Two placements ran the same month, for the same client, in outlets of comparable size. One named the client and its category; the other covered the client's category and never named the client. Gemini, Perplexity and Google's AI Overview cited the first. Nothing cited the second.
That single pair is not the finding. It is what the finding looks like at the scale of one placement, and a two-run test built to check whether it holds at scale says it does.
What was tested: methodology and two runs #
Two independent measurements, four days apart in the same week, on live non-listicle earned-media placements already running for paying clients, across the six surfaces that expose a source URL in their answer: ChatGPT, Perplexity, Microsoft Copilot, Google's AI Mode, Google's classic AI Overview and organic result, and Gemini. Each surface documents its own citation mechanism differently — Google grounds an overview in supporting web results it links to, Copilot attaches inline citations to the web sources it retrieves — but all six expose a checkable exact-URL citation, which is what this test reads.
Run one. 12 placements. Each article was read first, then three buyer questions were written from what it actually said: the topic the way a buyer would phrase it, the article's most distinctive claim, and a narrow question about the company itself. That produced 36 questions, asked once each on all six surfaces (the AI Mode reads came through DataForSEO's AI Mode SERP endpoint) for 216 answers, each checked for an exact-URL citation back to the placement, matched on URL, outlet domain and organic rank.
Run two. 57 placements, the same three-question construction per article, the same six surfaces, 995 answers total. This run tracked three question categories explicitly: the article's own angle with no company name, the article's angle plus the company name, and a question about the specific story the article told.
Every answer was read for an exact-URL hit, not a mention of the outlet or the client's name in passing. A domain appearing in an answer's prose without a link to the placement did not count.
The numbers #
Run one: 17 exact-URL citations landed on 9 of the 36 questions, spread across 7 of the 12 placements. By surface: Perplexity 4 of 6 outlet citations, ChatGPT 3 of 3, Google's AI Mode 3 of 3, the classic AI Overview 3 of 4, Gemini 2 of 2, Copilot 2 of 2 — no single surface carried the result, which matters, because a test run on one engine would have reported a different number by accident of which engine it picked.
Run two, at five times the sample: 41 of 57 placements were cited by exact URL on at least one of their three questions.
| Question category | Run one (of 12 placements) | Run two (of 995 answers, hit rate) |
|---|---|---|
| Names the company or a term the article coined | 6 of 7 placements cited | 17% of answers |
| Article's own angle, company unnamed | — | 11% of answers |
| The specific story the article told | — | 7% of answers |
| Generic category question | 3 of 29 questions cited | not separately broken out |
The company-naming column and the generic-category column in run one are the two ends of the same test: 6 of 7 placements that had a question naming the company were cited at least once; the 29 generic questions across the full set of 12 placements returned a citation on 3 of them. Run two's category breakdown, measured per-answer rather than per-placement, lands in the same order — branded-angle questions outperform unbranded-angle questions, which outperform the narrowest single-story questions — and confirms the pattern at a sample five times larger.
The clearest pair in run one ran through two outlets by name. A Family Handyman placement covered a client's home-services category with no company name in the piece; it was cited by nothing across its three questions. A USA Today placement for the same client, the same month, named the company; Gemini, Perplexity and the AI Overview all cited it. Comparable outlet tier, same client, same month — the variable that moved was whether the question and the article both named the company, not which publication ran the piece. The other cited placements in run one ran in The Next Web, Inc. and SourceForge; the placements that were asked about on generic category questions and never cited included outlets in LA Weekly, TechTimes, The Hacker News and TechBullion.
One outlier ran the pattern the other direction: a placement that explained and named a term the client had coined for its own category was cited on the company-free version of that question too, because the term itself functioned as a unique identifier the way a company name does. A generic category question with no coined term and no company name behaved like every other generic question in the set: unanswered by a citation.
What Google's AI Overview did differently #
Google's AI Overview was the one surface in run one whose citations lined up with classic organic ranking: all three of its citations went to placements that already ranked in Google's own top 10 for that question, and it cited nothing outside that top 10 across the other 33 questions. ChatGPT, Perplexity, Copilot and Google's own AI Mode all cited placements that did not rank organically at all. The AI Overview's citation behavior in this sample tracked one input the other five surfaces did not need.
A finding neither run was built to produce #
Run one's third question per placement asked about the general topic the placement covered, not the client. On the three questions built this way for one particular placement, the placement itself was cited zero times. Two owned research and blog pages — one on how AI search engines decide what to cite and one on the source-role coverage gap in AI citations — were cited six times across the same three questions on the same six surfaces, including one verbatim quote from Gemini. Neither page was written to compete with the placement; both simply already answered the general version of the question the placement was trying to answer indirectly. A placement that explains a category is competing, without knowing it, against an owned page built to do exactly that.
Limits #
Run one is one answer per question-surface cell, no repeat runs, so single-cell results carry no variance estimate; run two adds scale but keeps the same one-run-per-cell design. Four placements in run one sat behind bot walls and returned no readable content, so they were excluded rather than scored as non-citations. September placements in both runs may not be fully indexed by every surface yet. One answer in run one failed to return and was excluded. Neither run tested whether a citation, once earned, persists at 7, 14 or 30 days — that is the next measurement, not this one.
What this does and does not say about outlet choice #
Nothing here measures whether a national outlet outperforms a trade outlet at equal specificity — the pairs that varied outlet tier in this sample also varied company-naming, so the two are confounded in run one's data. What both runs isolate cleanly is the company-naming and category-generic contrast, because that variable was varied directly by the questions written, not incidentally by which placements happened to run where. This is the fidelity counterpart to a pattern already measured on the source-selection side: the question shape decides which source an engine reaches for in the first place. This test shows the same shape effect decides whether a specific already-published page gets reached for at all.
FAQ #
Does the publication a placement runs in determine whether AI search cites it? Not on its own, in this sample. The variable that separated cited from uncited placements was whether the buyer's question named the company or a term the article owned — not which outlet published it.
Do generic questions about a category ever return a placement citation? Rarely. 3 of 29 generic-category questions in the smaller run returned a citation; the larger run's unbranded-angle questions were cited in 11% of answers, versus 17% when the same angle included the company name.
Which AI surfaces cited placements that did not rank in Google's organic results? ChatGPT, Perplexity, Microsoft Copilot and Google's own AI Mode all cited placements with no organic ranking at all. Google's classic AI Overview cited only placements that already ranked in the top 10 for that question.