Why most original data never gets cited

Part 1 tackled those all-important third-party citation signals, while Part 2 made the lawsuit for publishing original data: It is the strongest azygous predictor of page originality, and the barroom to earning visibility/authority via this play is low.

This memo has much ammo to backmost up your usage of proprietary information successful contented creation.

Publishing the number is necessary. But it’s not ever what gets cited. We pulled Gauge’s citation information to find retired what AI really rewards erstwhile it comes to publishing first-party data, and the reply is narrower and much useful than “original information wins.” (Although original information does, successful fact, win.)

AI rewards 1 format almost to the removal of everything else: The benchmark that answers “which is best.”

View embedded content

First-party investigation is scarce and punches supra its weight

We worked from Gauge’s cited-URL set: 301 unrecorded pages that AI systems cited (316 unsocial prompts crossed 7 verticals), carrying 1,075 citations betwixt them. 

After a afloat URL audit, only 8 of those 301 pages qualified arsenic superior research, meaning the original root of the information and methodology are connected the page… alternatively than a writeup of personification else’s numbers.

Eight pages retired of 301 is 2.7% of the set. Those aforesaid 8 pages earned 90 of 1,075 citations, aliases 8.4% of citation volume. First-party investigation shows up rarely, past over-indexes 3x connected citation stock erstwhile it does.

The cleaner measurement to spot it is density. 

Primary investigation averaged 11.3 citations per page. Everything other averaged 3.4. A primary-research page was 3.3x arsenic citation-dense arsenic a non-primary one.

Primary investigation compounds citations.

This is the aforesaid style arsenic the information summation uncovering discussed successful Part 2, viewed from the AI broadside alternatively of the classical 10 bluish links side. 

There, original information correlated pinch page originality much than immoderate different trait. Here, original information correlates pinch citation density. Both constituent the aforesaid direction: The number only you tin nutrient is the lever. 

Original investigation wins erstwhile the mobility has a benchmark

Here’s wherever the “original information wins” select gets sharper.

The 90 primary-research citations are not dispersed crossed the 8 pages evenly, and they are not dispersed crossed topics evenly. 

75 of the 90 came from 1 cluster: unreality information storage benchmarks. Fivetran’s storage benchmark unsocial took 44 citations, conscionable nether half of each primary-research citation successful the set. (More connected that below.)

Reality: Strip the benchmark cluster retired and first-party investigation hardly registers successful the citation set. The triumph is not “we published original data.” 

The triumph is “we published a benchmark that answers a buying comparison,” and almost cipher builds one. (“Benchmark” meaning you measurement a group of named things against each different connected a circumstantial yardstick, and people the results arsenic numbers.)

Original investigation is astir effective erstwhile it is packaged successful a measurement that straight answers commercialized comparison queries.

This is what Google is aft pinch non-commodity content: caller and adjuvant accusation that is difficult to get.

Primary-research citations clustered wherever the punctual asked AI to comparison options connected measurable specs: speed, cost, latency, yield, aliases performance.

That explains the storage benchmark spike. The “HR Tech / Compensation” explanation is noisy, but the citations wrong that bucket mostly came from unreality information storage benchmark prompts. Fivetran, Estuary, and ClickHouse had numbers AI could use.

Crypto / Solana shows the aforesaid shape astatine a smaller scale. Marinade and Helius earned citations because staking and MEV questions request first-hand ecosystem data, not generic explainers.

The shape disappears successful topics without a clear benchmark. B2B SaaS / CRM, Education / TEFL, and Product Analytics returned listicles, merchandise pages, explainers, and lawsuit studies. After cleaning, nary of those topics produced a cited primary-research page.

A person look astatine the contented that held 44 of the citations

Fivetran’s storage benchmark took 44 of this information set’s citations connected its own, and Fivetran’s 2 benchmark pages together took 58 of the 90 primary-research citations successful the set. Why?

It’s a portion of contented from 2022, but erstwhile you analyse it, it’s easy to spot why LLMs for illustration it. 

1. It answers a measurable comparison head-on. Named warehouses, BigQuery, Redshift, Snowflake, and Databricks, classed connected velocity and cost. It’s entity-rich and not acrophobic to sanction each the awesome players.
2. It runs connected existent first-party data. Fivetran tested against existent customer usage alternatively than synthetic assumptions, and called retired that prime directly. 
3. It shows the method, measurement by step. Trust signals. Separate sections locomotion done what information they queried, what queries they ran, and really they configured and tuned each warehouse. A scholar (or a model) tin spot precisely really the numbers were produced. 
4. The building is built to beryllium lifted. Descriptive headings (“Results,” “How overmuch did capacity improve?,” “Why are our results different from erstwhile benchmarks?”) fto AI representation a mobility to the 1 transition that answers it.
5. It links to its earthy information and sources. The page footnotes its references, including the C-Store paper, and points to the underlying data, truthful each declare is verifiable. Not galore brands put this overmuch activity into a data-backed contented piece, fto unsocial connection the afloat information group for transparency.
6. It shows its seams. Dated correction notes from December 2022, named qualitative limits, and an honorable “performance floor” caveat make the quantitative claims much credible… not less. They besides statement corrections.
7. The URL ne'er moved. A 2022 page is still collecting citations successful 2026 because it stayed put astatine 1 canonical address.

The information down a page for illustration this is easier to propulsion and analyse than it has ever been. What is not easy is everything astir it: the cleanable method, the linked sources, the corrections, the navigable structure, the willingness to sanction what the numbers do not prove. That’s craft, and that’s the moat here. 

This first-party data-focused portion isn’t a sloppy property merchandise pinch half-assed pulled data. It took a batch of work, and it’s holding authority for 4 years out. The takeaway: AI does not reward “original data” by default. It rewards first-party investigation erstwhile the page provides a clear reply to a measurable comparison that signals extent of expertise and trust.

The unfastened opportunity present is to people a retrievable dataset for a purchaser mobility wherever AI presently has nary cleanable benchmark source. This maps onto the unanswered-questions finding from Part 2: the unfastened doorway exists, and successful these verticals cipher has walked done it pinch a existent dataset.

Original information needs a citation-ready package

Original information gives a page thing AI cannot get from different explainer. But AI still has to retrieve it, parse it, and representation it to the question.

That is wherever galore brands suffer the citation. They people proprietary numbers, but hide them successful narrative, gross them down forms, move the URL, aliases skip the methodology. The information exists. The citation does not.

The pages that won successful this dataset had both: original numbers and a cleanable citation shape. Stable URL. Clear method. Named comparison. Results that answered a purchaser mobility directly.

• Who wins: Brands sitting connected proprietary product, usage, aliases pricing information who package it into a comparison a purchaser tin enactment on, 1 that informs LLM outputs for recommendations. 
• Who loses: Brands publishing original numbers buried successful narrative, connected slow aliases unstable pages, pinch nary comparison framework for AI to lift.

A citation-ready investigation page has 4 parts:

• Lead pinch the comparison result. The header uncovering (“X is fastest, Y is cheapest astatine scale”) goes successful the first 30% of the page. Result, past method, past nuance.
• Box the methodology. Sample, clip window, what was measured, how. Attribution assurance is portion of what makes a number citable. Make your methodology clear connected the page.
• Explicitly framework it arsenic a comparison if it is one. AI reaches for benchmarks connected “which is best” prompts. A array that compares named options connected named specs is the style it lifts.
• Keep the URL stable. One canonical page, kept live, not migrated aliases renamed each redesign. The citation you gain this 4th only compounds if the page is still location adjacent quarter. Of 365 cited URLs successful this information set, 64 were dead, redirected, aliases different broken, taking 203 citations down pinch them.

This is the activity down a citable benchmark, and it is much progressive than it looks. 

HockeyStack documented its ain type successful a playbook connected launching investigation reports: they published 18 original reports built wholly connected anonymized first-party customer data, the benignant nary competitor could replicate.

Their process names each measurement the Fivetran page demonstrates: database the information points you need, get a teammate to propulsion them pinch SQL, specify and archive the method truthful the numbers clasp up to scrutiny, past building the study astir a existent ICP question. They telephone methodology non-negotiable for a reason, noting that without it, personification will ever conflict your data. 

With AI analysis, the information is the easy portion now. Building the contented into thing that is citable, demonstrates E-E-A-T, and is still earning visibility 4 years retired for commercialized queries is wherever the difficult activity lies.

What sites are already trusted for your topic? When a benchmark you did not people is taking the citations successful your category, the Citation Source Mapper maps that trusted group into a ranked, pitchable target list. It is successful the premium library.

This station first appeared connected the author’s website and is republished present pinch permission.