini the same 1,200 designed English questions twice through their official APIs, for 7,200 completed answers in total.
ReportedSupportedSource: Kingy AI Citation Index, Wave One, as summarised on dev.toView cited source - [2]
The Kingy study compared cited domains between runs with Jaccard overlap, defined as shared domains divided by all domains in either run.
- [3]
Same provider, two runs, cited-domain overlap: 0.664 for OpenAI, 0.515 for Claude, 0.319 for Gemini.
- [4]
Different providers answering the same question had cited-domain overlap of about 0.06 for every pair.
- [5]
Share of answers with any native citation: OpenAI 71.6%, Claude 53.7%, Gemini 25.2%.
- [6]
The Kingy author states three caveats: the panel was designed for the study and not sampled from real user demand; API results can differ from the consumer apps; two observations in one window cannot describe future behavior.
ReportedSupportedSource: Kingy AI Citation Index, Wave One, author's stated caveatsView cited source - [7]
Kingy.ai commissioned and funded the Kingy AI Citation Index study.
- [8]
In the Is My Brand in AI AI Answer Evidence Index round two (Minel Gunesoglu, published 8 October 2026), fifty price and cost questions went to five engines, and the first ten were asked a second time on 7 October to Google AI Overviews, ChatGPT and Gemini.
- [9]
On questions where both answers cited a page, Google AI Overviews cited again 32 of the 51 pages from its first answers, ChatGPT 5 of 17, and Gemini 17 of 29.
- [10]
One ChatGPT answer about local moving prices went from 5 of 6 figures found on its cited pages to 0 of 6 between the two runs.
- [11]
The study's own reading: "One answer is one draw."
- [12]
The round-two caveats are: ten repeated questions, English, New York location, and the author sells an AI visibility product.
- [13]
On cloro's production monitoring corpus, Gemini attached a source to 68.2% of answers in the 30 days to 7 October 2026.
ReportedSupportedSource: cloro AI Search Index freshness check dated 7 October 2026View cited source - [14]
On cloro's corpus, ChatGPT attached a source to 99.6% of answers in July, 79.6% in the 30 days to 19 August and 97.6% in the 30 days to 7 October.
- [15]
cloro's own line is that no single figure for those engines "survives long enough to plan around".
- [16]
cloro sells a search data API and notes its own two methods disagree: Gemini 32.9% against 49.5% in the same July fortnight.
- [17]
The write-up recommends freezing five buyer questions written without the brand name, never editing the wording, and running each prompt at least twice per engine per session in fresh chats; two is the minimum and three is better.
- [18]
The recommended log has one row per run with these cells: date, engine, prompt id, run number, answer had any citation (yes/no), named (brand appears in text), cited (URL is in the sources), and figures correct.
- [19]
The write-up's rule is that a change counts when it shows up in at least two fresh runs in two separate sessions, and in the same direction.
ReportedSupportedSource: dev.to write-up (robin_geo), described as the author's own ruleView cited source - [20]
The write-up says to report rates, not single results, and to keep engines in separate columns because a blended AI score averages things that barely share sources.
- [21]
The write-up's author says the any-citation cell is the one most people skip: if Gemini gave no sources at all, you did not lose a citation, because there was none to win that time.
- [22]
Share of the combined cited-domain list that appeared in only one of two runs: 0.336 for OpenAI, 0.485 for Claude, 0.681 for Gemini.
- [23]
Page re-cite rates in the round-two repeat: Google AI Overviews 62.7%, ChatGPT 29.4%, Gemini 58.6%.
- [24]
Five frozen prompts, each run twice per engine, across the three providers Kingy tested, is 30 logged rows per session at the minimum.
- [25]
In Kingy's panel, 74.8% of Gemini answers carried no native citation.
- [26]
cloro's Gemini source-attach rate (68.2%) is about 2.7 times Kingy's share of Gemini answers with any native citation (25.2%).
- [27]
A missing citation can mean the engine gave no citations that time, picked a different handful of sources from a pool that changes on every run, or genuinely stopped using you; one check cannot tell those apart.
Sources
1 independent publisher whose own reporting we read for this story.
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- AI citation measurementFollow
- Answer and Generative Engine OptimizationFollow