Why ChatGPT, Perplexity, and Google AI Mode Choose Different Sources for the Same Question
Five AI engines, same question, almost entirely different sources. Only 2.7% of domains get cited by all five. Here is what the data shows and what to do about it.
Ask the same buyer question to ChatGPT, Perplexity, Claude, Gemini, and Google AI Mode. You will get five answers that look similar on the surface. Underneath, the engines are pulling from almost completely different sets of sources. A study of 127,198 AI citations found that only 2.7% of cited domains appeared in all five engines. Seven in ten were cited by exactly one.
That is not a rounding error. It is the structural reality that every brand visibility program has to deal with right now.
"AI search" is not one thing
I run campaigns across these engines every week. The instinct most teams have is reasonable: treat AI search as a single channel, build one content strategy, measure one citation share number. The math says that instinct is wrong.
The evidence comes from multiple independent studies. Foglift's Q2 2026 benchmark put the same 75 buyer-intent prompts to ChatGPT and Google AI Overviews on the same day with the same wording. The average overlap between the two engines' cited domains was 4.1%. In 64% of those prompts, the two engines shared zero cited domains. Not low overlap. Zero.
BrightEdge's analysis found that for the same topics, the overlap between two engines' cited sources ran as low as 16%. An academic study of 55,936 queries across six LLM search engines confirmed that 37% of cited domains are unique to LLM engines and never appeared in traditional search at all. The divergence is not a quirk of one study. It is structural.
The same question. The same web. The same day. Almost completely different answer surfaces.
If your team is optimizing for "AI" as a single target, you are optimizing for an average that no engine actually is.
Every engine has a citation personality
This is not random variation. Conductor tracked seven AI engines for seven months (September 2025 through March 2026, 1,056 data points) and found that every engine has what they call a "persistent editorial identity": a default source type it reaches for, intent by intent, month after month. Semrush's analysis of more than 100 million AI citations confirmed similar patterns, with each engine weighting source types differently even for identical queries.
Here is what the combined data from these studies shows about each engine:
ChatGPT cites 3.7 sources per answer on average. It leans heavily on vendor first-party domains (41.5% of citations in Foglift's buyer-intent sample). When ChatGPT decides to attribute a claim about your product category, it goes straight to the brand's own site rather than the listicle that ranked it. It is also the only engine family that anchors on Wikipedia across multiple intent categories.
Google AI Overviews cites 9.3 domains per prompt in the Foglift sample. It leans aggregator: review sites, listicle hubs, and press media make up 22.3% of citations versus ChatGPT's 13.4%. AIO inherits Google's index breadth and its long-standing bias toward video, forum, and editorial results. 92.1% of the domains AIO cited were never cited by ChatGPT for the same prompts.
Google AI Mode sent 11.2% of its citations to YouTube and 4.0% to Reddit in the SurfacedBy study. It also routes users back to Google's own properties for purchase queries, something no other engine does. Google itself has acknowledged this fragmentation: the company recently brought Preferred Sources into AI Overviews and AI Mode, letting users pin favorite websites so they surface in AI answers, a tacit admission that the default source selection differs enough across Google's own products that users need a manual override. Our own Machine Relations Index data shows that market databases and research firms earn 1.6 to 1.9 times more citations in AI Mode than in AI Overviews. Same parent company, structurally different source logic.
Perplexity cites 8.6 sources per answer. It favors YouTube across most intent categories and pulls from a broader mix of editorial and documentation sources than either Google product. We wrote about how Perplexity selects its sources earlier this year, and the data here confirms the pattern: Perplexity runs its own crawler rather than relying solely on another engine's index, which is why its citation profile looks different from both Google products and ChatGPT.
Claude cites 6.8 sources per answer but sends just 0.02% of citations to YouTube and 0.01% to Reddit. Practically zero. Claude pulls from documentation, vendor pages, and editorial sites. If your buyers use Claude, a YouTube strategy does close to nothing. If they use Google AI Mode, it may be the whole game.
The comparison that matters
| Engine | Sources per answer | Top source type | YouTube share | Reddit share | Editorial identity |
|---|---|---|---|---|---|
| ChatGPT | 3.7 | Vendor first-party (41.5%) | Low | Low | Goes to the brand's own site |
| Google AI Overviews | 9.3 | Aggregators, listicles, press | Medium | Medium | Inherits Google Search DNA |
| Google AI Mode | 7.8 | YouTube, Google properties | 11.2% | 4.0% | Routes to Google's own ecosystem |
| Perplexity | 8.6 | YouTube, editorial, docs | 8.8% | Low | Broad, YouTube-forward |
| Claude | 6.8 | Docs, vendor pages, editorial | 0.02% | 0.01% | Almost never cites UGC |
| Gemini | 11.0 | YouTube across every intent | High | Low | YouTube-dominant |
Source data: SurfacedBy, Foglift, Conductor
Three of those engines are Google products. They share infrastructure, a user base, and a parent company. They do not share a source preference. Not for education queries. Not for purchase queries. Not for support queries. Google built three different editorial identities into three different AI surfaces.
What Reddit and Wikipedia actually contribute
The conventional wisdom puts Reddit and Wikipedia near the top of the AI citation food chain. 5WPR's index of 680 million citations ranks them prominently, and Pew Research found Wikipedia, YouTube, and Reddit among the most-cited sources in Google's AI summaries (with government sources appearing at 6% versus 2% in standard results). But the SurfacedBy data on commercial questions tells a different story: Reddit accounted for 1.8% of all citations. Wikipedia was under 0.6%. YouTube, at 4.9%, was the only major consumer platform that showed up strongly.
Where did the other 90% go? Vendor docs, product pages, and a very long tail of category-specific sites. The top 10 domains accounted for 20.6% of citations. The top 100 covered 42%. Nearly 43% of cited domains were cited exactly once.
If you are a B2B brand chasing Reddit threads and Wikipedia edits as your AI visibility strategy, the data says you are fighting for scraps that make up less than 2.5% of the citation pool for buyer questions.
There is one more thing the data breaks. Ranking number one on Google does not mean getting cited by AI engines. Ahrefs studied 15,000 prompts and found that only 12% of URLs cited by AI assistants rank in Google's top 10 for the same query. The overlap ranged from 28.6% for Perplexity down to 6 to 8% for the others. The AI citation game is not an extension of the SEO game. It is a different game with different rules on every surface.
The buyer intent layer makes it worse
The divergence between engines is not constant across the funnel. Foglift broke it down by buyer intent:
| Intent stage | Avg overlap | Prompts with zero overlap |
|---|---|---|
| Discovery | 4.1% | 56.0% |
| Shortlist | 3.5% | 72.0% |
| Variation | 4.6% | 64.0% |
Shortlist queries are where buyers compare options. They are also where the engines disagree the most: 72% zero overlap. AIO defers to third-party listicles (G2, review hubs, comparison sites). ChatGPT bypasses the listicle layer and goes straight to the vendor's own pages. Same "best X for Y" question, completely different citation logic.
That is the moment in the buyer journey where being cited matters most. And it is exactly where the engines diverge hardest.
What to do about it
Four moves follow directly from this data.
1. Pick the engine, then build the strategy. Find out which AI engines your buyers actually use. Then measure each one separately. ZDNET's practical guide walks through how to check each engine individually. A "share of AI citation" number that blends all five engines together tells you nothing actionable because the engines share almost nothing. The 2.7% overlap makes a blended metric a fiction.
2. Match the source type to the engine. YouTube and Reddit move Google AI Mode and Perplexity. Documentation and vendor-owned pages move Claude and ChatGPT. Aggregator listicles move AIO. Gracker's engine-by-engine citation pattern analysis breaks this down by content format. Build the content format that fits the engine your buyer is asking, not the format that feels easiest to produce.
3. Stop chasing Reddit for B2B buyer questions. At 1.8% of commercial citations, Reddit is a rounding error. Clear product pages, strong documentation, and category coverage do more for your citation count than any Reddit seeding campaign.
4. Mind the long tail. 43% of cited domains were cited exactly once. A specific page that answers your exact buyer question can get pulled in where one broad thought-leadership post never would. Specificity beats scale in the citation economy.
This is what Machine Relations is built for
The old model was simple: optimize for Google, rank, get traffic. One engine, one strategy, one channel. That model is over.
What replaced it is five separate AI engines with five separate editorial identities making five separate source-selection decisions on every buyer question. Each engine has its own retrieval logic, its own index biases, its own default source types. Managing those relationships, understanding what each engine values, and building the content that earns citation across the right ones for your buyers: that is the work. We have written about why AI search engines recommend some brands over others, and the answer is the same every time: the brands that win understand the source-selection logic of each engine they care about.
That is Machine Relations.
It is not one optimization. It is the discipline of managing how machines decide whether your brand is worth citing. And the data in this piece proves that "machines" is plural in every way that matters.
You either build the strategy per engine or you hope that the one engine you happen to be visible on is the one your buyer happens to use. Hope is not a strategy. The 2.7% overlap means it is barely even a guess.
FAQ
Do all AI engines use the same web index?
No. While the underlying web is the same, each engine retrieves from structurally different subsets. SurfacedBy's study of 127,198 citations found 69.6% of cited domains appeared in only one engine. The retrieval and ranking logic creates almost entirely separate citation universes.
Which AI engine cites the most sources per answer?
Gemini cites an average of 11.0 sources per answer, making it the most citation-dense engine. ChatGPT is the least dense at 3.7 sources per answer. More sources per answer means more chances for a given page to appear, but the source type has to match what that engine prefers.
Should B2B brands focus on Reddit for AI visibility?
For buyer-intent questions, Reddit accounts for just 1.8% of AI citations in the SurfacedBy commercial sample. Vendor documentation, product pages, and category-specific content make up the vast majority. Reddit may be visible on Google AI Mode (4.0% share), but it is nearly invisible to Claude (0.01%) and ChatGPT.
How different are Google's own AI products from each other?
Google AI Overviews, AI Mode, and Gemini share infrastructure and a parent company but have developed distinct editorial identities. Conductor's seven-month study found they choose different source types for the same intent categories. AI Mode routes to Google's own properties for purchase queries. Gemini defaults to YouTube across every intent. AIO inherits traditional Google Search ranking signals.