Blog · Content that gets cited

Does AI-written content get cited less by AI engines

Published 15 September 2026 · Updated 12 October 2026 · 8 min read

Does AI-written content get cited less by AI engines
Photograph: Alicia Christin Gerald / unsplash

The short answer: No, not for being machine-written. Google's published guidance says content is judged on quality however it was produced, and none of the engine documentation we checked describes an authorship check in the retrieval path. What gets cited less is content with no first-hand evidence in it, and that correlation is strong enough that people mistake it for a penalty on the tool. One retrieval study points the other way: in a controlled corpus, machine prose ranked higher.

That is an uncomfortable finding for anyone who has been told to sound human. It is worth understanding properly, because the real risk sits somewhere else entirely.

There is no detector in the retrieval path

Walk the pipeline an assistant runs. A prompt becomes one or more queries. Candidate documents are pulled from a search index, a cache, or a live fetch. Those documents are split into passages, embedded, scored against the query, re-ranked, and the few survivors become the answer. Citations are attached to some of them.

None of the published descriptions of that loop include a classifier for authorship. The ranking signals that are documented are lexical and semantic match, the freshness and authority inherited from the index, and whatever the re-ranker was trained to prefer. We cannot open a closed re-ranker, so we cannot prove an authorship feature is absent. What we can say is that it would be expensive to compute at query time and that the detectors which exist are wrong often enough to make it a poor signal.

How wrong? Work published in Cell Patterns in 2023 ran seven widely used detectors over 91 TOEFL essays written by non-native English speakers and 88 essays by US eighth-graders. The non-native essays were flagged as machine output far more often, while the native-writer essays were mostly classified correctly. That is one paper on two small essay sets, not a survey of every detector on the market, but a bias that large in a published test is enough to make the class of tool a poor ranking input. The authors also found that prompting the text into more literary language removed much of the signal.

The retrieval literature points the other way

Researchers at KDD 2024 tested what happens when human-written and LLM-rewritten versions of the same documents sit in one corpus. Neural retrievers ranked the machine versions higher, an effect the authors named source bias, and it persisted through the re-ranking stage. Their explanation is mechanical rather than mystical. Generated prose carries more focused semantics and less noise, so the embedding sits closer to a clean query vector.

Read that carefully, because it is easy to over-claim from it. The study is about retrievers over a controlled corpus, not about ChatGPT deciding which brand to name. It does not say machine text is better. It says the shape of machine text happens to suit dense retrieval, which is the same reason answer-first writing works. The lesson is about structure, and structure is copyable by a human writer in ten minutes.

So what actually suppresses the citation

Three things, none of which is a tool choice.

Nothing new in the document. An engine assembling an answer has five or six candidates saying roughly the same thing. It keeps the one that adds a number, a date, a constraint or a named condition. A post generated from the top ten results for its own query contains, by construction, zero information that was not already in those ten. It is a summary of a summary, and it loses every tie. What makes a page citable by an AI engine is mostly this point in longer form.

Duplication at scale. Google's spam policies name the behaviour directly under scaled content abuse: using generative tools to produce many pages without adding value for users. That documentation was last updated on 28 August 2026. The policy is about the outcome, not the tool, and the same sentence catches a human content farm. Worth remembering that a classic search penalty also removes you from the index that several assistants retrieve from, so one enforcement action costs you twice.

Claims with nothing behind them. A generated draft states things confidently with no source attached, because that is what the training objective rewards. The nearest public evidence that this costs you is the GEO benchmark work below, where adding citations and statistics to a source page raised its visibility in generated answers. A sentence with a figure, a year and a link beside it is easy to lift intact. A sentence asserting that something is important is not. We cannot see a re-ranker's preferences directly, so treat this as the best available inference rather than a documented rule.

The format tax, which is the part people miss

Citation is not awarded to a page. It is awarded to a passage that was clean enough to lift. Generated drafts fail that test in a specific, repeatable way: they front-load throat clearing, define terms the reader already knows, and bury the answer under four paragraphs of context. The chunk that gets retrieved then contains preamble rather than substance.

The KDD 2024 work on generative engine optimisation measured the gain from fixing exactly this. Adding citations, statistics and quotations to source pages lifted visibility in generated answers by up to 40 percent, measured on GEO-bench, a query benchmark the authors built themselves, with the gain varying a lot by domain. That is a benchmark result from 2024, not a figure from a live engine today. Nobody rewrote the author. They rewrote the evidence density.

This is also why comparison pages that get cited tend to beat generated listicles covering the same ground. A comparison page states a position, attaches a condition, and can be quoted in one sentence. A generated listicle hedges every row, so no single row survives on its own.

A check you can run this week

Take ten published pages. For each one, do four things, and keep the result in a sheet.

  1. Count the sentences containing a number, a date or a proper noun that is not your own

    2. brand. Fewer than one in six, and there is nothing in the page to extract.

  2. Copy the first 60 words into a blank document. Do they answer the title, or do they

    3. introduce the topic? Introductions do not get cited.

  3. Search a distinctive sentence in quotes. If near-identical phrasing appears on three

    4. other domains, you have a duplication problem rather than an authorship problem.

  4. Ask two engines the question the page targets, three times each, on different days.

    5. Record who was named and which sources carried it. One run tells you nothing, because answers move between runs.

Budget about an hour for ten pages. Our audit runs the same logic across a prompt set of roughly 45 prompts and five engines with retrieval on and off, and the sample audit report shows what the evidence looks like when it is written down. The manual version answers the question well enough to act on.

Where the authorship question does bite

Two places, both indirect.

Platforms have their own rules, and some are stricter than the engines. Several community spaces remove accounts that post generated text at volume, under the platform's rules rather than an engine's, and that matters because those spaces are heavily cited. If you are working on why Reddit gets cited so much, a generated comment is a removal risk long before it is a ranking question. The same applies to professional networks, where LinkedIn as a cited source depends on a post staying up and attracting real replies.

Then there is your own credibility with a human reader who arrives from the answer. Being cited and then losing the visitor in the first screen is arguably a worse outcome than not being cited, and no retrieval metric we know of will tell you it happened.

Honest limits

This post does not establish that engines never check authorship. It establishes that we found no published documentation of such a check and no detector accurate enough to serve as one, which is weaker. The two KDD 2024 results are benchmark findings on controlled corpora, so we do not know whether source bias or the 40 percent visibility gain survives in a live engine with authority and spam signals layered on top. The Cell Patterns detector result rests on two small essay sets from 2023 and detectors have changed since. Everything we say about what re-rankers prefer is inference from what moves when we change a page, not observation of the ranker itself.

FAQ

Do AI engines detect and demote AI-written content? We have found no published evidence of an authorship classifier in the retrieval path, and detector accuracy in the research we can open is too poor to use as a signal. What correlates with low citation is thin, unsourced, duplicated content, which generated drafts produce by default.

Will Google penalise me for publishing AI-written pages? Only under the scaled content abuse policy, which is about publishing many pages with no value added for users. The policy text is explicit that it applies no matter how the content was created.

Should I disclose that a page was AI-assisted? It is a trust decision rather than a ranking one. We have not found a disclosure requirement in the engine documentation we checked, and we know of no measured citation effect that depends on it.

Does editing a generated draft fix the problem? Only if the edit adds something. Changing the wording leaves the information content identical. Adding a measurement, a date, a named limitation or a real example changes what the passage is worth quoting for.

Sources

Related reading

What makes a page citable by an AI engineComparison pages that get citedWhy Reddit gets cited so much, and what to do about itLinkedIn as a cited sourceAnswer-first writing, and why the opening paragraph carries the answerWriting case studies an engine can quote