AI Search Engines Cited Different Websites in a 960-Answer Study.

Four AI search engines received the same 80 questions three times. They returned 8,579 cited URLs, but they did not search at the same rate or draw from the same source market. Nearly three quarters of the 1,329 cited sites appeared in only one engine.

The clearest contradiction appeared around video

In Prefer’s September 13 snapshot, Gemini cited YouTube in 113 of 240 answers. Perplexity cited it in none of 240. That contrast is difficult to explain as a general judgment about video quality because the same question set produced both results.

A separate Epovest longitudinal study found the same directional split over time. Perplexity’s share of sourced answers containing a video citation fell from 61.5% on August 10 to 0.7% on August 17. On nine travel trackers, 39 of 54 answers cited video on August 10 and none of 54 did one week later. From August 18 onward, only two of 276 tracked answers cited video. Epovest measured Gemini at 46.5%.

The two studies used different question sets and denominators, so their percentages should not be merged. Their agreement is narrower but still useful: both observed a near-zero Perplexity video presence after mid-August and a much higher Gemini rate.

The 960-answer snapshot held the questions still

Prefer published the study and its downloadable dataset under a CC BY 4.0 license. The test contained 80 questions across ten AI-search topics, three fresh runs per question, and four engines: ChatGPT using gpt-5.6-luna, Perplexity Sonar Pro, Gemini 3.8 Flash and Claude Sonnet 5.

Collection ran from 17:56 to 18:12 UTC on September 13 through the DataForSEO LLM Responses API. Web search was requested on every call, but the API did not report that every engine actually searched. These were API answers rather than captures from the consumer applications.

A source in this dataset is a unique cited URL inside one answer. The site-level counts remove www and merge subdomains only for Reddit, YouTube, LinkedIn and Wikipedia. Other subdomains remain separate because they can represent distinct properties.

Search activation controlled whether citations could appear

The four engines produced 960 answers. They searched on 861 and skipped search on 99. Every answer that skipped search contained zero cited sources.

Once search ran, a citation appeared in 858 of 861 answers. All searched Perplexity, Gemini and Claude answers included at least one source. ChatGPT had three searched answers with no source. This does not prove that search activation alone determines publisher visibility, but it identifies an upstream gate that citation-only monitoring cannot see.

The distinction supports the lifecycle described in our comparison of how conversational agents search the web. A missing citation can mean the engine never searched, searched but did not retrieve the domain, retrieved it without selecting it, or selected it without displaying it. Those are different failures.

The source markets separated after search

The repeated-run overlap is Jaccard similarity between pairs of runs for the same question when both runs cited at least one site. It measures short-window consistency, not long-term stability. Perplexity repeated nearly the same cited-site sets, while ChatGPT and Gemini changed much more between fresh calls made seconds apart.

The large difference in source volume matters too. Perplexity exposed about 6.4 times as many cited URLs per answer as ChatGPT. That does not establish that Perplexity’s evidence was better. It means a raw citation count gives engines with larger visible source lists more opportunities to credit a domain.

A larger B2B panel tests a different question

A Foundation and AirOps study updated September 26 classified about three million domain-level citation events in roughly 380,000 AI answers. Its prompts covered commercial-intent questions in 81 B2B software categories from June 22 to July 24. The six surfaces included Google AI Overview and AI Mode as well as ChatGPT, Claude, Gemini and Perplexity.

In that panel, social domains accounted for about 7% of citation events on Google’s two AI surfaces, but no more than 0.3% on ChatGPT, Gemini and Claude. ChatGPT’s community-domain share was 9.26%, compared with Claude’s 2.67%. The authors define a citation event as one domain appearing as a source in one answer. These figures therefore describe a particular buyer-prompt panel, not clicks, visits or sales.

The difference from the Prefer snapshot above is informative precisely because the numbers cannot be pooled. Prefer asked 80 AI-search-related questions through an API on one day, repeated each question three times, and counted cited URLs and answers citing specific sites. Foundation and AirOps tracked B2B buying prompts for several weeks and classified domain-level citation events. Google AI Overview also does not appear for every query, so its answer count has a different trigger condition from the standalone assistants.

For example, Prefer recorded no YouTube citations in its 240 ChatGPT answers, while Foundation and AirOps found a very small social-domain share across their much larger ChatGPT panel. That is not a reversal or a contradiction: zero in a narrow question set can coexist with a low nonzero share in a different topic mix. Likewise, Foundation’s category data show review-domain share ranging from 7.64% for managed databases to 25.07% for recruiting and applicant-tracking software. Topic selection can change the apparent source market before any platform behavior changes.

What to test next: run the same disclosed B2B prompts on each surface, record whether a search or AI Overview was triggered, and count both answer-level presence and domain-level citation events. Then repeat the prompts over time. Until a common prompt set and unit are available, the defensible conclusion is that both studies expose surface- and topic-specific citation mixes, not a universal ranking of source types. Foundation and AirOps have a commercial relationship, and the public article does not include the full prompt panel or complete methods.

The long tail diverged while the citation head converged

Across all four engines, 966 of the 1,329 cited sites appeared in one engine only. Another 205 appeared in two engines, 129 in three, and 29 in all four.

That looks like extreme fragmentation until citation frequency is considered. Of the 25 most-cited sites, 23 appeared in three or four engines. Among sites cited in ten or more answers, only 14 of 170 were exclusive to one engine. Much of the one-engine-only count came from the tail: 338 exclusive sites appeared in just one answer.

The useful reading is not “every engine uses a different internet.” Popular sources overlapped, while the long tail and platform mix differed sharply. A publisher can therefore face both a shared competitive head and an engine-specific discovery problem.

Build the denominator ledger before the visibility score

A cross-engine report should preserve at least five counts for every engine and prompt group:

1. answers requested;
2. answers that activated web search;
3. answers that returned any visible source;
4. answers citing the measured domain or platform;
5. unique cited URLs and sites.

Keep the model identifier, API or consumer surface, geography, timestamp and run number beside those counts. Otherwise, a change in search activation can be misread as a source-quality change, and a longer citation list can look like greater visibility even when the selection rate is unchanged.

Keep crawler access as a separate diagnostic layer too. Our AI crawler guide maps which agents request pages, but a successful crawl does not prove that an answer engine searched, retrieved or cited the page for a specific question.

Download the cross-engine source-market comparison CSV

The file separates the Prefer snapshot from the Epovest panel and leaves non-comparable fields blank. It is a compact comparison aid, not a merged benchmark.

What these studies cannot establish

• Prefer’s questions all concern AI search, GEO or AEO. The results do not represent every query class.
• Prefer measured API responses on one day, not the consumer applications or long-term behavior.
• Eight questions named Profound or Peec. Prefer reports that 29 of the 103 answers citing tryprofound.com came from those questions.
• Prefer and Epovest both sell AI-visibility products or services. Their disclosed methods and data improve inspectability but do not remove that commercial context.
• Neither study graded the factual correctness of every answer or proved why an engine selected a source.
• The two studies use different prompts, schedules and denominators. Their video results can be compared directionally, not pooled into one percentage.

The defensible conclusion is engine-specific

The same questions did not produce one shared citation market. Search activation, visible source capacity, repeated-run stability and platform preference all varied by engine. The strongest overlap appeared among frequently cited sites, while the long tail remained highly fragmented.

For publishers, that means one citation share should be treated as a summary after diagnosis, not as the diagnosis itself. Preserve the search gate and source-list denominator first. Then compare engines separately and repeat the test over time. Our YouTube citation comparison method applies the same discipline when the question is whether a cited video also appeared in a separately captured Google result set.

Study links

• Prefer AI answer study, September 2026
• Prefer downloadable JSON dataset
• Epovest YouTube citation panel