Do AI engines prefer citing content written by AI?

99Visibility Research · Updated October 21, 2026 · 5 min read

Short answer

There's no solid evidence that AI engines prefer content written by AI. A 2025 study found generative engines cite text that their underlying models find more predictable. Its causal test, though, had an LLM rewrite whole pages, so style was never isolated. The result is tied to specific models and predates current ones. Google says you don't need to write in a special way for its AI features.

Key takeaways

The one study behind this claim (arXiv:2509.14436, September 2025) found a citation preference for model-predictable text, but its causal arm changed many things at once.
Predictability is measured against one specific model, so the finding doesn't transfer cleanly to the models engines run as of October 2026.
Google states you don't need to write in a specific way for generative AI search, and it says appropriate use of AI is not against its guidelines.
Controlled tests of citation-oriented rewrites found them largely ineffective or harmful, so rewriting pages to sound like a model is not a supported tactic.
The levers with evidence are access fixes and clearly stated facts on fetchable pages, not prose style.

The short answer and the reasoning

No published evidence shows that AI engines reliably prefer citing AI-written text. The idea comes mostly from one paper, and that paper can't carry the claim.

The 2025 study "When Content is Goliath and Algorithm is David" (arXiv:2509.14436) reported that generative engines "exhibit preferences for citing content characterized by significantly higher predictability for underlying LLMs." Predictability here means low perplexity: text the model could have guessed word by word.

Three problems stop this from becoming advice:

The causal test rewrote whole pages. It changed register, structure, hedging, terminology and length together. The effect belongs to "an LLM rewrote the page", not to predictability alone.
The observational data is confounded. Pages that read smoothly also tend to come from well-funded publishers and to sit in training data.
Perplexity is model-specific. A score under one model says little about a different model, and the paper predates the 2026 frontier releases.

So the honest reading is: there's a correlation in one study, at the citation stage, for models that are no longer current. Our digest of the predictability study walks through the design in more detail.

Where in the pipeline this finding sits

AI answers pass through four stages. The model's frozen knowledge (parametric). Whether a crawler can fetch your page (access). Whether your page enters the context for a query (retrieval). Then which of those pages the model uses and cites (generation).

The predictability result lives in the last stage. It describes a preference among pages already retrieved. It says nothing about whether a page gets fetched or retrieved in the first place.

That matters because the generation stage is where most GEO evidence lives. A July 2026 critical survey of 45 GEO studies (arXiv:2607.14035) concluded that already-retrieved content can change its citation, but no technique shows a stable, cross-platform causal effect on discoverability. The same survey warns that "citation-oriented rewrites can impair retrieval."

The paper itself adds a useful contrast: its authors found perplexity had "minimal influence on ranking within conventional Google search results."

What controlled tests say about rewriting

If rewriting prose for machines worked, controlled benchmarks should show it. They mostly don't.

C-SEO Bench (arXiv:2506.11097, NeurIPS 2025) tested conversational SEO rewriting methods across several models. Its authors found that "most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking." A separate 252,000-trial factorial (arXiv:2605.25517) found that formatting-only edits have little impact, and that topical relevance and list position drive which source gets cited first.

Both studies supplied documents directly into the model's context. Neither queried a live index. That makes them generous tests, and the rewrites still failed. Our comparison of on-page rewrites versus access fixes lays out the evidence side by side.

What Google says about writing for AI

Google's guide to optimizing for its generative AI features is direct: "You don't need to write in a specific way just for generative AI search." It also says there's no ideal page length and no need to chase every long-tail variation.

On who or what writes the page, Google's position is that "appropriate use of AI or automation is not against our guidelines." What it acts against is scaled content abuse: many pages produced to manipulate rankings without adding value. Our summary of Google's guidance on AI-generated content covers the full wording.

These are statements from Google, not independent measurements. They are still the best first-party evidence for Google's surfaces.

Edge cases: volume, and the misread version

Two edge cases trip people up.

First, "AI likes AI text" can slide into "publish lots of AI text." OtterlyAI tested that in 2026 with two new domains of roughly 1,000 AI-generated posts each. Google deindexed both, and Otterly reports no manual action on either. Google's AI surfaces then almost stopped citing them, because a page must be indexed and snippet-eligible to appear there. Microsoft Copilot, which grounds on Bing, kept citing them. It's one pair of sites on a low-stakes topic, so treat it as a warning, not a rule. Our guide to where AI-generated SEO at scale breaks covers the pattern.

Second, some tools now report a "perplexity score" for your page. That score comes from an open-weight model, not the model doing the citing. Treat it as an uncalibrated proxy, not a visibility factor.

What to do instead

Write for your buyer in your normal voice, with or without AI help. Then spend effort where the evidence is strongest.

Check access first. Make sure your robots.txt allows the retrieval bots (OAI-SearchBot, PerplexityBot, Claude-SearchBot), and that your firewall or CDN doesn't return 403 to them.
Put key content in the HTML. Most AI crawlers don't execute JavaScript, so text that only appears after scripts run may be invisible to them.
State specific facts plainly. A price, a number or a one-line definition on a fetchable page gives an engine a sentence it can quote. See whether publishing original data helps and whether engines cite whole pages or sentences.
Measure per engine. Engines cite differently, so one blended number hides a lot. The free checker on this site shows whether engines name you for up to three buyer questions, as a snapshot.

None of this guarantees a citation. It removes the blocks you control.

Frequently asked questions

Should I run my existing pages through an LLM to make them more predictable?

The evidence doesn't support it. The study behind the idea couldn't separate predictability from everything else a full LLM rewrite changes. Controlled benchmarks found citation-oriented rewrites largely ineffective and sometimes harmful, and a 2026 survey warns they can impair retrieval. You'd risk the page's accuracy and voice for an effect nobody has isolated.

Does Google penalize content because AI wrote it?

Not for that reason alone. Google says appropriate use of AI or automation is not against its guidelines. Its spam policies target scaled content abuse, meaning many pages generated to manipulate rankings without helping users. The method of creation isn't the test. Value to the reader and the pattern of publishing are.

What is perplexity in the context of AI content?

Perplexity measures how surprised a language model is by a piece of text. Low perplexity means the model could predict the next words easily. It's always relative to one model, so the same page scores differently under different models. That's why a score from an open model can't tell you what a closed engine prefers.

Do AI detectors or perplexity scores predict AI citations?

No published evidence shows that. A perplexity score you compute uses a model other than the one choosing citations, so it's an uncalibrated proxy. The 2025 study's own authors found perplexity had minimal influence on conventional Google rankings. Treat any such score as an observation, not a visibility metric.

Is writing style irrelevant then?

Style isn't a proven lever for AI citations, but clarity still matters for readers. The strongest on-page signal in the evidence is stating specific facts, like prices or definitions, on a page crawlers can fetch. Topical relevance to the query also shows up consistently in controlled studies. Neither is the same as sounding like a model.

Sources

This page is AI-assisted. It was written from a verified evidence base and checked against our editorial policy. Engine behaviour changes often; facts are as of October 21, 2026.