Every guide to getting cited by AI search tells you to add schema, write in question form, and structure your content for machines. Almost none of them show you what an AI engine actually cites when you ask it a real question.
So we ran it. Eight prompts, taken straight from topics this site has published articles about, put to Google's AI Mode logged out on 1 September 2026. We recorded every domain cited in every answer — 55 citations across 48 unique domains — and checked whether BuzzRiding appeared in any of them.
It did not. Zero of eight. That part was expected. What the 55 citations show about who does get picked is the useful part, and it contradicts the thing everyone assumes.
🔬 How this was run
Google AI Mode (udm=50), logged out, from Ireland, on 1 September 2026. Every prompt and every cited domain is in the raw CSV, alongside the design note written before collection. Perplexity and ChatGPT both demanded a login and were dropped rather than run under an account you cannot reproduce.
Finding 1: AI Mode cites tiny sites constantly
The assumption behind most SEO advice is that big domains win. In these eight answers, they mostly did not.
Alongside Wikipedia, Forbes, Coursera and Semrush, the engine cited gowishy.com, mersel.ai, rankinllm.ai, aiclicks.io, brandviz.ai, geoscout.pro, syntermedia.ai, alkemyedtech.com, ankitpareek.com and leadde.ai — domains most marketers have never heard of, several of them clearly built in the last year.
Of 48 unique domains cited across the run, the large majority were small or unknown. Domain authority is a ranking constraint. On this evidence it is a much weaker citation constraint.
"Being small didn't stop a site being cited. Being absent from the topic did."
Finding 2: only three domains were cited more than once
Across 55 citations, exactly three domains appeared in more than one answer: youtube.com, linkedin.com and adlibrary.com. Every other domain was cited once and never again.
That is a fragmented citation graph. There is no small club of sources the engine keeps returning to on AI-marketing questions — which is good news for anyone trying to enter, and bad news for anyone assuming one strong article will compound across many queries.
Finding 3: YouTube is in five of eight answers
Video was cited on GEO ROI, GEO tools, model comparisons, AI skills and video tools. For a text-only site, that is five of eight answers where a meaningful share of the citation slots are simply unavailable.
We are not going to pretend that makes video a casual next step. It does mean any honest read of "how do we get cited" has to include the format question, not just the schema question.
Finding 4: the query we optimised for means something else
This one stung. Asked for best GEO tracker tools 2026, AI Mode returned a blend of generative-engine-optimization platforms and IP geolocation APIs — ipinfo.io, ip-api.com, apifreaks.com — because "GEO tracker" is ambiguous and the older, denser meaning of GEO pulls hard.
BuzzRiding has an article targeting exactly that phrase. It is competing in a query the engine itself does not resolve cleanly. Checking for that kind of collision takes one search, and we had never done it.
Finding 5: the cited set is not the ranking set
Several cited domains do not appear on page one of a normal Google search for the same phrase. The AI answer is assembled from a different, wider pool than the blue links. Reading the SERP and assuming you know the citation landscape is a mistake — they are related but visibly not the same list.
What we are changing because of this
- Re-run this monthly. Same eight prompts, same engine, logged out, published either way. One run is a snapshot; the trend is the actual asset.
- Add Perplexity and ChatGPT properly — with a stated account context, since logged-out access is closed.
- Check query ambiguity before targeting a phrase. One search, every time, before an article is briefed.
- Stop assuming authority is the blocker. On this evidence the blocker is being absent from the topic, which is a fixable problem.
💡 Key takeaway
Eight prompts, 55 citations, 48 domains, and a large share of them small and new. If AI answers are your route in, the question is not "are we big enough" — it is "are we on the topic, in the format the engine is citing".
The obvious caveat
n=8, one engine, one country, one day. This is a first run, not a study, and it is published as such. The prompts, the raw citation lists and the pre-registered design are all in the repository so anyone can re-run it and disagree with us.