
What Actually Earns an AI Citation
Neither keywords nor length. The pages AI ignores are longer and carry more markup than the pages it quotes.
// ORIGINAL_RESEARCH · AUGUST_2026
For one full month we recorded every source that ChatGPT, Google AI Overview and Perplexity used to answer real business questions. Not a sample of the web. Every answer, every day, for 31 days.
89
questions tracked
6,601
AI answers recorded
61,536
sources quoted
9,702
pages crawled and stored
31
consecutive days
Most research into AI search starts by inventing questions, then reports what AI says about them. The questions get chosen because they are easy to gather, not because anyone needed the answer.
We started from the opposite end. Every question here was agreed with a chief executive or business owner as one their company genuinely needs to monitor, because the answer decides whether a buyer shortlists them or never hears of them. They come from business software categories and run the length of the buying journey.
That makes this a smaller set than a scrape of thousands of keywords, and a far more serious one.
Each question was put to all three engines every day for the whole of August 2026. We stored the complete answer along with every source the engine quoted, producing 6,601 answers and 61,536 citations.
We then visited all 9,702 unique pages that had been quoted and kept a copy of each one exactly as the server delivered it, so we could examine what was on the page rather than guess from the link.
Because every question repeated daily for a month, we can see something a single snapshot cannot: which sources an engine returns to again and again, and which appear once and vanish.
Grouped by the part of the business they affect. Each is written for a decision maker, not a specialist.

Neither keywords nor length. The pages AI ignores are longer and carry more markup than the pages it quotes.

Ask about price and AI quotes a pricing page 10.6 times more often than usual. Product pages are one in eighteen citations.

Pages cited for comparison questions are 3.2 times more likely to carry a real data table. Bullet lists separate nobody.

16% of pages earn half of all citations, by answering a cluster of related questions inside one topic.

One in four ChatGPT answers names no source at all. Perplexity names one every time, averaging nineteen.

A redirect keeps 94% of a page's staying power. A broken page keeps 59% and is 7x less likely to hold a long lived citation.

Google AI Overview points at a bare website three times more often than ChatGPT or Perplexity.

The original, 99.8% of the time. A long standing worry that does not show up in the data.

Having markup changed nothing measurable. One specific type was the exception.
These findings describe what AI genuinely quotes, at scale, every day, for a month. That is not the same as proving what causes any single citation. What it gives you is a clear picture of what quoted pages have in common and what ignored pages do not, which is the practical foundation for understanding why a page gets chosen and what to change about your own.
Every brand in these articles is anonymised. Numbers, methods and sample sizes are published in full so the work can be checked.
Run your brand through our free AEO audit and see which AI engines cite you today.
Get a free audit →