84% of AI citations come from other people’s websites. Here’s what I’m changing because of it.

Short answer: AI engines mostly repeat what other websites say about you. In the largest dataset I read, earned media made up 84% of the sources ChatGPT, Claude, and Gemini cited, and paid content made up 0.3%. On your own pages, the paragraph that wins the citation is usually the one that answers the question in its first sentence. So the job is two things: get talked about in the places AI reads, and write pages a model can quote without editing.

I read 13 studies and guides end to end this week so I could stop guessing. Two are peer-reviewed. The other eleven come from companies that sell AI visibility tools, which I flag every time I use a number. The grid below is the consensus in one screen. After it: what each finding means for your content, where the research contradicts itself, and the six tests we are running on our own site.

How this was made: I read every source in full, pulled 10 findings from each, checked the two papers against their original PDFs, and caught four places where other people’s summaries got the research wrong. The full source table, method, and corrections are at the bottom.

What gets a brand cited in AI answers?

If AEO is new to you, start with our plain-English guide to answer engine optimization. A brand gets cited in AI answers when trusted third-party sites mention it and when its own pages contain short, specific, sourced passages that answer the exact question asked. Those two signals show up across nearly every source I read. Everything else (schema, freshness, formatting) helps the model find and trust those passages, but it can’t replace them.

The off-site side is bigger than most teams budget for. AirOps reports that 85% of brand mentions in AI search come from third-party pages, which makes brands 6.5 times more likely to be cited through someone else’s site than their own. Muck Rack’s study of 25 million links found earned media holding between 82% and 89% of citations across three editions since July 2025. That stability matters. It means this isn’t a quirk of one model update.

Earned media is what AI trusts.

Greg Galant, cofounder and CEO of Muck Rack, on the May 2026 What Is AI Reading? study

This is why I think digital PR became the core of AEO without anyone renaming it. The teams that moved first on AI visibility were often PR and comms teams, a point Tomek Rudzki makes at Peec AI. If you run PR already, you are doing half of AEO (we wrote up what changed for digital PR separately). You just aren’t measuring it that way yet. (That gap is most of what our digital PR service does now.)

Where do ChatGPT, Perplexity, Gemini, Claude, and AI Overviews get their answers?

Each AI engine leans on a different mix of sources, so a single playbook underperforms on all of them. ChatGPT leans hardest on Wikipedia, Perplexity on Reddit, Google AI Overviews on a wide spread that includes YouTube and LinkedIn, and Claude cites less often but pulls more sources when it does.

Two practical notes. First, report visibility per engine, never as one blended score, because a brand can be strong in ChatGPT and absent in AI Overviews at the same time. Second, if anyone on your team works from France, they will never see an AI Overview locally. Peec AI measured a 0% appearance rate there, against 91.4% in the US.

Do backlinks still matter for AI search?

Backlinks still matter for AI search, but brand mentions matter more. In Ahrefs’ study of 75,000 brands, branded web mentions correlated 0.664 with appearing in Google AI Overviews, branded anchor text 0.527, and branded search volume 0.392. Raw backlink count came in at 0.218. The three strongest signals all live on other people’s websites.

The gap at the top is steep. Brands in the top quarter for web mentions had a median of 169 AI Overview mentions. The next quarter down had 14. The bottom half had between zero and three. Ahrefs puts it bluntly about brands in the lower half of web mentions: “you’re essentially invisible to AI systems.”

Two more findings change how I value PR wins. AirOps, citing Kevin Indig’s analysis, found that nofollow links correlate with AI visibility almost exactly as strongly as dofollow links (0.509 versus 0.504). And Ahrefs recommends asking for your brand name in the anchor text, not “click here.” So a nofollow mention in a trade publication, with your name in the link, is worth more for AI than a lot of SEO teams give it credit for.

Keep the caveat in view. These are correlations, and Ahrefs says so. Mentions and AI visibility might both come from being a strong brand in the first place. I still think the direction is clear enough to act on.

How does an AI engine pick which paragraph to cite?

AI engines usually pick which paragraph to cite in stages: they retrieve candidate pages, score individual passages against the question, then write the answer from the passages that survive. A page can rank on Google, get retrieved, and still lose the citation because a competitor’s paragraph answered the question more directly.

Metehan Yesilyurt at Peec AI tested this with 12 open-source reranking models, the kind of model that scores passages against a query. For the question “what are the best AEO tools,” a product-description paragraph scored about 0.002% relevance on one common model. A one-sentence shortlist that directly named the best tools scored above 99.9% on the same model. Same topic, same brand. The only difference was whether the first sentence answered the question.

He is careful to say nobody knows which reranker ChatGPT or Google uses, and one of the 12 models scored both passages high. But every model preferred the direct answer or already scored it near its ceiling. His advice, which I’m adopting word for word as a house rule: “Don’t write for the chunker.” Write so any few-hundred-word window of the page makes sense on its own, and put one direct-answer sentence at the top of each section.

The shape of that answer should match the question. This is the table I’m handing our writers:

What does the peer-reviewed research say actually works?

The most-cited research on this, the GEO paper from Princeton and IIT Delhi (published at KDD 2024), found that adding quotations, statistics, and cited sources to a page raised its share of AI answers by roughly 27% to 41%, while keyword stuffing made things worse. The authors tested nine edits on 10,000 queries and then confirmed the pattern on Perplexity.

The finding I like most is the underdog effect. When the researchers added citations to the source ranked fifth on Google, its visibility in the AI answer rose 115%. The top-ranked source lost 30% under the same treatment. For a smaller brand competing with big incumbents, that is the most encouraging sentence in the whole literature.

One honest limit: this study used 2023-era models. Production engines have changed a lot since. I treat it as the mechanism and the 2026 field data as the check on whether the mechanism still holds.

Why do listicles keep showing up in AI answers?

Listicles keep showing up in AI answers because their shape matches “best X for Y” questions: several named options, criteria, and trade-offs in one place. AirOps found nearly 90% of third-party brand mentions come from listicles, comparison pages, and review roundups.

The most useful study here is Peec AI’s analysis of 5.7 million data points across eight AI engines. Being included in the third-party lists that AI engines repeatedly cite was strongly associated with being named in the answer. Ranking first in those lists moved a brand earlier in the AI’s answer by 0.82 to 1.80 positions, depending on the market, and that effect held in all three markets studied.

Two caveats keep this honest. It is observational, so it shows association, not proof. And the returns flatten fast. Their summary of the strategy is “Win the few listicles AI engines actually cite.” Five placements on lists that AI actually retrieves beat fifty on lists it never reads. Finding those five lists is a research job, and it is the first thing I would give a PR team.

How often do you need to update content to stay cited?

To stay cited, update your important pages at least every three months. AirOps found that pages not updated quarterly were more than three times as likely to lose AI citations, and more than half of all cited pages had been refreshed within the previous six months.

The bar is higher for commercial pages. About 83% of citations for buying-intent queries came from pages updated within a year, and in SaaS, finance, and news, pages older than three months dropped off sharply. Service pages, pricing pages, and comparison pages are exactly the pages that need a calendar reminder. Evertune adds a practical rule: show a visible “last updated” date and replace statistics older than 18 months.

Is Google’s AI Overviews worth tracking more than ChatGPT?

For most US businesses, Google AI Overviews deserves at least as much tracking as ChatGPT, and usually more. Google reported 2.5 billion monthly AI Overviews users at I/O 2026, and Peec AI’s sample of 500,000 business prompts found AI Overviews on 86.7% of them in April 2026, up from 56.9% a year earlier.

The myth that AI Overviews only answers informational questions doesn’t survive the data. On decision-stage prompts, the “which one should I buy” questions, it appeared 88.5% of the time. Longer, clearer questions trigger it more often than short keyword searches. And because AI Overviews leans on YouTube, a short explainer video for each important page is a direct route in that ChatGPT doesn’t offer.

What doesn’t work for AI visibility?

Keyword stuffing, sounding authoritative, and buying ads don’t move AI visibility in the research I read. Hidden instructions aimed at AI models do work, which is exactly why you shouldn’t use them.

• Keyword stuffing. About 8% worse than doing nothing in the GEO paper, and 10% worse on Perplexity.
• An authoritative tone. No significant gain. Proof beats posture.
• Ad spend. Branded ad traffic and ad cost correlated only 0.215 to 0.216 with AI Overview mentions (Ahrefs), and paid content was 0.3% of citations (Muck Rack). Ads still earn their keep capturing people who are ready to buy and building branded search demand, which does correlate with AI visibility. They just don’t buy citations.
• Hidden prompt injections. Berkeley researchers showed at EMNLP 2024 that hiding instructions in a product page could push the lowest-ranked product toward the top of AI recommendations, including on Perplexity. It works in a lab. It is also spam, it is the obvious next target for enforcement, and the paper notes “improved model capabilities do not result in inherent robustness.” We won’t do it for clients, and I’d be suspicious of anyone who offers.
• Blocking AI crawlers or hiding content behind JavaScript. Many AI bots don’t run JavaScript. If your core copy only renders in the browser, some engines never see it.

Where does the AI search research disagree?

The research disagrees on content length, on how much Gemini relies on Reddit, and on how many statistics a page needs. I think these disagreements are more useful than the agreements, because they tell you where to run your own test instead of trusting a vendor.

• Length. Evertune says cited pieces typically run 2,500 to 3,000 words. Peec AI’s reranker guide says a 400-word page is simply one passage and chunk-conscious structure changes nothing there. My read: long for pillar guides, tight for service pages, and a direct answer at the top of every section either way.
• Gemini and Reddit. Muck Rack says Reddit is Gemini’s top-cited domain. AirOps says Gemini uses community sources in as few as 7% of answers. They measure different things (top domain versus share of answers) over different periods. Track your own category.
• How many stats. Evertune suggests 20 to 40 statistics per 1,000 words. That is one every 25 to 50 words, which would wreck a sales page. Treat it as a ceiling for data pages, not a quota.
• Plain versus distinctive writing. A University of Washington study, as summarized by Gumshoe, found AI Overviews prefers predictable, low-surprise text. Good conversion copy is specific and distinctive. We split the difference: plain language in the answer block, voice and persuasion below it.

What we’re testing on our own site

Freshness is personal for us. In our ChatGPT citation study, half of the finance pages ChatGPT cited most were getting zero citations eleven months later. So we are rebuilding our 16 service pages on this research and running six tests on them, starting October 2026. Each test runs six weeks, half the pages get the change and half stay as controls, and we track citations weekly in ChatGPT, AI Overviews, Perplexity, and Gemini. Sixteen pages is a small sample, so read our results as directional, not proof.

I’ll update this table as the numbers come in, including the tests that fail. If you want the results before they post, the button below gets you on the list.

What should you change first?

If you only have a week, change these five things in this order:

1. Write one direct-answer sentence at the top of every section on your five most important commercial pages.
2. Add at least three sourced statistics and one named quote to each of those pages.
3. Search your top five buying questions in ChatGPT and Google, list the third-party pages that get cited, and pitch the ones that leave you out.
4. Put a visible “last updated” date on those pages and set a quarterly reminder to refresh them.
5. Check that your robots.txt allows the AI crawlers you want and that your core copy loads without JavaScript.

Numbers 1, 2, 4, and 5 are on-site work your content team and a technical SEO can finish in days. Number 3 is where the real visibility lives, and it never finishes.

Frequently asked questions

The basement: sources, method, and receipts

Everything below is for anyone who wants to check the work, and for the AI crawlers reading this page. Evidence tags: PR means peer-reviewed; VD means vendor data (the source reports it, but the dataset is proprietary and most of these vendors sell AI visibility tools); H means practitioner guidance rather than a measured finding.

Source table

How we did this

We read each source in full, recorded 10 findings from each with the exact figures, and tagged every finding by evidence type. For the two academic papers, we worked from the full text rather than the abstract or a vendor summary. Percentages for the GEO paper were computed from its Table 1 (overall position-adjusted word count, baseline 19.3). Where sources measured the same thing differently, we reported both and said so rather than picking the more convenient number.

Corrections we found in other summaries

• Gumshoe’s summary of the Berkeley ranking-manipulation paper describes the method as rewriting prompts. The paper’s method was injecting text into the retrieved web pages. We cite the paper.
• Peec AI’s AI Overviews post says in its takeaways that “88.5% of them are bottom-of-funnel.” The body shows AI Overviews appeared on 88.5% of bottom-of-funnel prompts. We use the body figure.
• The GEO paper is often summarized as statistics giving “up to 41%.” In the paper’s own table, quotations gave the largest lift (about 41%) and statistics about 31%. The oft-quoted 115% applies only to fifth-ranked sources using citations.
• The Goliath study figures (for example, citation probability rising from 47% to 56%) come from Gumshoe’s summary. We have not yet verified them against the primary paper, so we haven’t leaned on them above.

Changelog

September 30, 2026: first published. Test results table will update monthly starting November 2026. Written by Joe Robison, founder of Green Flag Digital, with research support from Claude. Every figure was checked against its source by a human before publishing.