Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines

In this paper, we explain how to measure brand visibility and analyze responses across AI search engines. We look at what these engines appear to value when they cite brands, what types of sources they rely on, and what kind of content LLMs are more likely to process and surface. The top brands that are already authoritative tend to get cited naturally. The harder and more important problem is for everyone else, including SMEs, D2C brands, creators, early-stage startups, and brands without much of an online footprint.

We demonstrate our findings by analyzing 100K+ prompt responses across 100+ brands tracked on Ranqo between March and May 2026. The very first visibility runs form a clear three-tier brand-stature ladder: global household names such as Stripe and Nike appear in 73% of relevant AI answers on their first tracking run; established mid-market and regional brands such as Olipop and Klaviyo appear in 44%; niche and small brands in just 11% — each tier about 30 percentage points below the last. When the AI search engines cite any sources, about 78% of those citations go to corporate websites, including the brand’s own pages and third-party pages from companies in the same space. Among non-corporate sources, YouTube is cited most often — ahead of Reddit, editorial media, and Wikipedia. At the page level, the highest-leverage citation surface is the ranked “best-of” listicle, the most-cited content format at about 21% of all citations: a single “best X tools” page can surface a brand across many AI answers at once. Sentiment is the unstable part: whether the AI frames a brand positively or negatively flips about 6.7 times more often than whether the brand is mentioned at all.

Together, these findings provide a first large-scale baseline for measuring GEO. They show that AI brand visibility can be measured, differs by platform, and varies strongly by brand maturity. We end the paper by proposing seven v1.1 protocols to test whether specific recommendations can causally improve AI visibility.

Introduction. More and more, people get their answers straight from an AI assistant instead of scrolling through ten blue links. They ask ChatGPT, Claude, Perplexity, Gemini, or Grok what to buy, what tool to use, or who to trust, and they act on what comes back. For a brand, this quietly changes the question that matters. The old question was whether you rank for a keyword. The new question is whether an AI model, when someone asks about your category, actually names you. This is what GEO is about. Generative Engine Optimization (GEO) is the practice of measuring and improving how AI search engines represent, cite, and recommend a brand (Aggarwal et al., 2024). Two related terms get used alongside it: Answer Engine Optimization (AEO) and AI Search Visibility. We treat both as part of the broader GEO scope in this paper, because they describe the same job from slightly different angles. GEO is also not a replacement for SEO. It is built on top of good SEO, since the content an AI engine ingests is, for the most part, the content you already publish for search. There is one more thing worth saying up front. The brands that are already big and authoritative tend to get cited no matter what. The hard, unsolved problem belongs to everyone else: the SMEs, the D2C brands, the creators, the early-stage startups, and any brand without much of an online footprint. That is the audience we care about most, and the gap this paper sets out to measure. A growing body of academic work already tells us that the AI-search playbook is not the old SEO playbook. Generative engines pull from a broader and very different set of sources than classical search, and they lean on their own internal knowledge to different degrees (Kirsten et al., 2025). When they do cite, the citations concentrate on a small set of outlets, and different engines pick different outlets (Yang, 2025). LLMs also reward sources that are already widely cited, a Matthew effect that makes the rich richer (Algaba et al., 2025). And the first careful benchmark of “conversational SEO” tactics found that most of them do not help, and several actively hurt (Puerto et al., 2025). We build on these findings with production data.

1. A way to measure AI brand visibility in production. We describe Ranqo (also Ranqo AI), a platform that issues controlled queries to five AI engines and measures, for each brand, whether the engines mention it, where they rank it, how they frame it, and its share of voice against competitors. It classifies the sources behind those answers, runs a six-dimension page audit, and re-measures on a recurring cadence.

2. A category reference study. We summarize Ranqo’s published CRM study (Ranqo, 2026b): 50 unbranded prompts sent to all five engines ten times each (2,500 responses) across ten CRM brands.

3. Findings from production data. A day-1 recognition finding (§6.2); mention behavior that is mostly deterministic, with flat trajectories absent any intervention (§6.4); a source breakdown showing the brand’s own site is a small slice and most citations go to other companies (§6.5); sentiment that is far noisier than mention (§6.6); the brand-stature visibility ladder, which is the headline result (§6.1).

4. A practical read, especially for smaller brands. We turn the findings into plain guidance for the brands that do not get cited automatically, and we propose seven v1.1 protocols to test whether specific recommendations causally move AI visibility.

The rest of the paper is laid out as follows. §2 covers the related work briefly. §3 describes the Ranqo platform and §4 the measurement methodology. §5 gives the CRM reference study. §6 presents the findings and the v1.1 protocols. §7 discusses what they mean, §8 lists the limits, and §9 concludes.

Related work. A handful of terms now describe overlapping work, so we fix what we mean by each, following the practitioner usage in (Ranqo, 2026a,c). SEO is about blue-link ranking on classical search engines. AEO is about being the single source a featured snippet or direct-answer box pulls from. GEO is about being cited and recommended when a large language model synthesizes an answer across many sources. AI Visibility is the broader question of whether an entity is represented inside a model’s parameters and retrieval indexes at all. In this paper we use GEO as the umbrella term and keep AEO and AI Visibility under it, because for a brand the practical surface is the same: what the engines ingest, and what they say back.

The academic picture is young but consistent on a few points, and we lean on it throughout. The foundational benchmark is the Princeton GEO study (Aggarwal et al., 2024) (KDD 2024), which ran 10,000 queries and tested content changes for their effect on whether an answer cites a page. The biggest wins came from adding machine-extractable provenance: quotations, statistics, and citations, each worth roughly 25–40% more visibility. That is still the core of the modern GEO playbook. The most useful corrective is C-SEO Bench (Puerto et al., 2025), the first systematic benchmark of “conversational SEO” tactics. Its headline is blunt: most of those tactics do not help, and several hurt, while plain source relevance keeps working. This shapes how we run our own experiments. We test changes to page-quality signals on the source side, not prompt-engineering tricks on the response.

On how engines actually cite, three recent studies line up with what we see. Kirsten et al. (2025) show generative engines pull from a broader and very different source set than classical search, and rely on their own internal knowledge to varying degrees. Yang (2025), across roughly 366,000 citations, find that citations concentrate on a small set of outlets and that different engines pick different ones, so cross-platform agreement is low. Algaba et al. (2025) show LLMs prefer sources that are already heavily cited, a Matthew effect. Together these explain a pattern we keep coming back to: the brands with the most existing web presence get cited the most, which is exactly the hill smaller brands have to climb.

Method. 3 The Ranqo Platform For each tracked brand, Ranqo generates a set of naturallanguage prompts spanning six categories (discovery, problem_solution, comparison, use_case, expert, and brand_research) and issues each prompt to all five engines via official APIs. Table 1 summarizes the per-engine configuration.

3.2 The six prompt categories The six categories are not interchangeable. A brand can look strong in one and be almost invisible in another, so we track all six separately. Using running shoes as the example throughout:

Discovery Broad, top-of-funnel questions where the buyer is still exploring and names no brand. “best running shoes,” “good running shoes for beginners.”

Problem/solution The buyer has a specific problem and wants something that fixes it. “I have ankle pain, what kind of running shoes should I wear,” “shoes for flat feet.”

Use case A specific activity or setting. “best shoes for trail running,” “shoes for standing all day at work.”

Comparison Weighing named options against each other.

“Nike Pegasus vs Brooks Ghost,” “Hoka or Asics for a marathon.” These usually name brands.

Expert Deeper, more technical questions from an informed buyer. “what heel-to-toe drop suits a stability shoe,” “does a carbon plate actually make me faster.”

Brand research Questions about one named brand. “is Nike a good running brand,” “Nike running shoes reviews.” These name the brand by definition.

The split matters for measurement. The first four (discovery, problem/solution, use case, and expert) are normally unbranded, so they test whether an engine surfaces the brand when nobody has named it. That is the hard, GEO-relevant case, and it is the one we lean on throughout the paper. Only comparison and brand research name a brand, which makes a mention much more likely and tells us far less about discovery. For that reason the starter set we generate for a new brand is deliberately unbranded.

Prompt themes. Categories describe the shape of a single prompt. Themes sit one level up. A theme is a topical cluster (a product line, a use-case area, a competitor set, a buying occasion), and the platform generates a batch of prompts under each theme while keeping the category mix balanced inside it. Themes let a brand track a whole topic at once (say, “trail running shoes” or “shoes for nurses”) instead of writing prompts one at a time, which is how tracking scales from a handful of prompts to broad topic coverage.

3.3 Citation extraction For each engine response, Ranqo extracts:

• Brand-mention signals: a Boolean brandMentioned, an integer brandPosition (1st, 2nd, 3rd, ...), a sentiment label (positive / neutral / negative) accompanied by a continuous sentimentScore ∈[−1, 1] and an array of sentiment-bearing text spans.

• Source citations: every URL the engine grounded the answer on, normalized to canonical domain and classified into a type (news_media, community, review, video, social, developer, academic, government, reference, corporate).

• Source-to-brand relationship: each cited source is labeled as you, competitor, non_competitor, editorial, forum, review, social, reference, institutional, or other.

• Competitor mentions: an array of competitor names with their ordinal positions.

All signals are stored at the level of a single (prompt, platform, run) tuple.

3.4 Scoring methodology Each (prompt, platform, run) tuple carries three brand signals: a Boolean mention, an ordinal position when mentioned, and a sentiment label. The three behave very differently — mention is stable while sentiment is far noisier (§6.6) — so we report them as separate metrics rather than collapse them into one weighted score.

3.5 Six-dimension page audit In parallel with the query layer, Ranqo runs a sixdimension page audit against each tracked URL, scoring each dimension 0–100 and combining them into a weighted overall page score:

1. Crawlability — whether an AI crawler can reach and parse the page at all: robots, sitemap, canonical, HTTPS, JS rendering, and mobile-friendly checks.

Discussion. 1. Know which tier you are in before you spend on GEO. The right GEO program looks different in each tier; we frame these as hypotheses consistent with the data, not established commercial claims. Tier 1 global brands sit at 73% day-1 visibility, effectively at saturation (the n = 11 Tier-1 cell makes strong sub-claims fragile); their need is share-of-voice against competitors rather than basic recognition. Tier 2 brands sit at 44% with the largest absolute headroom, but the nointervention baseline (§6.4) shows unbranded visibility drifting flat-to-down rather than climbing on its own; whether targeted intervention closes that headroom is the v1.1 RCT question (Protocol P3), not a baseline finding. Tier 3 brands sit at 11%; the fastest Tier-3 movers we observed (§6.1) still travel only 10–20 absolute points across the window, so the highest-confidence move is brand-mass investment (Wikipedia, mainstream press, sustained YouTube presence) to clear the Tier 2 threshold before running heavy per-engine optimization. The specifics differ by sector; see our SaaS playbook (Ranqo, 2026f) and DTC guide (Ranqo, 2026g).

2. Treat each engine as a separate market. Crossplatform source overlap is low. Yang et al. (Yang, 2025) found this across 366,000 news citations; we find the same pattern across 102,025 brand-tracking responses. A Perplexity win does not transfer to Gemini, and the working playbook for each engine differs accordingly (Ranqo, 2026e).

3. Put your facts up front, not your schema markup. Aggarwal et al. (2024) showed that content-quality additions (quotations, statistics, citations) drive the largest GEO lifts. Heavy structured-data markup, by contrast, does little on its own, which is what Protocol P4 sets out to confirm. The practical move is to state your key facts clearly and early on the page, where the engine is most likely to read them; the page-level factors that most affect whether a brand gets cited are detailed in (Ranqo, 2026d).

4. Close the loop in weeks, not quarters. Our day-1 recognition finding (§6.2) shows established brands are recognized within days, while brand-new entities take longer. Either way the loop runs on weeks: make a change, then re-audit and iterate at 14–30 days.

7.2 The honest causal status of this dataset The empirical section reports baselines, trajectories, and source and sentiment composition. It does not report a randomized controlled trial of the recommendation engine. We separate three claim tiers carefully. What the data supports. The per-tier visibility ladder, the day-1 recognition, the deterministic-cell finding, and the source-composition and sentiment shares. These are direct readings of production telemetry. What the data hypothesizes. That the Tier 3 brand-mass ceiling can be lifted with targeted content investment. This is a hypothesis derived from the data; it needs RCT-grade validation. What the data does not yet support: any causal claim that Ranqo’s recommendations cause a measured visibility lift on a brand. The recommendation engine’s closedloop machinery is built into the data model (§3.6), but population-scale closed-loop measurement is left to v1.1 (Protocol P3). What it most clearly establishes. The sharp three-tier day-1 visibility ladder of §6.1 (robust to the Tier-1 leaveone-out reclassification). Whether intervention accelerates those trajectories is the v1.1 RCT question (Protocol P3); we do not pre-commit to a multiplier we have not derived from a controlled experiment.

7.3 A note on white-hat measurement One stance is worth stating plainly. Ranqo measures and recommends on page-quality signals only. It does not generate hidden text, ranking payloads, or anything designed to trick an engine, and the recommendation

Conclusion. In our 102-brand cohort, brand stature is the dominant determinant of day-1 AI visibility. The five engines diverge on sources, position, and sentiment, and they respond differently to schema markup, bylines, freshness, and web-search grounding. On the trajectory side, unbranded visibility mostly holds when a brand does nothing, drifting down only slowly on ChatGPT and Perplexity over our window. In practice, this means a brand’s GEO program has to measure each engine separately, intervene on the upstream signals the engines actually retain, and remeasure on a cycle of weeks; a quarterly cadence is too slow for a medium that re-trains on its own outputs. Three lines of future work follow:

1. A multi-vendor replication track. We publish our protocols as an invitation.

2. An RCT-grade closed-loop study (Protocol P3, §6.7), in the shape of Aggarwal et al. (2024), with random assignment to treatment vs. control recommendations on Ranqo’s brand fleet. Its results, the measured visibility lift a brand can expect from acting on the platform’s recommendations, would be the focus of follow-up work.

3. A training-vs-retrieval separation benchmark. The web-search-on/off design (Protocol P6, §6.7) generalizes to a public benchmark.

Limitations. We name the constraints on the empirical claims directly. A reader should weigh the headline findings against this list.

1. No causal claim on the recommendation engine. Every trajectory in §6.1 is an observational baseline, not a treatment effect; closed-loop measurement on customer brands remains future work (Protocol P3).

2. Convenience-sampled cohort. The brands skew toward SaaS, retail-execution, fintech, and Indian DTC, so we do not claim category representativeness. Several cuts run on small cells (the Tier 1 group is n = 11, and the Agency-tier engines Grok and Claude run on far fewer brands than the rest). We report n next to every aggregate and flag small-sample readings.

3. Manual tier classification. The tier assignment is hand-coded from public signals of brand stature (Wikipedia, press, funding). Because those signals are themselves proxies for web prominence, the ladder should be read as quantifying the size of an expected effect, not as independent proof that prominence drives citation. The rubric and rosters are in Appendix B; a formal inter-rater check is left to v1.1. The day-1 ladder is robust to the reclassifications we tried (Tier-1 leave-one-out range [70.5, 76.7]%).