Organic

Why Most Original Data Never Gets Cited (And the One Format That Does)

New citation research shows that original data is rare in AI answers, but hugely over-weighted when it appears, and almost all of that advantage comes from one specific format. This piece breaks down what a citable benchmark actually requires and why most B2B brands are sitting on the raw material for one without knowing it.
Kathryn Kleist
VP of Content Strategy, Arcalea
Jul 10, 2026 · Updated Jul 10, 2026 · 10 min read

Quick Answer: A July 2026 Growth Memo analysis of Gauge's citation data found that primary research accounted for only 2.7% of AI-cited pages in the sample, yet those pages earned 3.3 times the citation density of the rest. The advantage was not spread evenly. Most of it concentrated in a single format: a benchmark that directly answers a "which is best" comparison. Publishing proprietary data is necessary, but not sufficient. The data has to be packaged as a named, measurable comparison, not buried inside a narrative report, or it will not get picked up.

Most brands that publish original data assume they have done the hard part. They ran the survey, pulled the usage numbers, built the report. The data is real, proprietary, and exactly the kind of thing every GEO guide says AI systems reward.

Then, it isn't cited. Not because the data is weak, but because almost none of it was built in the one shape that actually earns citations.

What the Data Actually Shows

The Growth Memo team audited 301 live pages cited by AI systems across 316 prompts and seven verticals, totaling 1,075 citations. Only 8 of those pages, 2.7% of the set, qualified as genuine primary research, meaning the page itself was the source of the data and method rather than a summary of someone else's numbers.

Those 8 pages earned 90 of the 1,075 citations, more than 8% of total citation volume from under 3% of the pages. Primary research averaged 11.3 citations per page against 3.4 for everything else, a 3.3x density advantage. Original data is rare in AI answers, and it punches far above its weight when it shows up.

Original Data Isn't the Lever, the Benchmark Is

Here is where the finding gets more specific than "publish original data, and you will get cited." The 90 primary research citations in the study were not evenly distributed across topics. The large majority, 75 of the 90, came from a single cluster: cloud data warehouse benchmarks that ranked named vendors on speed and cost. Strip that cluster out and first-party research barely registers in the rest of the dataset.

The pattern held wherever there was a clear head-to-head comparison. A staking and blockchain infrastructure benchmark earned citations at a smaller scale for the same reason. Categories without an obvious benchmark question produced no cited primary research pages at all, even when brands in those categories had published original survey data or usage reports.

The mechanism is not complicated once you see it. AI systems reach for a benchmark when a prompt asks which option is best on a measurable spec: speed, cost, latency, accuracy, yield. A page that answers that question with named comparisons and real numbers gives the model something to lift directly. A page that reports the same underlying data as a narrative essay does not, no matter how original the numbers are.

What a Citable Benchmark Actually Requires

The research points to four structural requirements that separate a benchmark that gets cited from proprietary data that sits unused.

The comparison finding has to lead the page rather than wait for it. Which option is fastest, cheapest, or highest-performing belongs in the first section, not buried after several paragraphs of setup, with the method and nuance following.

The methodology needs to be as visible as it is accurate, since what was measured, over what time window, and using which sample deserves its own clearly labeled section rather than a footnote. A model treats a documented method as a trust signal distinct from the number itself.

The comparison itself has to be made explicit, with named options evaluated against named specs, ideally in a table, since that is the shape AI systems lift for "which is best" prompts. A narrative description of the same finding is much harder to extract cleanly.

And the URL has to hold still. One of the benchmark pages behind the largest citation share in the study was published in 2022 and is still earning citations today, in part because it never moved, while the study also found that of 365 cited URLs, 64 had gone dead, redirected, or otherwise broken, taking 203 citations down with them. A citation earned this quarter only compounds if the page is still standing next quarter.

The strongest example in the dataset, a cloud data warehouse benchmark, also openly showed its limitations: dated correction notes, named caveats about what the numbers did and did not prove, and links to the underlying data and sources. That transparency appears to build trust rather than undermine it, much like a well-sourced analyst report earns more confidence than one that hides its assumptions.

Why This Matters for B2B Specifically

Most B2B organizations already sit on the raw material for a benchmark. Usage data, pricing data, performance data, and customer outcome data exist in product analytics and CRM systems, but no one has packaged them as a named comparison. The data is proprietary, just buried in a dashboard instead of published as an answer to the comparison question buyers are actually asking an AI tool.

This is the same principle behind Arcalea's development of measurement products like the AEO Index and Compass, which serve as benchmarks rather than blog posts. A benchmark that names the comparison, documents the method, and stays at a stable address is the asset that compounds in AI citation over time. A blog post making the same underlying argument, without the named comparison and the visible methodology, competes for attention instead of earning a citation.

The Data You Already Have

Most B2B organizations are not short on proprietary data. They are short on a rationale for packaging it as an answer rather than an appendix. The brands earning outsized AI citations right now are not the ones with the most original research. They are the ones who turned whatever data they already had into a benchmark someone else's model wanted to cite.

Suppose you want help identifying which of your product or usage data could become a citable benchmark; that is the kind of GEO work we build. If you want to see whether your existing content is already earning AI citations or getting quietly skipped, the AEO Index can show you where you stand.

Frequently Asked Questions

Answers to the questions we hear most often about the original data and why it's rare in AI answers.

Original data helps, but publishing it alone is not sufficient. Research from Growth Memo found that primary-research pages earned 3.3 times the citation density of other content, but almost all of that advantage concentrated in pages structured as a direct comparison benchmark rather than a narrative report of the same numbers.

A GEO benchmark is a page that measures a set of named options against a specific, quantifiable standard, such as speed, cost, or performance, and publishes the results as a direct comparison. It differs from general original research by explicitly answering a "which is best" question rather than reporting findings in narrative form.

Most first-party research pages are rarely cited because they report findings as narrative rather than as a named, measurable comparison. AI systems reach for benchmarks when a prompt asks which option is best on a specific spec, and a page without a clear comparison table or head-to-head structure gives the model nothing clean to extract.

AI systems favor comparison pages that lead with the result, document a clear methodology, explicitly name the options being compared, and remain at a stable URL over time. Pages missing any of these, especially a documented method or a fixed address, lose citation share even when the underlying data is original.

Start by identifying a comparison question your buyers are already asking, then structure your existing usage, pricing, or performance data to answer it directly. Lead with the finding, document exactly how you measured it, name the options being compared, and publish it at a permanent URL you do not plan to move or redesign away.

Ready to Put a Framework Behind Your Strategy?

Arcalea's Measure, Accelerate, and Amplify model starts with strategic clarity before any tactic is executed.