ChatGPT, Google AI Overviews, and Perplexity do not cite the web in the same way. Before trusting an AI visibility percentage, inspect its denominator, query set, time window, and citation unit.
AI visibility reports increasingly arrive with precise-looking percentages: Reddit accounts for one share of citations, Wikipedia another, and a publisher or review site appears to dominate a particular answer engine.
The numbers look comparable. Often, they are not.
A citation percentage is only meaningful after we know what was counted, which prompts were tested, when the observations were collected, and what population sits under the percentage sign. Without that context, two accurate studies can appear to contradict each other while measuring different things.
That is the denominator problem, and it matters for every team investing in answer engine optimization (AEO), generative engine optimization (GEO), or large language model optimization (LLMO).
Consider three useful pieces of research now circulating in the AI visibility market.
Nick Lafferty's 2026 comparison of AI visibility platforms ranks products using an AEO score built from criteria such as answer-engine coverage, citation tracking, prompt monitoring, analytics, and workflow capabilities. It is a buyer's framework for evaluating software platforms. It is not, by itself, a universal census of every citation produced by every AI system.
Profound's citation-pattern study analyzes more than 680 million citations collected between August 2024 and June 2025. Its central finding is that the major answer platforms have distinct sourcing preferences. Google AI Overviews leaned more heavily toward Reddit, ChatGPT frequently cited Wikipedia and Reddit, and Perplexity drew from a broader mix that included community and editorial sources.
Semrush later published a separate three-month study based on more than 100 million citations collected from July through October 2025. It examined citation behavior across ChatGPT, Google AI Overviews, Google AI Mode, and Perplexity, with breakdowns by platform and industry.
All three resources are valuable. But their percentages should not be placed in a single table and treated as if they came from one instrument.
Every reported share can be expressed as:
Citation share = matching citations / eligible citations in the measured dataset
The numerator may be easy to understand. The denominator is where comparability breaks.
“Eligible citations” could mean:
Change any of these choices and the percentage can change, even when the underlying behavior has not.
The denominator is only part of the story. The products themselves use different retrieval and answer-generation systems.
ChatGPT may answer from model knowledge, browse the live web, or combine retrieval with synthesized reasoning depending on the product mode and query. A study that collects citations from browsing-enabled answers is measuring a narrower behavior than “everything ChatGPT knows.”
Google's AI experiences sit close to its search index and ranking systems. They can surface sources that satisfy search intent, freshness, locality, and authority signals. But AI Overviews and AI Mode are not identical interfaces, and grouping them together can obscure meaningful differences.
Perplexity is designed around visible web citations. Its answers often contain more source links than other systems, which can change both the absolute number of citations and the distribution of domains inside the sample.
If one platform returns eight citations per answer and another returns two, a cross-platform percentage can be heavily influenced by the platform that emits more links.
AI products change quickly. Retrieval partners, ranking logic, model versions, interfaces, and citation policies can shift within weeks.
The Profound dataset spans August 2024 to June 2025. The Semrush study covers July to October 2025. That sequence is useful because it offers observations from adjacent periods, but it does not create an apples-to-apples replication.
A domain's citation share can move because:
A percentage without a date range is not a benchmark. It is an orphaned statistic.
The prompt set is the research equivalent of an audience.
Ask mostly software-development questions and GitHub, Stack Overflow, technical documentation, and developer publications may dominate. Ask consumer product questions and review sites, forums, retailers, and video platforms may rise. Ask health questions and institutional or clinical sources should become more prominent.
Even within one industry, prompt intent matters:
A domain-level percentage can therefore reveal as much about the study's prompt design as it does about the answer engine.
Teams should ask whether a report counts URLs, domains, answers, prompts, or citation events.
Suppose Wikipedia appears three times in one answer and once in another. Depending on the methodology, that may count as four citation events, two answers containing Wikipedia, or one unique domain. Each calculation answers a different question.
The same issue appears with subdomains and syndication. Are support.example.com and blog.example.com one source or two? Are copied articles consolidated under a canonical publisher? Are tracking parameters removed? Methodological choices determine the output.
Before repeating a percentage in a strategy deck, ask seven questions.
Is it citation events, unique URLs, unique domains, or answers containing a source?
Is the share calculated within one platform, across all platforms, within one industry, or across the entire dataset?
“Google” could mean classic results, AI Overviews, or AI Mode. “ChatGPT” could refer to browsing-enabled responses or another product configuration.
Look for industry, geography, language, intent, brand status, and prompt volume.
Prefer a precise collection window over a publication date alone.
Check treatment of repeated links, subdomains, canonical URLs, redirects, and duplicated content.
The strongest reports explain enough of their method for another analyst to test a comparable sample.
The wrong response is to chase whichever domain leads the latest chart. The better response is to understand why that source type satisfies a particular query and platform.
If community sites appear often, the lesson is not “manufacture Reddit mentions.” It may be that answer engines seek first-hand experience for comparison and recommendation prompts. If Wikipedia is prominent, the lesson is not “copy an encyclopedia.” It may indicate demand for stable, well-structured entity information. If technical documentation performs well, clear task-oriented answers may be the underlying advantage.
A durable AI visibility program should therefore:
Citation visibility is a diagnostic metric. It is not the business outcome.
When I present AI visibility data, I recommend attaching a compact methodology label to every percentage:
12.4% of citation events, within Perplexity answers, from 2,000 non-branded English prompts in the North American SaaS category, collected weekly from May 1 to June 30, 2026, with duplicate URLs counted once per answer.
That sentence is longer than “12.4% citation share,” but it is also usable. A colleague can understand it, challenge it, and attempt to reproduce it.
For executive reporting, keep both layers:
Precision should reduce ambiguity, not hide it.
The emerging research from Profound, Semrush, and independent platform reviewers points to a consistent strategic truth: AI answer engines do not share one universal source hierarchy.
But the percentages inside these studies are not interchangeable. Dataset size, collection period, platform mix, prompt design, geography, and counting method all shape the result.
Before asking, “Which domain wins AI search?” ask a better set of questions:
That shift turns a striking statistic into a defensible decision.
Methodology note: the studies above use different datasets and observation windows. Their percentages should be interpreted within each study, not combined as a single longitudinal series.
They may use different answer engines, prompts, industries, geographies, collection periods, and counting units. Two valid percentages can disagree because their denominators are different.
The denominator is the eligible population against which matching citations are divided. It may include citation events from one platform, all citations across several platforms, answers containing a domain, or unique URLs in a defined prompt set.
Not without normalization. The cited Profound and Semrush reports use different dataset sizes, collection windows, platform mixes, and methodologies, so each percentage should be interpreted within its own study.
It should disclose the numerator, denominator, answer experience, prompt population, geography, language, collection dates, model or product configuration, and citation-normalization rules.
Use it to diagnose where and why a brand appears, compare changes within a consistent methodology, and improve useful content and third-party authority. Do not treat citation share alone as the final business outcome.
