ChatGPT SEO and the primary sources that keep getting cited

What the study measured, and what it did not#

Ahrefs published its study on 12 October 2026. The headline finding is a rank move: “PubMed climbed from 64th to 4th among ChatGPT’s most-cited domains”. The study compares two samples, one in March and one in September, and the authors say the sample covers “across 2.4 million answers in March and 3.8 million in September”.

The method is worth stating plainly. According to the study, it analysed logged-out, free-tier ChatGPT answers and the fan-out queries behind them, then used separate citation data from Ahrefs’ Brand Radar tool to track which domains were cited. That makes it a study of one tier, observed through one vendor’s data, over two months.

It is not peer-reviewed, and it is not a census of every answer the assistant gives. Anyone quoting it should say so, and should say that the sample describes the free tier.

The study also ranks the “200 most-cited domains” rather than the whole web of citations, so a move in its table is a move among leaders. Ahrefs’ own monthly tracker tells a different story for the same month. Its September update, which uses a broad US sample and is labelled “Data: broad US queries, September 2026”, places PubMed 17th with a mention share of 1.8%. Two Ahrefs datasets for September give two different PubMed ranks, so the sample decides the answer. Ahrefs’ most-cited domains tracker is the document to read beside the study.

Why original sources matter to a chatgpt seo plan#

The study reads the rank move as a preference. In its own words, ChatGPT appears to put original sources ahead of sites that summarise them. We paraphrase that finding here rather than quote it, because quoting the study’s own wording would print an American spelling in our copy. For chatgpt seo, the lesson is about origin rather than format, and the scent of a primary source is the strongest signal in the data.

Put plainly, a page that repeats a fact from a research paper is competing with the paper itself. When an assistant chooses which source to cite, the paper has the advantage of being the place the fact came from. For an agency, that changes the job from writing a better summary to becoming the source worth summarising.

The social reaction to the study followed the same line. Search Engine Journal’s headline made the point directly, and the reaction on X picked up the same argument about primary sources.

This matters for geo optimization work because it points at the content that earns citations, not at the tricks that pretend to. A primary source has a named author, a dated method and a number a reader can check. An explainer has none of those, and an assistant has less reason to cite it when a primary source is available.

What an agency should stop selling as aeo seo#

Some of what is sold under the heading of answer engine work is ordinary search work with a new label. The study does not support a claim that a file, a markup tag or a special format will move a page into an assistant’s citations, and it does not show that any such tool works. A chatgpt seo deliverable should be judged the same way, and the undergrowth of format tricks is easy to clear.

The sensible position is the one Google describes for its own generative features. Its documentation on AI features in Search, last updated on 10 December 2025, says: “You don't need to create new machine readable files, AI text files, or markup to appear in these features.” It adds that there is “no special schema.org structured data that you need to add”. An agency that sells a separate AI file as the key to visibility is selling something the public documentation does not require.

Google’s guidance on generative AI content, updated on 1 October 2026, says generative AI “can be particularly useful when researching a topic, and to add structure to original content”. It also warns that mass-produced pages can break its rules, and its spam policies define scaled content abuse as “when many pages are generated for the primary purpose of manipulating search rankings”. Volume with no added value is the thing to sell against.

Our own search and generative engine work starts from the same place: a page a machine and a reader can both verify. We would rather a client published one fact with a named source than ten summaries with the same fact and no source.

Making a page worth citing, step by step#

The working list below is ours. It follows from the study’s finding that original sources win, and it does not claim the study tested any of these steps. Each step is a chatgpt seo habit rather than a format trick, and the trail from fact to citation is short.

Each step is small. Together they change what the page is for, which is the change the study points toward.

What to measure over the next thirty days#

Measurement is where most GEO plans go wrong, because the easy numbers are not the right ones. A rank in one assistant on one day is noise. A pattern across the sources an assistant cites, across several weeks, is closer to evidence. Measure chatgpt seo as a citation pattern, not a single rank, because the trail of citations is what changes.

Track the chatgpt cited sources behind a client’s priority questions, and classify each one as primary, reference or summary. Then track whether the client’s own pages move from summary to primary over the period. The aim is not a vanity number. It is a clear answer to whether the content is becoming the source.

Keep the study’s own limit in view when you read the results. The data is a free-tier sample from one vendor, so a result on one platform does not prove a result on another.

Where the chatgpt rankings argument is still thin#

The study is useful, and it is also one study. It does not test whether primary sources win citations on paid tiers, in other languages or for queries about the same topic asked in different ways. Any chatgpt seo result should be read with the same caution, and the honest prowl for evidence continues.

Academic work points the same way on sources, and in a different direction on structure. A 2025 preprint finds that news citations “concentrate heavily among a small number of outlets” (arXiv:2507.05301). A 2026 preprint reports “evidence of AI-generated sources being cited across all four generative search engines” (arXiv:2605.23684). A third study finds that “pillars related to Metadata and Freshness, Semantic HTML, and Structured Data showed the strongest associations” with citation (arXiv:2509.10762). These are associations, not proof of cause.

Context matters too. A year earlier, Ahrefs found that 67% of ChatGPT’s Top 1,000 Citations Are Off-Limits to Marketers, in a study of the pages ChatGPT cited in September 2025. Different years, samples and methods give different answers, which is the strongest argument for reading each study with its method beside it.

It does not show that an article which cites a primary source will rank higher in search, and it does not show that a primary-source strategy beats a strong summary that is well linked. Those are reasonable hypotheses. They are not findings.

Anyone who reports this result to a client should say which claim the study supports, which it does not, and when the next measurement will come. That discipline is worth more than the headline.

Frequently asked questions#