Citation Watch: who ChatGPT cites, measured weekly

Citation Watch is Prefer's own instrument pointed at our own category: every week we run the same 50 buyer prompts through ChatGPT (model pinned, web search on), record every source behind every answer, classify each one as owned, earned, or competitor ground, and publish the counts. This is edition 1, built from the first two runs. The numbers are small, single-engine, and honest; the value is the trend line and the method, which you can copy. To run it for your own brand, Prefer tracks your buyer prompts on ChatGPT, Gemini, Perplexity, Google AI Overviews and AI Mode and shows the sources behind each answer.

The baseline run

Run 1 (2 September 2026): 47 of 50 prompts produced grounded answers, citing 177 sources across 92 distinct domains. Prefer was mentioned in 2 answers and cited in 2, both on prompts that named us; we earned zero unprompted presence at baseline, which is exactly the honest starting line this study exists to move.

Fig. 01

Where the citations went

Prefer

All 177 citations from run 1, by source class

69%

64%

30%

7%

3%

3%

1%

Week over week

Run 2 (7 September 2026, five days after baseline): 48 of 50 prompts grounded, 169 citations across 94 domains. Prefer moved from 2 prompted appearances to 5 mentions and 5 citations, and 3 of them were unprompted: ChatGPT cited our AthenaHQ-alternatives and AirOps-alternatives pages on those category prompts, and named Prefer first on the done-for-you-with-a-guarantee prompt, citing the Managed page. All three cited pages went live between the runs. Citations to our own pages doubled from 3 to 6.

The cited set churns hard, which is the standing caveat: 44 domains entered, 42 left, and only 50 appeared in both runs. Competitor-owned citations fell from 69 to 34, but read that gently: 41 of run 2's citations came from newly seen domains not yet classified (mostly small AI-SEO tool and agency sites), and they enter the split next edition once each has been read.

Method and caveats

One engine (ChatGPT), one model, one run per prompt per week, so counts carry run-to-run noise and we publish trends, not single-run precision. Source classes follow the taxonomy in our loop repository (owned, earned-participatory, earned-editorial, earned-profile, competitor-owned, competitor-seeded, neutral-reference). Our own results are reported unprompted-vs-prompted; we do not count answers to prompts that name us as earned visibility. For the wider cross-engine picture, see the July 2026 four-engine study (1,237 citations across 47 questions) and the Reddit citation study.