How to Get Cited by AI Search: What ChatGPT and Perplexity Actually Link To

The short version

AI search engines read almost everything and credit almost no one. Cloudflare measured Anthropic's crawler pulling roughly 38,000 pages for every single visitor it sent back to a website in mid-2025. A separate analysis of 548,534 pages that ChatGPT actually retrieved across 15,000 prompts found that 85% were never cited in the final answer.

So the real question is not whether AI will read your content. It will. The question is what lands in the 15% that gets named and linked. The 2026 research converges on one answer: original, dense, extractable data, placed where the model looks first.

AI takes far more than it gives back

The old search bargain was simple. A crawler indexed your page, and in return the engine sent you visitors. That bargain is breaking.

Cloudflare, which sits in front of a large share of the web, published the ratios in August 2025. For every visitor an AI company referred back to a site, here is how many pages its crawlers had already taken:

• Anthropic: 286,930 pages per referral in January 2025, falling to 38,066 by July.
• OpenAI: about 1,091 pages per referral.
• Google: about 5 pages per referral.

Google still roughly honors the old deal. The AI-native crawlers do not. They consume orders of magnitude more content than they return in traffic. That shift changes what you should optimize for. When clicks from AI answers are this scarce, the unit that matters is no longer the visit. It is the citation: being the source the model names when it answers.

Most retrieved pages are never cited

Getting crawled is not the goal, because getting crawled is nearly universal. Getting cited is the goal, and it is rare.

AirOps studied 548,534 pages that ChatGPT retrieved across 15,000 prompts (reported by Search Engine Land in March 2026). Only 15% of retrieved pages were cited in the final answer. The other 85% were read and discarded. Pages ranking number one on Google were cited 3.5 times more often than pages outside the top 20, so classic search authority still feeds AI citation. But authority alone does not explain the gap. Something about the cited pages makes them easier to quote.

What separates the cited 15%: data density

The clearest signal comes from Kevin Indig and Amanda Johnson's Growth Memo analysis in June 2026, which scored pages on "information gain," a proxy for how much novel information a page adds beyond what is already common on the web. AI engines reward novelty, and novelty tracks with data.

• Pages with 15 or more unique figures averaged an information-gain score of 62.1.
• Pages with one figure or fewer averaged 40.2.
• The top organic results for a typical query carry only about 4 unique data points.

That last number is the opportunity. The bar is low. Most pages that rank are thin on actual data, so a page carrying real, specific numbers stands out to a model looking for something novel to quote.

Where you put the data matters as much as having it

Having the numbers is half the job. Placement is the other half. The same Growth Memo analysis found citations cluster heavily at the top of a page:

• 44.2% of all citations come from the first 30% of the page.
• The middle 30% to 70% earns 31.1%.
• The final 10% earns between 2.4% and 4.4%.

Bury your best statistic in a conclusion and the model is far less likely to reach it. Lead with it.

The academic backing

This is not just correlation from field data. A Princeton-led paper, "Generative Engine Optimization," presented at ACM KDD 2024, ran controlled tests across roughly 10,000 queries in nine domains. It found that the right on-page changes lift a source's visibility in generative answers by up to 40%. The highest-impact, lowest-effort moves were consistent: add statistics, cite your sources, and add quotations. Fluency edits and authority signals helped less than simply putting verifiable numbers and citations on the page.

The playbook

The research points to five concrete moves, in rough order of leverage:

1. Publish original data. First-party numbers, surveys, benchmarks, and pricing you gathered yourself are the strongest source of information gain. You cannot be the novel source if you are repeating everyone else's figures.
2. Make it dense. Clear the 4-data-point baseline and keep going. Pages with 15+ figures scored 50% higher on information gain than near-empty pages.
3. Front-load it. Put your headline number and key figures in the first third of the page, where 44% of citations are earned.
4. Cite your sources. Attribution and direct quotations both lifted visibility in the Princeton tests. Sourced claims read as more trustworthy to a model deciding what to repeat.
5. Make it extractable. Write one fact per sentence. Use tables and short, self-contained statements. A model quoting your page should be able to lift a line without untangling it from three clauses of hedging.

The catch: owning the data is not the same as being cited

Growth Memo's sharpest finding is a warning. Owning original data does not guarantee the citation. If a competitor repackages your numbers in a cleaner, more extraction-friendly format on a more trusted domain, they capture the citation your research generated. The credit follows the clearest presentation, not the original source.

So the work does not end at running the survey. Format the result so plainly, and place it so prominently, that yours is the version a model reaches for first.

Why this shapes how we publish at Toolradar

We take our own advice. Toolradar's tool pages and comparisons lead with first-party data: editorial scores from hands-on testing, pricing pulled from official sources, and side-by-side numbers you cannot get by paraphrasing a vendor's homepage. It is dense, front-loaded, and sourced, which is exactly why AI answer engines cite it.

The same principles work on any site. Stop optimizing for a click that AI increasingly will not send, and start earning the citation instead. Publish something true and specific that nobody else has, put the number first, and make it impossible to misquote.