Apr 22, 2026

Three AI Search Engines. Zero Agreement.

The phrase "AI search" gets used as if it describes one thing. It doesn't.

In the last 30 days alone, GetMentioned tracked over 411,000 citations across 26,000+ unique domains, drawn from ChatGPT, Perplexity, and Gemini. That's one month of AI citation data from our platform - a rolling snapshot of how today's AI answer engines actually decide who to recommend.

What came back wasn't three versions of the same engine. It was three fundamentally different retrieval systems - with different source universes, different content preferences, different levels of stability, and different blind spots.

If you're building a GEO strategy around "get cited by AI," you're optimizing for a category that doesn't exist. The better question is: which AI, in which content format, for which category?

Here's what the data shows.

Part 1: The three engines are working from three different internets

Start with the top-line numbers:

Gemini behaves as if it has a ~1,500-domain whitelist. ChatGPT pulls from 13× more sources. Perplexity sits in the middle on source count but cites more URLs total - meaning it recycles a tighter set of trusted domains heavily across queries.

The overlap between models is almost non-existent:

Only 474 domains were cited by all three models - less than 2% of every unique domain in the dataset.
Of ChatGPT's ~19K unique domains, 13,854 are exclusive to ChatGPT (never cited by the other two).
Of Perplexity's ~12K unique domains, 6,826 are exclusive to Perplexity.
Of Gemini's ~1,500 unique domains, 556 are exclusive to Gemini.

When two models are asked the same question about the same brand, they agree on only 12-19% of their sources. That's the base rate: most of the time, ChatGPT and Perplexity are pulling from different places to answer the same question. And they barely talk to the same sources Gemini does.

The practical implication: an AI visibility audit that only checks one of the three models is roughly as representative as auditing one search engine in 2005 and calling it a day.

Part 2: The source taxonomy shows three different content philosophies

Aggregate the 412K citations into source types, and each model's signature shape emerges clearly. Here's the full breakdown - platform-specific callouts preserved, content types split - as share of each model's total citations:

Three distinct retrieval philosophies fall out of this table:

ChatGPT is the authority-first engine. 68% of its citations point to brand and competitor product/homepage content - the default "go to the company's website" move. On top of that, it's the only model that meaningfully cites Wikipedia (0.27%), LinkedIn (0.41%), and traditional news (2.02%). If a source looks officially indexed and editorially legitimate, ChatGPT surfaces it. This is the model most legible to traditional SEO and PR playbooks. (For a deeper look at how ChatGPT chooses its sources, we've broken down the mechanism separately.)

Perplexity is the hybrid engine. It's the only model with meaningful citation share across every row: Reddit, YouTube, editorial blogs, marketplaces, listicles, news. It leans heaviest into third-party editorial content (11.3%) and is the only model that cites YouTube at scale (1.35%, ~10× ChatGPT's rate). Perplexity behaves like a social-aware search engine wearing an AI costume.

Gemini is the listicle engine. 36.77% of its citations are third-party "Top X / Best of" ranking articles. Add in its own-owned listicle preference (4.43%) and nearly 41% of Gemini citations are ranking content. Meanwhile it cites Wikipedia 0.01% (one time total across 14,350 citations), LinkedIn 0.00% (zero), and marketplaces 0.40%. Gemini ignores nearly every authority signal the other models trust, and leans hard on a narrow ecosystem of ranking content.

That last point deserves unpacking. The "Third-party listicle" bucket on Gemini isn't diffuse - it concentrates into a small cluster of domains. The top 15 listicle-publishing sites account for roughly 56% of all Gemini listicle citations; the top 50 account for ~79%. Many of these domains are small SEO-optimized agency sites that appear to have been built specifically to be cited - publishing "Top 10 [category] agencies of 2026" content aimed squarely at the AI retrieval layer.

There is, essentially, an AI-citation content farm forming. And Gemini is its biggest customer. (We've catalogued the seven content formats AI models actually cite - ranking listicles top the list, but the other six matter too.)

Part 3: Ten behavioral findings that break the "AI search" abstraction

Beyond source type, the three engines behave differently in ways that matter for strategy.

1. Perplexity has 2× the citation memory of the other two.

Day-over-day URL carryover - the share of URLs cited today that were also cited yesterday - averaged 67% on Perplexity versus 34% on ChatGPT and Gemini. Two out of three URLs Perplexity cited today, it cited yesterday. The other two engines reshuffle about two-thirds of their sources every day.

Once you earn a Perplexity citation, it tends to stick. ChatGPT and Gemini require continuous effort to stay in rotation - yesterday's visibility is not a reliable predictor of today's.

2. Gemini cites 13 sources per answer. The others cite ~50.

Median sources per individual query: ChatGPT 44, Perplexity 50, Gemini 13. At p90, the gap widens: ChatGPT 121, Perplexity 100, Gemini 34. Gemini is a ~4× narrower gate - much harder to get cited in - but each citation carries disproportionate weight on the final answer the user sees. (This is closely tied to query fan-out - how each engine breaks a single user question into multiple sub-queries before retrieving sources.)

3. Gemini almost never cites homepages.

Share of citations pointing to a bare root domain (homepage, no deep path): ChatGPT 9.4%, Perplexity 9.1%, Gemini 2.7%. Gemini wants a specific article, product page, or ranking list - it does not treat "the brand's front door" as a useful source. The other two will fall back to a homepage when they can't find anything more specific; Gemini appears to just skip you.

4. 64% of Gemini's top-100 domains don't appear in the other models' top-500.

Same measurement for ChatGPT: 36%. For Perplexity: 29%. Gemini isn't choosing different sources within the same shared universe. It's operating in a substantially different source universe - built primarily from ranking content and a specific ecosystem of SEO-built agency blogs that the other two models barely touch.

5. Own-domain capture rates across brands span 0% to 14%.

A consistent metric emerges from the data: what share of a brand's total AI citations come from the brand's own website? Across the ~25 well-measurable brands in the dataset, the distribution looks like this:

Zero. One brand with thousands of AI citations about their products had effectively none of those citations coming from the site they control. AI learned about that brand entirely from reviews, forums, publications, and competitors.

This is a measurable KPI any brand can now track: what percentage of your AI visibility comes from content you actually own? Most brands have never measured it. Most brands are at 1-3%. (Source attribution is how you actually surface this number.)

6. Two opposite content playbooks both work.

Among brands with high own-domain capture, the strategy splits cleanly in two:

Research-driven categories (B2B SaaS, professional services, diagnostics, dev services): ~70% of own-domain citations come from blog/editorial content.
Transactional categories (real estate, local services, e-commerce): ~99% of own-domain citations come from product, location, or listing pages - not blog content.

There is no universal "AI content playbook." Publishing a blog in a transactional category is wasted effort; publishing only product pages in a research category is wasted opportunity. Match content structure to the underlying query intent in your vertical.

7. A single URL can outperform an entire brand domain.

The top 5 individual URLs in the dataset each earned 480-750 citations in 30 days - and in several cases, a single third-party ranking article generated more AI citations about a brand than that brand's own homepage did across the same period.

The shape of these "citation magnet" URLs is consistent:

"Top X" style listicles on third-party blogs
Partner/directory pages on platform-owner sites (SaaS marketplaces)
Deep-linked forum threads with directly relevant titles

A single well-placed ranking article can be worth more AI visibility than a year of on-site content work. (We've broken down what these high-performing URLs have in common in the anatomy of content that gets cited by AI.)

8. Rankings and business publications outperform Wikipedia and G2 as cross-industry authority signals.

Excluding social platforms, the sources cited across the widest range of distinct brands in the dataset rank (by source type) roughly as follows:

A single placement in a business publication's ranking article feeds AI visibility for 20+ different brands in our dataset. A Wikipedia placement does much less. The traditional PR playbook (earn business-press coverage) and the traditional SEO playbook (get listed on review sites) both outperform the traditional "authority content" playbook (Wikipedia, encyclopedic references) for AI visibility.

9. News and Wikipedia combined are under 2.5% of all AI citations.

Across every model, the traditional "trusted authority" layer - Wikipedia, .gov, mainstream news - accounts for 2.34% of citations in aggregate. That's less than Reddit and YouTube combined (2.05%) and roughly the same as "Owned editorial content" (0.69% + some brand blogs).

The legacy authority hierarchy has been largely bypassed. Editorial content on brand blogs (9.63%) is roughly 4× more influential in AI citations than mainstream news. The "earn press coverage" mindset that most SEO and PR teams default to isn't wrong - business publications are genuinely high-leverage, per finding #8 - but it's working through a different mechanism than most teams assume.

10. Topic fragmentation is the real opportunity map.

For any given topic, you can measure how much citation share the single most-dominant domain holds. The distribution across topics in the dataset is wide:

Some topics in the dataset have a top-1 share of 2-3% (open field - any well-structured piece of content can earn share). Others sit at 17%+ (one domain owns the category). The variance is enormous, and it's predictive: the first diagnostic of any new GEO project should be to measure the top-1 citation share for the target topic. If it's over 15%, the strategy is displacement and is expensive. If it's under 5%, the strategy is arrival and is cheap.

Most teams skip this diagnostic and apply a generic "publish more content" prescription to topics that require very different approaches. (Topic segmentation makes this diagnostic trivial to run.)

Part 4: What this actually means for strategy

Pulling the findings together, three concrete implications fall out.

These are not three marketing channels that overlap. These are three genuinely different retrieval systems with minimal shared infrastructure. A strategy that optimizes only for the most-cited engine in your category will likely underperform on the other two. (For a practical framework on running all three in parallel, see how to dominate AI search in 2026, and for monitoring how competitors are doing on the same queries, competitor benchmarking is the starting point.)

Measure your own-domain capture rate. Track it monthly.

Of all the findings, the own-domain capture rate is the most actionable individual KPI. It requires no special tooling - just a source audit - and it gives a single number that summarizes how much of your AI visibility is under your control. (Mentions monitoring and weekly reports let you track this automatically instead of running manual audits.)

A reasonable benchmark from this dataset:

Under 2% - you do not control your own narrative. AI is describing your brand using third-party sources, likely with factual drift and uncontrollable framing.
3-7% - typical. You have some narrative control but most of the story is being told by others.
10%+ - strong. Your content strategy is being rewarded by AI retrieval.

Most brands measured here sit at 1-3%. The difference between that and 10%+ is not random - it correlates with structured publishing of the right content type for the vertical (per finding #6) and with deep-linked, specifically-titled URLs (per finding #7).

Diagnose topic fragmentation before building content.

The temptation is to pick the topic with the highest search volume and publish against it. In AI citation terms, this is often exactly wrong - high-volume topics are disproportionately likely to have an entrenched gatekeeper already. Lower-volume adjacent topics with fragmented citation distributions are significantly cheaper to win.

A practical workflow:

List 10-20 target topics in your category.
For each, pull the actual citation data for a week and measure: how many distinct domains are AI citing, and what's the top-1 share?
Sort ascending by top-1 share. Your most efficient wins are at the top of that list - the fragmented topics where no one has won yet.
Only tackle the high-concentration topics once you've built up authority on the fragmented ones.

The takeaway

"AI search" is a marketing abstraction that hides three engineering realities.

ChatGPT reads brand sites, publications, and reference material. Perplexity reads forums, blogs, videos, and marketplaces - and it remembers what it reads. Gemini reads ranking content and almost nothing else. These three systems do not, for the most part, cite the same sources, structure answers the same way, or respond to the same optimization signals.

If your current AI visibility work is built on a single, generic "GEO" playbook, the 412,000-citation view strongly suggests you're leaving most of the opportunity on the table - optimizing for a composite engine that doesn't exist, while three real engines with three very different appetites look elsewhere.

The content layer that AI actually cites is narrower, more idiosyncratic, and more measurable than the industry discussion implies. The good news: that makes it addressable. The bad news: addressing it requires running three strategies, not one.

See your own numbers

Every chart in this report was built from the same kind of data we generate for each brand on our platform. If you want to see where your brand sits on the own-domain capture index, which of the three engines is citing you most, and which topics in your category are fragmented opportunities versus entrenched ones:

Run a free AI visibility report - a 5-minute check across ChatGPT, Perplexity, and Gemini.
Start a free trial - track citations, sources, and competitors continuously.
Book a demo - walk through the data with our team.

Built for agencies, marketing teams, and in-house SEO teams.