Ask five people what "AI search" means and you'll get five different answers. Ask five AI engines to pick their AI citation sources for the same query and you'll get five different answers too, literally, not figuratively. ChatGPT, Google, Claude, and Gemini do not share a trust model. Each one has built its own idiosyncratic rulebook for which pages it's willing to name, and the rules don't just differ by industry or topic. They differ by engine, full stop.
One post on this site already went deep on the sharpest version of this story: why we told clients to stop chasing Reddit AI citations walked through Reddit's ChatGPT citation share crashing and the 99.39% OpenAI rejection number this post reopens below. That finding stands, and we're not re-running it. What it doesn't answer is why the same platform behaves so differently everywhere else engines look for sources, and that's the more useful question if you're building an AI citation strategy that has to survive contact with more than one engine.
Most teams report on "AI visibility" the way they used to report on Google rankings: one number, one trend line, one dashboard tile. That habit made sense when there was effectively one search engine to optimize for. It stops making sense the moment four separate systems are choosing sources on four separate sets of rules, because a rising blended number can hide one engine collapsing while another compensates for it. You cannot fix what a single average won't let you see.
Why AI citation sources diverge by engine, not by topic
Most GEO advice still talks about "AI search" like it's one system with one preference list: get on Reddit, get on Wikipedia, get a G2 listing, and the engines will find you. Dan Petrovic, founder of the Australian SEO firm DEJAN, published research on July 21, 2026 that breaks that assumption directly. His study tracked grounding-source selection, meaning which retrieved candidate pages an engine actually turns into a citation, across OpenAI, Google, and Anthropic. The conclusion wasn't that AI prefers or avoids any one source category. It was that each engine has learned its own trust profile, built from its own training data, its own retrieval pipeline, and its own editorial guardrails, and those profiles don't converge.
That's a structural claim, not a stylistic one. It means the fix for "we're invisible on ChatGPT" is not the same fix for "we're invisible on Claude," even when the symptom looks identical on a dashboard. Two brands with an identical backlink profile and an identical Reddit presence can get treated as trustworthy by one engine and functionally ignored by another, for reasons that have nothing to do with the quality of either page.
It also means the vendor pitching you a single "AI search ranking" score is selling you the same mistake in a new wrapper. A blended score across engines that behave this differently isn't a simplification, it's a number with the useful information averaged out of it. If OpenAI cites you at a near-zero rate and Google cites you generously, a blended score just tells you "medium," which tells you nothing about which engine to fix or how.
Reddit's rejection rate, and what it reveals about AI citation sources
Start with the number DEJAN's study makes hardest to argue with. Across OpenAI's searches, Reddit showed up as a candidate source, meaning it was retrieved and available to cite, in 76% of cases. It got selected and actually cited in just 0.61% of those. Petrovic counted 491,024 retrieved Reddit pages in the sample; 488,012 of them were rejected, a 99.39% rejection rate. On OpenAI's free-style prompts specifically, Reddit turned up in only 16 of 103,974 query fan-outs, 0.015%, and had fallen to zero by June 2026, per Dan Petrovic's DEJAN research.
Google tells a completely different story about the same platform. Petrovic's inferred range for how often Google selects Reddit as a source sits between 9% and 60%, which he describes as at least 14 times more often than OpenAI. Anthropic sits at the opposite extreme from Google: across 139,601 grounding sources sampled between May and July 2026, Claude cited Reddit zero times. Same content, on the same platform, three completely different outcomes depending on which engine is doing the choosing.
The 76% figure is worth sitting with too, because it's easy to skim past. Reddit isn't hard for OpenAI to find. It's retrieved as a plausible candidate in three out of four searches where it's relevant at all. The rejection happens after retrieval, at the selection step, which means OpenAI is not failing to see Reddit's content, it is actively deciding, page by page, not to name it. That's a different failure mode than invisibility, and it calls for a different fix. You are not trying to get discovered. You are trying to clear a filter that is already looking directly at you and saying no.
How often each engine actually cites Reddit once it's been retrieved as a candidate (DEJAN, Jul 21, 2026)
Wikipedia, arXiv, and the myth of a shared trust model
Reddit isn't the only source with a number attached. Across the full set of grounding sources DEJAN sampled, Wikipedia was selected 5.64% of the time and arXiv 0.77%, both of which sound low next to the reputations reference sites and academic papers carry in classic SEO, and both of which still clear Reddit's 0.61% OpenAI figure. Line those three numbers up and a pattern that looked like a Reddit story turns into an engine-behavior story: selection rates are low across the board, they vary by source, and they vary again by which engine is sampling that source.
| SOURCE | SELECTION RATE | SAMPLED BY |
|---|---|---|
| Reddit (OpenAI) | 0.61% | DEJAN, Jul 21, 2026 |
| Reddit (Google, inferred) | 9%-60% | DEJAN, Jul 21, 2026 |
| Reddit (Anthropic / Claude) | 0% (0 of 139,601) | DEJAN, Jul 21, 2026 |
| Wikipedia (all engines sampled) | 5.64% | DEJAN, Jul 21, 2026 |
| arXiv (all engines sampled) | 0.77% | DEJAN, Jul 21, 2026 |
None of this means underlying content quality is irrelevant. DEJAN's argument isn't that quality doesn't matter, it's that a shared, cross-engine ranking of "trusted source types" doesn't actually exist the way most GEO checklists assume. If you optimized a single UGC-heavy citation play expecting it to travel evenly across ChatGPT, Google AI Mode, and Claude, the 9%-to-60% range on Google alone should worry you: even within "Google trusts Reddit more," the range is wide enough to make a single number useless for planning.
We've written before about the anatomy of an AI citation and about the six link types AI engines trust as if there were one shared list every engine reads from. There isn't, fully. Some source types travel better than others, and Wikipedia's consistent, if modest, selection rate across the DEJAN sample suggests reference-grade citations are the closest thing to a universal signal available right now. But "closest to universal" still isn't universal, and the gap between 5.64% and 0% is exactly the gap a single checklist can't close.
Think about what that means for a real content calendar. A team that spends a quarter building forum-style, community-flavored content on the theory that "AI likes UGC" is optimizing for a rule that doesn't exist in any of the three engines Petrovic measured, not even the one, Google, where Reddit performs best. A team that spends the same quarter on reference-grade, sourced, citable pages is closer to the one pattern that actually holds up across OpenAI, Google, and Anthropic alike, even if it's a weaker pattern than most GEO advice implies.
Gemini's separate divergence: order bias, not topic trust
Reddit, Wikipedia, and arXiv are all about topic trust, meaning which source category an engine is willing to name. Gemini's divergence runs on a completely different axis. In research published July 26, 2026, DEJAN found that Gemini defaults to whichever web page it reads first 92% of the time. That's not a judgment about the source's authority or category. It's an ordering effect: the page Gemini processes earliest in its retrieval sequence wins the citation in nine out of ten cases, regardless of whether a later-read page made a stronger or more current claim.
Only content specifically restructured around the text-ordering preferences Gemini responds to was able to override that default in DEJAN's testing. That's a second, independent lesson stacked on top of the Reddit findings: not only does each engine trust different sources, at least one of them barely evaluates trust at all for a large share of decisions. It evaluates sequence instead. A page that would win on Claude for its sourcing and lose on Google for its selection rate can still win on Gemini for showing up first in the read order, or lose there for exactly the same reason on a different query.
Stack the two findings and the shape of the problem gets clearer. OpenAI runs something close to a strict filter: high retrieval, near-total rejection. Google runs a wider, more permissive selection band. Anthropic runs close to a hard exclusion for at least this one source. Gemini barely runs a trust evaluation at all for the majority of cases; it runs an ordering heuristic. Four engines, four different mechanisms, not four dialects of the same mechanism.
Practically, that means a page built to win Gemini looks different from a page built to win the other three. Instead of asking "is this source category trustworthy," the Gemini fix asks "what does this page put first." Front-load the strongest, most current claim. Put your comparison verdict, your pricing answer, or your direct definition in the first block of text a crawler reads, not after three paragraphs of scene-setting. On the other three engines that's simply good practice. On Gemini, per DEJAN's 92% figure, it's close to the entire game.
“OpenAI filters. Google weighs. Anthropic excludes. Gemini orders. None of those are the same problem, and none of them share a fix.”
Build a per-engine citation strategy, not one GEO checklist
The practical mistake sitting underneath most "AI search" reporting is treating mention rate as one number instead of four. A brand can be well-optimized for OpenAI's strict filter and still lose on Gemini because nothing on the page accounts for read order. It can win Google's wider Reddit band and still get zero credit from Claude for the exact same content. Reporting AI visibility as a blended average hides which engine is actually moving and which one is stuck, and it makes the fix look uniform when it isn't.
This is the case for building generative engine optimization services around per-engine diagnostics instead of one undifferentiated GEO checklist. For B2B brands especially, where a single lost citation on a comparison query can mean a buyer never sees you at all, that distinction is the difference between a strategy that compounds and one that plateaus once the easy technical fixes are done. It's the same discipline we used on our work with Arnica, where treating each engine's citation behavior as a separate variable, not a shared score, was what actually moved mention rate.
None of this is an argument for running four separate playbooks with four separate budgets. It's an argument for measuring before you build. Pull your last thirty days of AI-cited traffic, split it by engine, and look for the one where you're already invisible. That's where the next sprint starts, not with a rewritten Reddit strategy, and not with one more line item labeled GEO.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Tyler leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.