AI Citation Sources: Engines Barely Cite the Same Sites

01The engines do not share a source list

The working assumption behind most AEO advice is that AI search is one channel. The citation data says it is several.

A practitioner working on this professionally described running a large comparison and finding far less common ground than expected:

The same 11 percent figure circulates in AEO communities attributed to a 2026 analysis, which is either corroboration or the same number travelling. We could not establish which, and are reporting it as a practitioner claim rather than an established finding.

What is corroborated across several independent accounts is the direction and the rough shape of each engine’s preferences:

Note the confidence column. Most of what circulates about per-engine citation preferences rests on one vendor study each, and vendors in this category have an obvious interest in the answer.

02Why retrieval architecture drives the difference

The divergence is not arbitrary. It follows from how each engine actually gets information at answer time. A practitioner guide summarising this put the mechanism plainly: the engines are not reading the same internet, because they have different retrieval architectures.

• An engine backed by a search index inherits that index’s authority signals, which is why encyclopedic and established-publisher sources appear so heavily.
• An engine that crawls continuously can surface recent discussion, which is why community sources rank higher in its citation mix.
• An engine leaning on training data has a knowledge cutoff and cites less live material, favouring what was well-represented in training.
• An engine built on top of organic search stays closest to conventional rankings, which is why traditional SEO carries over there better than anywhere else.

This is why the volume of citations per answer also differs so much. One analysis put Perplexity near 21 sources per response against roughly 7 for ChatGPT. That has a direct strategic consequence:

An engine citing three times as many sources is three times easier to appear in and considerably harder to dominate, because each citation carries less weight in the final answer. Entering the citation set is not the same as being the recommendation, which is the distinction covered in measuring AI search visibility.

03How to find your category's real sources

This takes about an hour and needs no tooling. The output is a ranked list of the domains that actually decide recommendations in your category, which is almost always different from the list you would have guessed.

A practitioner working on citation tracking described the manual method clearly, and the critical instruction is in the third sentence:

The procedure:

1. Write 15 to 20 unbranded buying questions. Category questions, not brand lookups. A brand-name query returns your own site and teaches you nothing.
2. Run them in each engine that matters to your buyers. Perplexity exposes numbered citations directly. For ChatGPT, phrase questions so they require current information and then ask which sources were used.
3. Log the domain behind every citation. Not the page. The domain is the unit that repeats.
4. Count and rank. After twenty questions the distribution is usually stark: a handful of domains carry most of the citations, with a long tail appearing once each.
5. Separate the reachable from the unreachable. Some of your top sources will be places you can realistically appear (review sites, community threads, industry write-ups). Some will not be. Only the first group is a work list.

04Why every published list expires

The most important thing to understand about citation source data, including everything on this page, is that it has a short shelf life.

The clearest recent example: practitioners tracking citation share reported that Reddit held roughly 3.8 percent of ChatGPT Search citations from mid-July into early August 2026, then fell to about 0.5 percent by 14 August. That is an approximately 86 percent relative decline within about two weeks, affecting a source that had been among the engine’s most-cited for two years.

One operator running per-brand citation tracking reported the same pattern in a live account, with Reddit citations at zero for a client in the current window and the engine leaning on primary and manufacturer sources instead. They were careful to caveat it as a single-account snapshot, which is the right level of confidence.

Two conclusions follow, and the second is the useful one. First, treat every most-cited-domains listicle as a dated snapshot rather than a structural fact. Second, the durable practice is not knowing the list, it is running the measurement. A team measuring its own category monthly would have detected the Reddit change within weeks. A team working from a published list would still be optimising for it.

05Building on the overlap set

Building separately for every engine is not affordable for most teams. The efficient structure is a foundation plus a layer.

The practitioner who ran the 100,000-prompt comparison drew the same operational conclusion, recommending starting with the sources that appear across engines and building platform-specific tactics on top of that base. Domains repeatedly named as appearing across multiple engines include Reddit, Wikipedia, YouTube, G2, and major business press.

Choosing the second layer requires knowing which engine your buyers use, which self-reported attribution answers better than any benchmark. That method is covered in AI search ROI and attribution.

One caution on the overlap set: it is exactly as unstable as anything else here. The August ChatGPT change removed a major overlap source from one engine without warning. A foundation built on cross-engine sources is more robust than a single-engine bet, not permanently safe.

06Reading a competitor's citation sources

The same method points at a competitor and answers a question no rank tracker ever could: not that they beat you, but what the engine read before deciding they should.

Run the unbranded buying questions, and when a competitor is recommended, log the citations attached to that answer. Over twenty questions this produces their effective source set, and comparing it against yours produces a specific, checkable work list:

The last row is the one that most often surprises teams, and it is where the work sits in categories dominated by review sites and community discussion. We cover reading those surfaces in competitor review analysis and the underlying question of why an engine prefers a rival in why ChatGPT recommends your competitors.

07The Reddit question, specifically

Reddit deserves its own section because it is simultaneously the most-cited community source across engines and the clearest recent example of how fast this can change.

The case for it: Reddit appears in nearly every cross-engine overlap list, and Perplexity in particular has been reported to lean on it harder than any engine leans on any single domain. One tracking programme covering more than 200 million prompts across five months described Perplexity citing Reddit in up to 20 percent of responses, against a peak near 5 percent for any single domain in ChatGPT.

The case against relying on it: the August 2026 drop in ChatGPT removed most of that engine’s Reddit citations in about a fortnight. If your entire AEO strategy had been Reddit presence, one engine went dark on you with no notice and no announcement.

The honest position: community discussion is a genuinely important citation surface in most B2B categories, it is not controllable the way owned content is, and its weight varies by engine and over time. Being discussed accurately in the threads where your category is debated is worth real effort. Treating any single platform as a durable strategy is what the August data argues against.

We should disclose our own position here: Linkeddit operates a Reddit data pipeline, so we have an obvious interest in Reddit mattering. That is exactly why this section reports the decline as prominently as the case in favour.

08What to actually do about it

The findings above collapse into a short list of things worth doing, in order.

1. Measure your own category before reading another study. Twenty unbranded questions per engine. Your list will differ from every published list, because those aggregate across categories that behave nothing like yours.
2. Fix the reachable sources first. A review-site profile that is thin or outdated is usually the fastest fix available, and review sites appear in most categories’ top citation sets.
3. Answer the question on your own domain. Engines cite pages that answer the buying question directly. Product pages rarely do; comparison and method pages often do.
4. Re-measure monthly with the same questions. This is the only defence against the volatility described above, and it is cheap once the question set exists.
5. Do not chase a single engine’s quirk. Any tactic aimed at one engine’s current preference is a bet that the preference persists. August 2026 is the argument against that bet.
6. Keep a dated log of what you found. The value compounds only if you can compare this month against three months ago. A spreadsheet with a date column beats a dashboard you cannot export, because the whole point is the trend rather than any single reading.

09Frequently asked questions