Content and evidence
Publish case data that a reader can verify
Compare a statistics-page experiment with GEO research, then build a verifiable case claim without mistaking citation volume for truth or causation.
TL;DR
- A number is useful evidence only when the reader can trace its definition, scope, period and source.
- Peec's statistics-page experiment included fabricated data; its reported exposure illustrates a verification failure, not proof that those claims were true.
- A historical GEO study found visibility changes in controlled tests, with results that should not be treated as current citation guarantees.
- Publish a complete measurement claim and its limits, rather than a percentage with no baseline.
Good case data helps buyers judge whether a result applies to their problem. A link or an AI citation makes a claim easier to find. It does not prove that the data behind it is sound.
Keep that gap in mind when reading advice to “add statistics for GEO.” A test of which text a system picks is not a test of whether that text is true. Publishers must check both.
What the statistics-page experiment does and does not show
In October 2026, Peec researcher Metehan Yesilyurt described publishing 50,000 fabricated statistics. He reported more than 10,000 ChatGPT sessions since launch, around 72,000 ChatGPT-User requests in a later 30-day window, and more than 330 referring domains. These are the author's reported observations, not an independent audit. See the original experiment account.
| Reported measure | Reported value and period | What it cannot establish |
|---|---|---|
| ChatGPT-attributed sessions | More than 10,000 since launch, about seven months. | The truth of the statistics or the number of purchases. |
| ChatGPT-User requests | About 72,000 in the final 30-day window. | The number of unique answers that cited or endorsed a claim. |
| Referring domains | More than 330 after about seven months. | Independent validation of the underlying data. |
The measures use different units and periods. Do not turn them into a conversion funnel. They also do not show that statistics are the top GEO factor. Make a claim easy to check before making it easy to repeat.
What independent research adds
The 2024 paper GEO: Generative Engine Optimization tested changes to source content. Its Perplexity evaluation used 200 samples, supplying source text as uploaded files and constraining answers to those sources. That design examines answer construction under supplied evidence; it is not an open-web ranking trial.
| Content treatment | Position-adjusted word-count score | Subjective impression score |
|---|---|---|
| No optimization | 24.1 | 24.7 |
| Statistics addition | 26.2 | 33.9 |
| Quotation addition | 29.1 | 32.1 |
The columns are study metrics, not percentages of users or sales. Statistics and quotations differed across the measures. The evidence supports testing useful source improvements; it does not establish a permanent winner or an expected return for a current website.
Peec's account and the paper address different questions. One describes the spread of published material, including false claims. The other tests source-content treatments in a constrained setup. Neither replaces an audit of your own case evidence.
Worked example: replace an unsupported conversion headline
This is a fictional case with invented numbers. A draft says, “Our page redesign increased conversion by 50%.” The worksheet shows 20 accepted inquiries from 1,000 eligible visits before the change, and 30 from 1,000 after it. Paid campaigns also changed during the second period.
Check the inquiry rule, visit filter, period lengths and campaign log. The rate rose from 2% to 3%: one percentage point, or 50% relative to the earlier rate. The data does not isolate the redesign's causal effect.
A defensible claim is:
In this illustrative comparison, accepted inquiries rose from 20 of 1,000 eligible visits to 30 of 1,000 across equal periods. The rate changed from 2% to 3%. Campaign changes occurred at the same time, so this comparison does not isolate the page redesign's effect.
For a real case, use approved records and real dates. State what counts as an accepted inquiry and link to the source. Explain how you handled duplicates, tests and spam. Keep any doubt visible rather than rounding it away.
The case owner approves what can be shared. The analyst checks the math against the source. The editor checks that the headline, chart and summary use the same measure. That tells you more than counting the numbers on the page.
Common questions
Can an anonymous case still be useful?
Yes, if the permitted context, measurement method and limitations are specific enough to judge. Do not invent a country, industry or client identity to make it sound more credible.
What if there is no outcome data?
Show a documented problem, the work performed and an approved output. A concrete work sample is more useful than an invented result.
Semanta's evidence review helps connect commercial claims with records a buyer can inspect. This is the Evidence dimension of our service-page framework. Bring a draft case and its underlying measurements to discuss a case-data review.