Why each AI engine cites different sources

By Miguel Maio · Published 15 October 2026

Source overlap, by engine pair

1. 1Perplexity and AI Overviews: 24% of sources shared
2. 2Gemini and AI Overviews: 22%
3. 3Gemini and Perplexity: 17%
4. 4ChatGPT and Perplexity: 13%
5. 5ChatGPT and AI Overviews: 13%
6. 6ChatGPT and Gemini: 12%

No pair reaches a quarter. Only 3.8% of sources are cited by all four engines.

Ask the same question to ChatGPT, Gemini, Perplexity and Google's AI Overviews. Read the sources each one cites. You'll find four lists with almost no sites in common. It's not a feeling: a study of 161,286 prompts measured the overlap, and only 3.8% of sources appear on all four at once, while 72% appear on just one engine.

This breaks an idea still sold in plenty of places: that there's a single "GEO", one set of best practices that gets you into the AI. There isn't. There are engines with different source-selection machines, rewarding different things. This article shows the numbers, the machinery behind them, and what to do with it.

The proof: what citation divergence shows

The Writesonic study analysed citations across ChatGPT, Gemini, Perplexity and Google AI Overviews. Of 161,286 prompts with citations, 70,879 had citations on all four engines - the comparison set. The metric is the domain-level Jaccard index: how much of one engine's source list overlaps with another's.

Source overlap between engine pairs, from the most similar to the most distant. Source: Writesonic, study of 161,286 prompts.

Three readings that matter. First: no pair reaches 25%. Even the two most similar engines - Perplexity and AI Overviews, both heavily anchored in the Google ecosystem - share less than a quarter of their sources.

Second: ChatGPT is the outlier. It has the lowest overlap with every other engine, around 0.12 to 0.13. Anyone concentrating effort only on ChatGPT is optimising for the most isolated of the four machines.

Third: the single source rules. 72% of cited domains appear on exactly one engine; about 15% on two; and only 3.8% on all four. Being cited on one engine does not reliably predict being cited on another.

Honesty about the method, because it matters: Jaccard is symmetric and doesn't capture that some engines cite more sources than others; at the exact-URL level, rather than domain, overlap would be lower still. And it's a third-party study, not a law of physics. But the order of magnitude - high divergence, low overlap - is consistent with what you see tracking citations day to day.

Why they diverge: the retrieval machinery, engine by engine

Overlap is low because, before writing the answer, each engine does something different to fetch sources. Three axes explain almost everything: which index it starts from, which crawler built it, and what signal it weighs when choosing.

Google AI Overviews

It starts from Google's main search index and largely reuses the organic ranking that already exists. It uses a form of query fan-out - breaking the question into sub-searches - lighter than AI Mode. Result: it tends to cite pages that already had organic visibility for the question, with a bias toward featured-snippet-type passages. If you already rank well, you're halfway there.

Google AI Mode

Same index, different machine. The fan-out is more aggressive: the question is split into several sub-searches and the results fused into one answer. In November 2025, with Gemini 3, Google said the process runs even more searches and understands intent better. Consequence: AI Mode orders sources differently from the classic SERP - a site that never showed on page one can be cited because it answers a side sub-question well.

ChatGPT Search

It no longer depends only on third parties. Today it crawls with its own answer crawler - OAI-SearchBot, distinct from GPTBot, which is for training - and builds its own index, historically leaning on Bing. It's the most isolated of the four in terms of sources, which makes sense: its own index and ranker, less aligned with Google's organic.

Perplexity

Its own crawler - PerplexityBot to index, Perplexity-User for the user's real-time fetch - and its own index, also with a historical Bing root. It visibly weighs freshness and cites a lot. It's the engine with the highest overlap with AI Overviews (0.237), a sign that both value recognised-authority sources.

Gemini

It has no index of its own: it grounds in Google's index and cross-references the Knowledge Graph (entities). And it decides, prompt by prompt, whether a search is worth it. In the old API this was explicit: each question got a prediction score from 0 to 1 and, by default, only at 0.3 or above did it ground. In current models Google removed that control from the docs - the model decides internally. The machine has closed. For anyone optimising, the read is: a strong entity - sameAs, Knowledge Graph, consistency - weighs more here than on any other engine.

Copilot (Microsoft)

The case where Bing is still the engine, not a historical backstop. Bing's index plus IndexNow for fast indexing. If you're not in Bing Webmaster Tools, you're invisible here.

There's a detail almost no one gets right and it changes decisions: the difference between a training crawler and an answer crawler. Blocking GPTBot does not remove you from ChatGPT Search - that's OAI-SearchBot. Blocking Google-Extended does not remove you from AI Overviews - that's Googlebot. Block the wrong bot and you vanish by accident. The full bot matrix is in AI crawlers and bots.

What each engine rewards, in practice

Each engine's dominant signal. The signals collide: what makes you strong on one doesn't make you strong on another.

The conclusion isn't comfortable: the signals collide. What makes you strong on Gemini (entity) isn't what makes you strong on AI Mode (depth of coverage), and neither guarantees ChatGPT, which pulls from sources outside Google organic.

Consequences for strategy

There is no single GEO. Keep that line for when someone sells you "showing up in AI" as if it were a button. The right question is: on which engine, with what machinery.

Prioritise by where your audience is, not by the engine of the month. Measure first where your AI traffic comes from and concentrate effort there. The map of signals by channel is in the pillar on media channels as GEO intelligence.

Cover the three axes, in this order of effort: solid organic, which serves AI Overviews, AI Mode and, by extension, Perplexity; a consolidated entity, which serves Gemini and reinforces all of them; and presence outside Google - mentions and citable sources - which serves ChatGPT, the most isolated one.

Treat divergence as risk. Depending on a single engine for visibility is platform risk: one ranker change and you vanish from a huge slice of your citations. The why and the implications are in AI citations and platform risk.

Check on each engine, not by sample. Ask the same key questions across all four and record who cites you. It's the only way to know where you are - and where you aren't. How to do it, engine by engine, is in how to get cited on ChatGPT and Perplexity and in getting cited on Gemini.

Treat AI engines as a single block and you optimise for the average of four machines that barely agree. And the average cites no one.

Sources

• Writesonic - Do AI Engines Cite the Same Sources? A Study of 161,286 Prompts
• Marie Haynes - AI Mode query fan-out
• Google AI for Developers - Grounding with Google Search
• dejan.ai - How Google decides when to use Gemini grounding

Frequently asked questions