What +400K AI citations reveal about how LLMs cite (based on real data)

Research: AI citations reveal about how ChatGPT, Gemini, AI Mode, Perplexity and Grok choose their sources

How LLMs cite? We pulled every source the five major AI engines cited across our tracked prompt set for 90 days. The five engines produce roughly the same volume of citations — but they draw them from pools that differ by a factor of forty, and they reward completely different formats. Optimising for “AI search” as one channel is the fastest way to waste a content budget.

Dataset: 403,588 citation events · 72,581 unique URLs · 15,558 domains Window: 12 June – 8 September 2026 Engines: Google AI Mode, ChatGPT, Gemini, Perplexity, Grok Source: Opttab citation tracking, AI visibility & GEO software category

The engines agree on volume and disagree on everything else

Split by citation volume, the four leading engines look almost interchangeable. AI Mode takes 27.8% of all citation events, ChatGPT 21.8%, Perplexity 21.6%, Gemini 21.6%. Grok trails at 7.0%.

Then you count how many distinct pages each engine actually pulled from, and the picture falls apart. AI Mode cited 30,628 unique URLs. Perplexity produced almost the same citation volume from 7,746. Grok produced 47,080 citation events from 730 URLs.

A page cited by Grok is cited, on average, 64.5 times. The same page in AI Mode gets 6.1 citations. Grok is not a bigger opportunity — it is a much narrower door.

The practical consequence is that the number of pages you need to win differs by an order of magnitude per engine. Half of everything Grok cited came from just 25 domains and 52 URLs. Half of everything AI Mode cited came from 84 domains and 766 URLs.

Grok52 URLs Perplexity413 URLs Gemini532 URLs ChatGPT704 URLs AI Mode766 URLs

Number of individual URLs that account for 50% of each engine’s citations.

What to do with this

• If you are chasing Grok or Perplexity, stop building pages and start targeting specific existing URLs — the ones already inside that narrow set. Getting added to one of them is worth more than ten new posts.

• If you are chasing AI Mode, the opposite holds. Breadth wins: it will happily cite the 10,000th domain in the tail.

• Report AI visibility per engine. A single blended “AI share of voice” number hides the fact that the same page performs 10x differently across engines.

Each engine has a different source diet

Here is the share of every engine’s citations by the type of source behind it.

AI Mode ChatGPT Gemini Perplexity Grok Vendor & operator sites Trade / industry media Video Social & UGC Research, news, other

Three splits in that chart are worth the whole study.

AI Mode cites video 178 times more than ChatGPT

Video is 5.56% of AI Mode’s citations and 0.03% of ChatGPT’s. Perplexity sits in between at 1.28%; Gemini is at 0.19%, Grok 0.53%. Effectively all of it is YouTube.

The gap is widening. Comparing the two complete months in the window, AI Mode’s newly discovered citations dropped 41% overall — but its video citations rose. Video went from 4.0% of AI Mode’s new citations in July to 8.4% in August.

Within that video pool, the format split is decisive: long-form /watch pages took 10,018 citation events against 426 for Shorts. AI Mode is citing full videos with substantial transcripts, not clips.

To become visible in Google AI Mode

• Publish long-form YouTube video on your core commercial topics — the same topics as your money pages, not brand content. Shorts are close to worthless here.

• Treat the transcript as the ranking asset. Say the definition, the number and the product name out loud in the video; AI Mode extracts from spoken content, so a claim that only appears in your on-screen graphics does not exist.

• Write a full description with the same claims in text, plus timestamped chapters, so the passage is retrievable both ways.

• Do not port this to ChatGPT. A video strategy aimed at ChatGPT is a 0.03% strategy.

ChatGPT is the only engine that reads homepages and research papers

Homepages are 11.9% of ChatGPT’s citations and 0.8% of AI Mode’s — a 14x gap. Add pricing pages, product pages and docs, and roughly 21% of ChatGPT’s citations point at brand-owned commercial pages. Every other engine sits between 3% and 5%.

ChatGPT is also the only engine that reads academic work at any scale. arXiv alone is 4.4% of its citations, making it ChatGPT’s single most-cited domain. Gemini’s entire academic share is 0.11% — ChatGPT cites research 44x more.

To become visible in ChatGPT

• Fix your homepage first. It is the one engine that will quote your own front page back to a buyer, so the page has to state plainly what the product is, who it is for, what it costs and how it differs — in text, above any interactive element.

• Keep pricing public and in HTML. “Contact us for pricing” removes you from the comparison ChatGPT is assembling.

• Publish docs. Technical documentation earns ChatGPT citations at 3–4x the rate it earns them anywhere else.

• If you have any research, data or benchmark work, put it on arXiv or in a citable paper. No other engine rewards this — ChatGPT rewards it heavily.

News media barely registers anywhere

Mainstream news is 0.32% of all citations in the dataset. Even in ChatGPT, which uses it most, it is 0.99%. Trade and industry media — the specialist publications in the category — is 9.4%, roughly 29 times larger.

This inverts the standard PR playbook. A placement in a national title generates almost no AI visibility. A placement in the five specialist publications your category actually reads generates most of it. Grok is the extreme case: 14.3% of its citations come from trade media, the highest of any engine.

Where to spend earned-media budget

• Rank your category’s trade publications by how often each is already cited, and pitch those. Ignore tier-one national press if AI visibility is the goal.

• Reddit and LinkedIn are small but not zero, and they split by engine: LinkedIn is 2.1% of AI Mode and 1.6% of Grok, but 0.00% of Gemini. Reddit is strongest in ChatGPT at 2.0%.

• Don’t buy generic guest posts. Buy inclusion in pages that are already being cited — see the next section.

Listicles are the one format every engine agrees on

“Best X” and comparison pages are 25–39% of citations on every single engine. It is the only format with no meaningful disagreement between them, and on Grok it is the single largest category at 38.9%.

They also outperform per page. A listicle earns 7.5 citations on average against 5.2 for a standard blog post — and the effect compounds on the narrow engines, where a listicle earns 86.3 citations on Grok and 27.0 on Perplexity.

Wording matters more than it should. URLs containing “best” average 8.4 citations; “top” averages 5.6, “alternatives” 5.6 and “vs” 5.2. If you are commissioning or writing one, “best” is the pattern the engines are retrieving against.

The compounding asset. Only 1.7% of URLs in the dataset were cited by three or more engines — but those 1,244 pages carry 22.6% of all citations. They average 73.2 citations each against 4.0 for a single-engine URL, an 18x difference. And they are disproportionately listicles: 31.1% of cross-engine pages are comparison content versus 19.2% of single-engine pages.

That is the highest-leverage target in AI visibility. One placement in a page that all five engines already trust is worth more than eighteen placements in pages only one engine reads.

To get cited across every engine at once

• Build a list of the pages already cited by three or more engines for your category’s prompts. That list is short — roughly 1,700 pages per 100,000 tracked citations — and it is your outreach target list.

• Approach those publishers to be added to the existing list. Being added to a cited listicle is faster and cheaper than ranking a new one.

• Publish your own comparison content with real pricing, honest trade-offs and named competitors. Vague roundups without specifics do not get retrieved.

• Keep the year current. URLs containing “2026” average 6.2 citations, “2025” 3.9, “2024” 2.4. Stale-dated URLs decay fast.

Gemini quotes passages, not pages

90.6% of Gemini’s citations arrive as a link to a specific highlighted passage rather than a page. No other engine does this at scale — AI Mode does it 2% of the time, ChatGPT and Perplexity effectively never.

We extracted every one of those 19,292 quoted passages. The median is nine words, with the middle half falling between 8 and 13. One in five is a plain definition (“X is…”, “X refers to…”). Three in ten are comma-separated lists of features or attributes. Only 14.7% contain a number.

Gemini is also the most domain-concentrated engine after Grok: 21,288 URLs from just 1,776 domains. It reads few sites, but reads them exhaustively.

To become visible in Gemini

• Write in extractable units. Every important claim should survive being lifted as an 8–13 word sentence fragment with no surrounding context.

• Lead sections with a direct definition of the thing being discussed, in one clause, before you elaborate.

• Use plain comma-separated lists of capabilities in prose, not only in bullet graphics or tables — that comma-series pattern is what Gemini lifts nearly a third of the time.

• Depth beats breadth on one domain. Gemini rewards sites that cover a topic exhaustively rather than sites that publish once about many topics.

Citation sets churn faster than anyone plans for

For AI Mode and Gemini, the median cited URL has a lifespan of zero days: it appears once and is never cited again. ChatGPT’s median is two days. Perplexity’s is fourteen. Grok’s is forty-six.

Read the other way: 91.7% of Grok’s citation volume comes from pages still being cited in the most recent three weeks, and 69.8% of Perplexity’s — but only 56–61% for AI Mode, ChatGPT and Gemini. The Google engines constantly recycle their source pool. Grok has effectively frozen its own.

“New URLs” compares July and August, the only two complete months in the window.

ChatGPT was the only engine expanding its source pool over the summer, adding 44% more previously unseen URLs in August than in July, while AI Mode and Gemini each cut theirs by roughly half. If your AI visibility fell in August and you only track Google surfaces, the cause may not be your content.

Why this matters for measurement. With a median lifespan of zero days on two of the five engines, a monthly visibility check is not sampling the same population twice. Anything measured less often than weekly is measuring noise as much as movement.

The short version

Underneath the per-engine tactics, three things are true across all five: comparison content is the universal format, specialist trade media beats national press by roughly 29 to 1, and the pages cited by three or more engines are worth 18 times an ordinary placement. Start there, then split the remaining budget by engine.

Method and limits

The dataset covers 403,588 citation events from 72,581 unique URLs across 15,558 domains, collected between 12 June and 8 September 2026 from live AI responses to a tracked prompt set in the AI visibility and GEO software category. Source and page types were classified from domain and URL structure. Month-over-month comparisons use July and August only, since June and September are partial.

Two honest caveats. First, this is one category. The absolute mix — the high share of vendor blogs, for example — reflects a software category where buyers ask comparison questions; a consumer or health category would look different. The relative differences between engines are the transferable finding. Second, per-engine coverage is uneven: Grok and Mistral produce far fewer citations in this prompt set, so their percentages sit on a smaller base.

Questions we get about this data

Which AI engine is the most important to optimise for? By citation volume, Google AI Mode is the largest at 27.8% of events, but the four leading engines are within six points of each other. The more useful question is which engine your buyers use, because the tactics do not transfer — a video strategy that works in AI Mode is worthless in ChatGPT. Does video really help AI visibility? Only in Google AI Mode, where video is 5.6% of citations and rising — it doubled as a share of new citations between July and August. In ChatGPT it is 0.03%. Long-form YouTube videos are cited far more than Shorts. Is press coverage worth it for AI search visibility? Mainstream news accounts for 0.32% of citations in this dataset. Specialist trade and industry publications account for 9.4%. For AI visibility specifically, pitching your category’s trade press is roughly 29 times more productive than national coverage. What is the single highest-leverage placement? A page already cited by three or more engines. These are 1.7% of all cited URLs but carry 22.6% of citations, averaging 73.2 citations each versus 4.0 for a page only one engine cites. How often should we measure AI visibility? Weekly at minimum. On AI Mode and Gemini the median cited URL has a lifespan of zero days, so monthly snapshots compare two largely different source pools and read as noise. Why do listicles get cited so much? They match the shape of the question. Comparison and “best of” pages are 25–39% of citations on every engine and average 7.5 citations per page against 5.2 for standard blog posts. URLs using “best” outperform those using “top” by roughly 50%.

See which pages are citing your competitors instead of you

Opttab tracks the sources behind AI answers across ChatGPT, Gemini, Google AI Mode, Perplexity and Grok — so you can see the exact URLs to target, per engine.

Explore Opttab’s AI visibility platform