AI citation patterns by engine: what ChatGPT, Claude, Gemini and Perplexity each lean on

We ran the check across 507 buyer questions and compared what ChatGPT, Claude, Gemini and Perplexity cited. Perplexity cited the most sources per answer, ChatGPT cited the fewest, and Claude named the tracked brand most often by a small margin. Here, the tracked brand means the company being audited in that scan, across 44 different sites in our sample, rather than one product measured across every category. The sample contains 64,667 citations across 9,791 AI answers, from 77 scans in 48 categories. You can read the weekly research on who owns AI answers and how CiteGraph measures before treating the result as a plan.

Engines cite at different rates

Perplexity averaged 5.5 sources per answer across 2,440 sampled answers and named the tracked brand in 10.5% of them. Claude averaged 4.9 sources across 2,120 answers and named the tracked brand in 10.7%. Gemini averaged 4.2 sources across 2,437 answers and named it in 9.4%, while ChatGPT averaged 3.5 sources across 2,434 answers and named it in 6.1%. A source is a page the engine cites in its answer. A source count is the average number of cited sources in one answer, while a named rate is the share of answers that named the tracked brand.

Those measures describe different jobs. More sources can mean a broader research trail, but it does not automatically mean more product recommendations. A high named rate can mean the tracked brand appears in the answer, while the cited pages may belong to other companies or publishers. Treating the source count and named rate as one score is a tidy way to lose the useful detail.

The gap matters when you plan content. A page intended to support Perplexity has more chances to enter a multi-source answer. A page intended to support ChatGPT has fewer citation slots in this sample, so it needs to answer the buyer's question with less wasted motion. These are useful working assumptions from this sample, rather than settings built into the engines.

Perplexity leaves room for evidence

Perplexity's higher source count suggests a useful publishing target: create pages that can supply a specific piece of evidence in a wider answer, rather than relying on one broad company page to carry every claim. The sample does not promise that Perplexity will cite your page. It does suggest making individual claims easy to find, check and reuse.

ChatGPT showed fewer sources per answer in the same sample. That makes concise, complete pages more important when you are trying to support its answers. The difference concerns the shape of the answer, not a separate content format for each engine.

Start with the questions buyers already ask. The scans covered 507 distinct buyer questions, so a useful content plan can begin with the decisions those questions represent. Make each page answer one decision plainly, state who your product suits, explain where it falls short, and give the reader enough detail to compare it with alternatives.

The page should also work when quoted in a sentence. Put the product category, audience, key capability and meaningful limitation in ordinary text. A page full of slogans gives an engine very little material to reuse. A page with clear statements gives it something closer to an answer.

Claude leads on product naming

Claude had the highest named rate in the sample, with Perplexity close behind. Gemini and ChatGPT followed. The result makes Claude an engine worth watching when the immediate aim is product recognition, while the source figures show why recognition and citation volume should stay separate.

It's worth remembering that naming and citing are logically separate outcomes, even though our current sample does not break out how often each happens alone. A tracked brand can appear in an answer without its own site being cited, or its site can be cited without the answer recommending it. Track both outcomes, then inspect the pages attached to each result.

The difference between Claude and ChatGPT is large enough to affect reporting. If you monitor only ChatGPT, you could conclude that a product is rarely named when the broader sample shows a higher named rate elsewhere. If you monitor only naming, you could miss the source pages that shape the recommendation. Engine citation behaviour is visible in the answer and in the page behind the link.

The practical publishing response is to make your product legible in two places. Your own page should state what the product does and who it is for. Other pages should be able to describe it accurately in a comparison or roundup, because those page types dominate the citation record.

The page mix is lopsided

Listicles and roundups accounted for 68.3% of recorded citations, or 44,164 citations. Competitor product pages accounted for 24.1%, or 15,575 citations. Together, those two page types made up 92.4% of citations. The tracked product's own site accounted for 1.7%, or 1,107 citations.

That distribution changes where you publish. Your product page still needs to explain the offer, but it is not the main citation surface in this sample. A comparison page, roundup or competitor page is more likely to sit in the path between a buyer question and an AI answer. This is awkward for the familiar content plan, which often gives the homepage and product page all the attention and treats the surrounding web as someone else's problem.

Here is a hypothetical before-and-after, not a page from our scans, to show the shift in practice. Before, a product page might open with the sentence, “Built for teams who want to move fast,” followed by a feature list. After, it could open with, “A workflow tool for small marketing teams that organises campaign tasks but does not replace project management,” then place comparison facts beside the relevant features. The illustrative change tells the reader what the product is, who it suits, what it does and where its boundary lies.

That change helps your own site, but the larger opportunity is often outside it. Build a short list of publishers, communities and comparison sites that already cover the buyer question. Give them precise facts they can check. Do not ask them to repeat your positioning word for word. Nobody needs another page describing a tool as powerful, flexible and polished.

Our wider sample included community threads at 3%, social posts and documentation at 0.8% each, review platforms at 0.8%, YouTube at 0.4%, news at 0.1%, and Wikipedia at 0%. Those shares are small beside listicles and competitor pages. They still show that source choice is not limited to your domain, which is why a publishing plan needs more than product page improvements.

Sample limits matter here

The obvious objection is that page type may only reflect the questions we asked. If the scans contained more comparison prompts, comparison pages would naturally appear more often, so the same percentages may not describe every category or engine. The objection is fair. It does not, however, make the pattern useless: across the recorded citations, listicles and roundups took the largest share by a wide margin, followed by competitor product pages. That is enough evidence to test those surfaces first, provided you keep the conclusion narrow and treat these as the pages our scans found, rather than a guarantee about every future answer.

There are other limits. The sample covers 44 sites and 48 categories, not every market. The engines cited 9,192 distinct domains and 21,465 distinct pages, so the source set is broad, but breadth does not remove sampling bias. We measure samples, print their margins and do not know what we have not measured.

The named rates also need care. A product name can appear because the answer is comparing products, answering a branded question or drawing on a source that mentions it. The rate does not tell you why the engine chose the product, whether the reader trusted the recommendation, or whether the product gained a customer. It tells you what appeared in the sampled answers.

Engine-level reporting is therefore more useful than one blended number. Compare source count, named rate, cited page type and the exact page behind the citation. The leaderboards can help you keep that view separate by engine and category.

Evidence travels beyond your domain

The findings point to different jobs for different engines. For Perplexity, produce clear evidence pages that can contribute one useful source to a fuller answer. For Claude, make the product easy to identify and describe, while checking whether the supporting citation comes from your site or from an outside page. For Gemini, build around readable product facts and credible surrounding coverage. For ChatGPT, make precise buyer answers and strong third party pages especially important.

Do not turn those observations into separate content machines. The same page can support several engines when it is specific, readable and honest about fit. The difference is where you look for the next opportunity. A Perplexity report may send you towards source breadth, a Claude report towards product recognition, and a ChatGPT report towards the small set of pages it actually cited.

Over time, group your results by page type as well as domain, so repeated buyer questions show whether the missing evidence sits on your site or elsewhere. This week, do three things. First, pull your engine results apart and compare source counts with named rates, rather than reporting one visibility number. Second, choose one buyer question and rewrite the relevant page so the category, audience, use case, comparison point and limitation are plain near the start. Third, find the external listicles, roundups and competitor pages appearing for that question, then decide which accurate fact or comparison you can make easier for them to verify.