Chapters throughout this book cite an OppAlerts study: roughly 403,000 prompts across 100 industries, answered by 13 models, with results correlated against Google rankings, backlinks, Wikipedia, Wikidata, Reddit, news coverage, and video. This chapter is the appendix for that study. It publishes the design so you can judge the evidence yourself, consolidates the findings the chapters cite, and lists the external research worth reading alongside it.The study is part of the AI Search Visibility research, published free.
Study design: prompts, models, personas, industries
The unit of the study is a recommendation prompt: a question a real buyer would ask, of the form "what are the best X for Y", run repeatedly against a model and parsed for which brands the answer names, in what order.
The dimensions:
- Industries: 100 consumer and business categories, chosen for search volume and commercial intent, from auto insurance to sneaker brands.
- Models: 13 models from the major providers, all queried through their APIs in pure-memory mode (no web search), of which 5 were also run with live search enabled, so memory and searched answers can be compared on the same prompts. The memory-versus-search comparison rests on 5,440 paired answers.
- Personas: 1,100 persona framings, 11 per industry: the same commercial question asked as different buyers (budget-conscious, luxury, senior, and so on), because who is asking changes the answer, and aggregate-only measurement hides that.
- Repetition: every prompt runs multiple times per model, because generation is sampled and a single answer is one draw from a distribution. Published measurement research reaches the same requirement independently.Don't Measure Once: Measuring Visibility in AI Search, which argues from empirical studies that AI visibility must be characterized as a distribution over repeated measurements, not a single observation.
Answers are parsed into brand mentions and resolved to registrable domains, roughly 138,000 recommended brand domains in total. Each domain is then joined against the external signal sets: Google organic rankings for the category's queries, backlink authority, Wikipedia presence and citation counts, Wikidata entities, Reddit discussion volume, presence in category news, and video presence. For the 5 search-enabled models, the study also logs which pages each model retrieved while answering and which sources it named, so what gets read and what gets credited can be separated.
The findings, consolidated
Each finding below is treated in depth in its owning chapter; the numbers are collected here so they live in one place and get maintained together. The owning chapter carries the interpretation and the caveats; this table is the lookup.
| Finding | The number | Owning chapter |
|---|---|---|
| Backlinks are the strongest single correlate of memory recommendations, and the strongest independent one | rho 0.23; partial correlation 0.077 with other signals held fixed | What Actually Correlates with AI Search Visibility |
| Wikipedia and Wikidata add almost nothing once backlinks are controlled for | partial correlations -0.011 and 0.002 | What Actually Correlates with AI Search Visibility |
| Wikipedia works as a threshold: presence predicts, depth barely does | rho 0.126 for presence vs 0.011 for citation volume | What Actually Correlates with AI Search Visibility |
| Live search rewrites a large share of the answer | 44.1% of brands in a searched answer absent from the same model's memory answer | Model Memory and Live Retrieval |
| Roughly half of recommendations come without the model reading the brand's site | 53.2% of recommended brands never fetched directly | Model Memory and Live Retrieval |
| What models read and what they credit are different lists | Reddit credited about once per 100 retrievals; brands' own sites at least 53.9% of named citations; Forbes credited in 51 of 100 industries | What Actually Correlates with AI Search Visibility |
| Video is a mirror of fame, not a citation channel | 0.5% of grounded-model citations, vs 9.2% for Reddit | Multimodal: Images, Video, and Audio |
| News coverage predicts memory answers more than searched answers | rho 0.169 vs 0.139 on the same brands; 12 of 13 memory models out-predict every search model | Digital PR, News, and the Gatekeeper Publishers |
| AI answers concentrate about twice as hard as Google organic | top five brands hold 49% of AI answer share vs 24% in Google organic | Building Your AI Search Strategy |
| Google visibility still predicts AI visibility independently | partial correlation about 0.17 | What Carries Over from SEO, and What Breaks |
Industry benchmarks: concentration, persona sensitivity, volatility
The study produces three per-industry scores that the strategy chapters use to size the fight. Concentration: how much answer share the category's top brands hold, which decides whether the game is displacing an incumbent or claiming open ground. Persona sensitivity: how much the recommended brands change when the asker changes, which decides whether the aggregate leaderboard is one competition or several. Volatility: how much answers vary run to run and model to model, which sets the sample sizes any honest measurement needs. How to compute each for your own category is in Building Your AI Search Strategy and Measuring AI Search Visibility.
What the study can and cannot tell you
The correlational findings are correlational. Famous brands have backlinks, Reddit threads, and Wikipedia pages because they are famous, and models recommend them partly for the same reason, so no correlation here proves that adding a signal causes a recommendation. The partial-correlation tests narrow the field to signals with independent relationships, and controlled experiments elsewhere cover the content side,GEO: Generative Engine Optimization (Aggarwal et al.), the benchmark paper that measured visibility changes from controlled content edits. but causal claims in this book are flagged as such where they appear.
The study is also a snapshot. Models ship, indexes move, and every number here carries an implicit "as of the study window." The guide's update cadence exists for that reason, and the findings table above is maintained with it, so a number in this table supersedes the same number quoted anywhere else if they ever disagree.
The external research library
Three papers are worth reading in full alongside this study. The GEO paper (Aggarwal et al.) coined the field's most-used term, built the GEO-bench query benchmark, and measured content-side visibility experiments, reporting gains of up to 40% in its best case. "Don't Measure Once" is the measurement-methodology paper: it demonstrates why single observations mislead and repeated sampling is mandatory, which this study's design follows. And the critical survey of GEO reviews 45 studies across the field, finding that topical relevance and context position are the most reproducible levers, that generic heuristics transfer poorly between settings, and that the foundational paper's gains are conditional on a source already being present in the model's context.Optimizing Visibility in Generative Engines: A Critical Survey of GEO, which also argues GEO is a stochastic, partially observable pipeline rather than a single ranking task.
Read the survey's framing even if you skip the rest: the pipeline it describes, from crawling through retrieval, context allocation, citation, and user behavior, is the same machinery this book's Part II walks through, arrived at independently. Where the survey and this study disagree, say on the strength of a content tactic, trust the controlled result over the correlational one and test it in your own category.