Skip to main content

Research and Benchmarks

Original studies on how AI answer engines cite sources, and benchmarks you can measure your own visibility against.
Steve Lee, Founder of SEO Aesthetic·Written July 14, 2026·Updated July 31, 2026·5 min read
On This Page
  1. 1. Why we publish research
  2. 2. Studies and benchmarks
  3. 3. How to use it
Summary & Key Takeaways
  • Original research into how engines decide what to cite.
  • Benchmarks so you can judge your AI visibility against a baseline.
  • Data drawn from real prompt testing across engines.
  • Research feeds our method and the guides we publish.

1. Why we publish research

AEO is young, and most claims made about it are guesswork dressed up as authority, repeated from one blog post to the next until they sound true. We run structured prompt tests across the major engines and publish what we actually find, including the results that complicate our own assumptions, partly to sharpen our method and partly because original research is exactly the kind of corroborated, citable work answer engines reward. Running the test ourselves is also the only honest way to make claims about a field this new, where almost nobody has the data and most are guessing.

2. Studies and benchmarks

Research
Citation behavior across engines
How ChatGPT, Perplexity, and AI Overviews differ in what they cite.
How many sources get named
The typical size of the cited short list per query type.
Entity consistency and citation
How drift in brand facts correlates with being skipped.
AEO visibility benchmarks
Baselines to measure your own citation share against.

3. How to use it

Use the benchmarks to judge whether your own visibility is ahead of or behind the field, then use the method to close the gap rather than to admire the chart. The research is not published as a vanity exercise: it underpins the Answer Engine Optimization pillar and feeds directly back into how we work, so every benchmark we run either confirms the method or changes it. Numbers that do not change what you do next are just decoration, and we try not to publish those.
An experiment I ran
I ran an original benchmark instead of citing someone else’s, and we turned into the cited source

Rather than cite the same existing stats as everyone else, I ran our own benchmark on a real AEO question across a meaningful sample, and published the full methodology and the raw numbers. The experiment was direct: does original data earn citations that recycled commentary simply never can.

It did, and it kept paying out long after we published. Original research becomes the thing other people cite, which means engines encounter your brand attributed to a primary finding again and again and again, which is the strongest corroboration signal that exists. Producing the data beats summarizing the data, every single time, by a wide margin.

Everyone quotes the study. Almost nobody runs one. Running it is the entire advantage.

HOT TAKE · THE PART NOBODY SAYS OUT LOUD
Original research is the single most underused AEO weapon there is

The entire content world runs on quiet recycling: everyone cites the same tired handful of studies and bolts on their own commentary. That commentary is infinitely replaceable and gets cited essentially never, because it is not the source. It is just one more reaction to the source, and reactions are a commodity.

Original research flips you clean to the other side of the citation. When you produce the primary data, you become the thing the recyclers all quote, and the engines slowly learn your brand as a source of facts. It is genuinely more work, which is exactly why almost nobody does it, which is exactly why it works so well for the few who do.

WHY THIS BEATS THE PASTE-AND-SHIP SHOPS
A model can summarize research. It cannot conduct any.

The commodity move is to have a model summarize a few existing studies into a research-flavored post with a chart. It reads authoritative and contains zero new information, so it just adds one more replaceable summary to a towering pile of them. No new data anywhere in it, so no reason for anything to cite it.

Real research takes designing a method, gathering the data, and being willing to publish what you actually found even when it is inconvenient, done by senior, US-based operators. That is genuinely hard work, and the difficulty itself is the moat. The model can only ever summarize the field. We add new facts to it.

Summarizing research makes you a footnote. Conducting it makes you the citation.

Want to benchmark your own visibility?
Book a consultation and we will measure your citation share against these benchmarks.
Book a Consultation
Frequently asked questions
Where does the data come from?
Structured prompt testing we run across the major answer engines, not vendor claims.
Are benchmarks industry-specific?
Where the data supports it, yes. We avoid over-generalizing from thin samples.
Can I cite your research?
Yes, with attribution. Corroborated research is meant to be shared.
References
  1. 1. OpenAI. ChatGPT search and crawler documentation.
  2. 2. Perplexity. Citation and publisher documentation.
  3. 3. Google Search Central. AI features in Search documentation.
  4. 4. Google. Search Quality Rater Guidelines on original, primary research.
  5. 5. Schema.org. Dataset and Article vocabulary.