This chapter identifies the evidence used in the guide and explains how to check it. Provider documentation supports claims about a product. Published experiments support findings within their tested conditions. OppAlerts report data provides observational comparisons.
A source must support the specific claim attached to it. A link to a report homepage is insufficient when the claim depends on a particular table, sample, or calculation.
The OppAlerts data used here
The available OppAlerts report snapshot was generated on May 11, 2026. Its industry index contains 145 entries. The file includes all-industry signal correlations and per-industry summaries. The generation timestamp identifies the output file, not the dates on which prompts were collected.
The report describes comparisons between external web measures and an LLM recommendation score. The table below reproduces selected values from the published JSON. These are pairwise correlations. They do not adjust for other variables or compare answers with and without web search.
OppAlerts report data. Snapshot generated May 11, 2026. The table contains pairwise correlations with the report’s LLM recommendation score; sample sizes differ by signal.Selected values from the report
Spearman rho measures how two sets of ranks vary together. A positive value indicates that higher ranks on one measure tend to accompany higher ranks on the other. It does not identify the reason for that relationship.
| Measure | Data field | Spearman rho | Domains with paired data | Reported coverage |
|---|---|---|---|---|
| Search engine appearances | serp_appearances | 0.241 | 10,914 | 36.9% |
| Backlink count | bl_backlink_count | 0.204 | 20,402 | 69% |
| Reddit comments | reddit_comments | 0.111 | 6,171 | 20.9% |
| Wikipedia citations | enwiki | 0.077 | 5,761 | 19.5% |
| Wikidata entities | wikidata | 0.120 | 1,619 | 5.5% |
The backlink-count row, for example, reports rho 0.204 across 20,402 domains. The Wikipedia row measures citations, not a yes/no test of article presence. Preserve these distinctions when quoting a result.
Industry comparisons
The public report contains industry and persona views. Use them to investigate the named category. Do not treat the all-industry table as a forecast for a particular business.
Concentration, sensitivity to buyer requirements, and variation across repeated answers require different calculations. The available snapshot does not provide those three benchmarks together. Use the underlying observations and a stated method before comparing industries on those measures.
What the study can and cannot tell you
The figures above were checked against the published JSON. That confirms what the report contains. It does not reproduce the collection or statistical analysis from raw responses.
Data coverage varies by measure. Missing observations must not silently become zeroes. Correlations calculated on different sets of domains do not provide a controlled comparison of which activity would help a business most.
The snapshot alone does not establish the total number of attempted prompts, an adjusted effect for a signal, or the effect of enabling web search. Those claims require the corresponding collection records and analysis.
A reproducible analysis needs collection dates, exact prompts, model identifiers, tool settings, attempted and completed counts, brand-resolution rules, scoring definitions, exclusions, and calculation code. Publishing these details lets another researcher evaluate how a number was obtained.
Related research
The GEO paper tests content interventions in a defined benchmark. Its results concern the experiment’s visibility measures. They do not establish a fixed increase in traffic or sales after applying the same edits to a website.
Don’t Measure Once examines variation across repeated AI answers. It supports retaining repeated observations and reporting variation. It does not make ten runs a universal requirement for every research question.
The July 2026 critical survey reviews the evidence across GEO studies. A literature review is useful for finding experiments and their limitations. It is not an additional independent experiment confirming every result it discusses.
For a project using this guide, write the proposed change and success measure before collecting the follow-up answers. Keep the original observations so the interpretation can be reviewed later.