ChatGPT, Perplexity, Gemini, Claude and Bing Copilot all answer the same prompts. They do not all cite the same sources. Each engine has a learned taste — a measurable bias toward specific kinds of domain — and a brand that wins on one engine can be entirely missing from the others.
We sampled 240 answers across the five engines over the last 90 days, 60 prompts spanning commercial-intent, how-to and fundamentals. This is the map of the skew, and what to do with it when planning content.
Why the engines disagree
Three reasons.
First, the retrieval indices are different. Perplexity has its own crawler and its own retrieval stack. ChatGPT’s browse mode taps Bing under the hood. Gemini taps Google. Claude leans on a similar stack to ChatGPT’s plus Anthropic-side curation. Bing Copilot is the most direct expression of Bing’s index.
Second, the rerankers are different. Each model was post-trained on examples of “good answers” curated by its own team. Whatever signals the curators valued — recency, snippet quality, brand familiarity, source authority — got baked into the reranker’s preferences.
Third, the safety and policy layers differ. Each model has a set of source types it actively prefers or actively discounts. Claude’s safety guidelines push it toward primary research and established publishers. Perplexity’s product positioning incentivises citing community sources (Reddit) more aggressively than the others.
You can argue with the reasons. You cannot argue with the data. The five engines pull from systematically different source pools and the differences matter for any AEO programme.
Perplexity — Reddit and primary sources
Perplexity’s signature in our sample. 21% of unique cited domains were either reddit.com or a Reddit subdomain. That is roughly 4× the next engine on Reddit citation share. The other big over-index — primary-research publishers and original-data blogs. If your content is “we ran a study, here are the numbers,” Perplexity rewards it disproportionately.
Two consequences for the planning.
If you are targeting Perplexity, Reddit is part of the playbook, not a bonus. A subreddit thread on your topic that names your brand — and is well-upvoted — frequently outranks your own product page on the same query. Engage in the subreddits, do not just publish blog content.
Original research is the second-highest Perplexity unlock. A 90-day study you ran in-house, a benchmark you computed, a survey you fielded — those pages get cited disproportionately on Perplexity. The investment is real but the unlock is engine-specific. If your portfolio has one piece of original research a year, route it through Perplexity-friendly structure first.
For the operational details, Perplexity citations tactics covers the cleaner methodology.
ChatGPT — established domains and Wikipedia
ChatGPT showed the highest combined share of citations to .edu, .gov, and Wikipedia — roughly 18% across our sample. It is the most conservative of the five on source selection. New domains take longer to earn citations on ChatGPT than anywhere else, but once earned, they hold longer (also the engine with the longest measured half-life on tactics content).
What wins on ChatGPT.
Authority signals matter more than freshness. A 14-month-old guide on a domain with strong inbound link authority will outperform a 14-day-old guide on a new domain — even if the 14-day guide is sharper. This inverts most of what works on Perplexity.
Wikipedia citation is a multiplier. Pages whose own Wikipedia entries are well-maintained tend to win disproportionately on ChatGPT, because the model can cross-reference the page against the Wikipedia summary and the consistency gives it confidence.
Bylined experts with strong off-site presence beat unbylined corporate content. Named experts with LinkedIn profiles, conference talks and verifiable credentials are explicitly weighted higher.
Gemini — news domains and AIO overlap
Gemini is the most news-biased of the five. 44% of cited domains in our sample were established-media outlets — Reuters, Bloomberg, sector trade press. This share is roughly double what ChatGPT or Claude show on the same prompts.
The cause is the Google retrieval stack underneath. Gemini’s retrieval is close cousin to the retrieval Google uses for AI Overviews, which is itself biased toward news domains for any prompt with a recency signal. If you are showing up in Google AIO, you are probably showing up in Gemini. The reverse is also true.
For brands without an established-media presence, two routes.
One — earn coverage on the trade press in your category. A single piece of independent coverage from a sector-relevant publisher carries more weight on Gemini than ten pieces of owned content. The HARO / Featured.com pitch templates in our outreach pack are calibrated for this exact lift.
Two — accept that Gemini will be your weakest engine until the off-site work matures, and weight your measurement and reporting accordingly. Tell the client honestly that month-1 Gemini citation count will lag the other engines by a quarter or more.
Claude — long-form and structured depth
Claude showed an average cited-page length around 2,100 words median. Compare that to Perplexity at 1,200 and ChatGPT at 1,500. Claude is the friendliest engine to longer-form content.
But length alone is not the signal. The longer-form pages Claude cites are also the most structured — proper H2 hierarchy, definition lists, Quick Facts tables, named-expert bylines. Long-and-rambling does not win on Claude either. Long-and-structured does.
The other Claude signature — original analytical reasoning. Pages that build an argument across multiple H2 sections, with each section adding a new observation or constraint, tend to win on Claude over equally long pages that simply list items.
Practical move. If your portfolio has one or two pieces that are 2,500+ words with deep H2 structure, those are your Claude bets. The shorter tactical pieces will not earn Claude citations as consistently as they earn Perplexity ones.
Bing Copilot — Microsoft properties and the Reddit secondary
Bing Copilot is the most direct expression of Bing’s index, so the bias maps to what Bing prefers — Microsoft properties (LinkedIn pulled disproportionately, MSN), and a strong secondary signal from Reddit.
The LinkedIn weight is the most actionable. A complete LinkedIn Company page, with named team members posting in their own voice, contributes more on Bing Copilot than on any other engine. Bing reads LinkedIn aggressively. We have watched no-name domains win Copilot citations through their team’s LinkedIn posts before their own site ranked.
The MSN bias is harder to act on directly. You do not write on MSN. But if a trade publication you are pitched into syndicates onto MSN, you get the lift twice — once on the publication, once on the MSN copy.
Bing also handles llms.txt well, possibly best of the five. The llms.txt spec piece covers the format. If you only ship llms.txt for one engine, ship it for Bing first.
How to plan for the skew
Three strategic options.
Specialise on two engines. The realistic move for most brands. Pick the two engines your buyers actually use — usually Perplexity + ChatGPT for B2B SaaS, Gemini + ChatGPT for content-marketing-heavy categories — and tune for them. Accept partial coverage on the other three.
Run a portfolio of bets. A bigger team can route specific content types to specific engines — original research to Perplexity, long structured analysis to Claude, news-shaped commentary to Gemini, LinkedIn-team-voice content to Bing Copilot. Each engine gets the content type it rewards.
Build the entity layer first. All five engines reward a brand with a clean entity record (Wikidata, LinkedIn, clean anchor text, schema sameAs). If you are not sure where to start, start there — the lift is universal. Then specialise.
The mistake is “we’ll write for everyone.” Nobody is the customer for content written for everyone, and no engine specifically rewards it.
What this changes about measurement
You probably already track citations per engine in your dashboard. The skew above adds nuance.
Watch the per-engine citation distribution, not just the total. A brand at 30 active citations is in different shape if those are 25 Perplexity + 5 ChatGPT (heavy on one engine) vs 6 per engine evenly (balanced).
Watch the trend per engine. Citation share shifting toward your weakest engine is a leading indicator that your off-site or structural work is paying off. Citation share consolidating on one engine is a leading indicator that you are getting brittle.
For the full measurement stack — what to log per engine, how to surface this in a dashboard, how to interpret the curves — see measuring AI citations and the citation half-life study. The skew above is the lens that makes those metrics actionable.
Each engine has a taste. Plan for the tastes you can serve. Accept the ones you cannot.