The available OppAlerts report compares observable web measures with an LLM recommendation score. Several measures have positive correlations. Those associations help identify questions to investigate; they do not show that adding a backlink, review, or encyclopedia entry causes an AI recommendation.
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.What the available table measures
The following values come from the report snapshot generated May 11, 2026. Each row names its actual field and sample size.
| 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% |
Search appearances and backlinks have positive associations in this table. The signals overlap, and their data coverage differs. These values alone cannot establish which signal has the largest association after adjustment.
Partial correlation is a separate calculation that adjusts for specified measured variables. An adjusted result needs its own sample and control definitions. The available table does not establish the adjusted values quoted in earlier drafts.
Presence and volume
A count of Wikipedia citations measures something different from the presence of a Wikipedia article about a company. The same distinction applies to Wikidata entities and simple presence flags.
Check the field definition before making a recommendation. A count correlation cannot show that a model stores a yes/no notability flag. It also cannot establish that creating an entry would change an answer.
What gets read and what gets credited
Record retrieved sources and visible citations separately. A system can consult more sources than it cites. OpenAI’s web-search API documentation distinguishes a sources list from citation annotations in the generated answer.OpenAI web search guide. The documented API exposes source records and URL citation annotations. These fields are specific to the API.
A citation to a review site does not prove that the business’s own website was fetched. An uncited page in a source list does not establish how much it affected the answer. The guide’s earlier source-share percentages require their original analysis before they can be reused.
What a news comparison would need
To compare news coverage with recommendations, record the period of the coverage and the answer collection. Separate the availability of a source from evidence that a model used it.
Current news cannot be assumed to have appeared in an older model’s training data. A relationship between coverage and answers may also reflect company size, market share, or other differences.
Correlation, causation, and how to read the rest of this book
Controlled content experiments, including the GEO paper, provide another form of evidence. Their findings apply to the tested systems, edits, and evaluation measures. They should not be treated as proof of a universal ranking formula.Aggarwal et al., GEO: Generative Engine Optimization, which introduced the GEO-bench benchmark. A later critical survey reviews which of the field's claims have held up and which remain thin; the honest summary is that content-side experiments exist and authority-side experiments largely do not.Don't Measure Once: Measuring Visibility in AI Search, which makes the empirical case for repeated measurement and distribution-level reporting.
Use the research to prioritize questions for your own category. State the proposed mechanism separately from the observation. Then measure whether the change you make is followed by a repeatable result under comparable conditions.
The finding-to-chapter map
| Question to investigate | Next chapter |
|---|---|
| Are relevant pages accessible and discoverable? | Technical accessibility |
| Which independent pages describe the business accurately? | Backlinks and AI search |
| Are identity and product facts clear? | Business information |
| Does the tested content answer the customer’s requirements? | Writing for retrieval and citation |
| Does the result repeat under comparable conditions? | Measurement |
Use a correlation as a reason to investigate. Assign work after identifying an observable problem and a way to verify the proposed correction.