Open data

The AI Citation Index

309 readings · 2026-08-27 to 2026-10-08 · rebuilt 8 October 2026 · free to cite

Every morning we ask ChatGPT and Gemini the buying question behind one of our ranked lists, and record every website the engine returns as a source. This page is that log, published in full. The studies of AI citation we have read report their conclusions and their aggregates. This is the layer underneath one: the raw source sets, with the ask counts attached.

How the question is phrased changes the answer more than anything on the page

We ask each category two ways. Once the way a search engine expects — best <category> <year> — and once the way a buyer who has already shortlisted actually types: <BrandA> vs <BrandB> vs <BrandC>: which is best in <year>?. Same category, same day, same engine.

The two framings return 775 and 1253 distinct source domains respectively — the same number twice is a coincidence, not a typo, and the two sets are not the same sets — and they share only 173 of them — an overlap of about 22%. 602 domains appear only under the generic phrasing and 1080 only under the comparison phrasing. Whatever these engines are doing, they are not consulting one ranked list of trusted publishers and reciting it.

These three rates measure one specific thing: how often the engine cited the Top 11 page the question was generated from. They are a property of this site as much as of the engines, and we publish them because the ask counts make them checkable, not because they generalise.

There is no oligopoly here

Across 384 comparison answers these two engines returned 5,500 domain citations spread over 2,631 distinct domains. The single most-cited domain in the whole dataset — g2.com — holds 3% of them. The top ten together hold 12.7%, and you have to go to a hundred domains before you have covered even 29% of the citations. 63.8% of every domain cited was cited exactly once.

The reason we think this is a real property of the engines and not an artefact of how we ask is that it reproduces on each engine on its own. ChatGPT and Gemini are built by different companies, retrieve differently, and are sampled here on different schedules — three asks per page against one — and they return almost the same curve.

Comparison-phrasing answers only. One citation is one domain appearing in one answer. topelevens.com and the ownership-related domains are excluded here exactly as they are from the leaderboards, so this describes the outside world and not us.

Share of citations and presence in answers are not the same question, and it is worth keeping them apart. The median answer here cites 14 different domains, so a domain can appear constantly and still hold a small share of the total. g2.com is the clearest case: 3% of all citations, but present in 42.7% of answers.

Alongside the review aggregators you would expect — G2, Capterra, Software Advice, TechRadar — the same answers cite domains with no visible profile at all. The practical reading is that there is no short list to get onto. Being an established publisher is not a precondition for being cited, and even the most-present domain in this dataset is absent from the majority of answers.

Method

• Engines. ChatGPT (OpenAI Responses API, web search enabled); Gemini (google_search grounding). Both are asked in the same run, on the same day, with the same question.
• Questions. Generated from a live Top 11 list page. The comparison question names the top three ranked entries on that page.
• Repeats. The comparison question is asked more than once per page, because we measured that a single “not cited” reading is only about 75% reliable while a “cited” reading reproduces. Hits and asks are both recorded.
• The two engines are not sampled equally. ChatGPT is asked the comparison question three times per page; Gemini once. Gemini was also asked the generic question until 1 September 2026 and it returned no citation of the asked-about page in 106 readings, so that ask was retired to stop spending on it. Read the Gemini rate as the less-sampled of the two.
• Counting. One source slot is one domain appearing in one answer. A domain cited twice in the same answer counts once.
• Failures. Rate-limited and errored calls are retried and are never counted as a negative reading. This has been true since 1 September 2026. Readings before that date predate retry handling, so some of them may record a call that quietly failed as a page that was not cited. Every rate here is therefore a floor.

What this dataset cannot tell you

• It is not a sample of what buyers ask. Every question is generated from one of our own list pages, so the category mix is our category mix — business software and fractional executive services, weighted the way our site is weighted.
• It is young. 309 readings starting 2026-08-27. The leaderboard tail is thin and individual domains near the bottom are close to noise. Treat counts under about five as indicative only.
• It measures citation, not traffic. Being cited is not the same as being visited, and we make no claim about the second.
• Our own numbers are confounded and we exclude them from the leaderboards. For the record: 18 generic and 55 comparison appearances for topelevens.com across the same readings. Reported separately and never ranked against the others. The questions are built from this site's own list titles, so a high count here measures the question, not the site.