The research foundation behind AI visibility scoring.

Finding

Coined the term GEO and introduced GEO-bench (10,000 queries across 9 domains). Tested 9 content strategies and showed that adding quotations, statistics, and cited sources measurably raises how much of an answer a source contributes. Lower-ranked sources gained most: with Cite Sources, +115.1% for sources ranked fifth and −30.3% for the top-ranked one (its Table 2). The headline gains are a relative maximum on one metric (position-adjusted word count) and are measured with the source already present in the model’s context. The paper gives absolute scores (its Table 1) and states only that the best methods improve on the baseline by 41% and 28% on its two metrics; the per-method percentages below are relative gains on position-adjusted word count computed from that table.

How GEO Optimizer uses it

The scoring engine’s categories (robots.txt, llms.txt, schema, meta, content, signals, AI discovery, brand entity) and the citability suite are derived from this signal taxonomy. The Google AI readiness category follows Google’s own published documentation instead. Direct quotations and concrete statistics carry the most weight in the content checks.

Finding

Reviews 45 GEO studies across a Nov 2023 – Jul 2026 window and grades the evidence. Strong support: query–document topical relevance and position within the retrieved context are the dominant citation levers, and generative engines differ substantially in their source ecosystems. Weak or unproven: no reviewed technique shows a stable, cross-platform causal effect on organic discoverability or on downstream clicks and conversions. The "+40% visibility" figure is reframed as a conditional relative maximum, not a universal rank gain.

How GEO Optimizer uses it

Confirms the design choice to score machine-readable infrastructure and passage-level structure rather than promise ranking gains. The survey’s "competition erodes individual gains" finding is why GEO Optimizer reports a readiness score, not a projected traffic number.

Finding

Human audit of four generative search engines. On average only 51.5% of generated sentences are fully supported by their citations, and only 74.5% of citations actually support the statement they are attached to. Fluent answers routinely contain unsupported claims.

How GEO Optimizer uses it

Motivates the citability checks that reward self-contained, quotable statements with inline evidence — the passages an engine can attribute cleanly — and the negative-signal checks that flag thin or unsourced content.

Finding

Large-scale audit of Google Search, AI Overviews, and Gemini. AI Overviews surface a markedly different set of sources than the organic results on the same query, cross-surface overlap is low, and answers change on repeated runs for the same query.

How GEO Optimizer uses it

Underpins the "audit recurring, not once" guidance and the State of GEO methodology: each monthly cohort is treated as an independent sample, and re-audits are expected to move as engines update.

Finding

AI-search visibility is a distribution, not a single value. Across four engines (ChatGPT, Gemini, Google AI Mode, Perplexity) over a 45–46 day window, the day-to-day Jaccard similarity of cited sources averaged 0.34–0.42, and 57.8% of ChatGPT runs returned no citations. A one-shot check is an unreliable estimate. The data come from Swiss-German verticals, so the authors caution against generalising.

How GEO Optimizer uses it

Directly shapes the State of GEO benchmark (aggregate distributions, medians and percentiles rather than point claims) and the monitoring product, which runs scheduled repeat checks and tracks score history.

Finding

Proposes treating citation-visibility metrics as sample estimators of an underlying response distribution, with confidence intervals and a minimum number of runs needed before a visibility claim is statistically meaningful.

How GEO Optimizer uses it

Informs how the score-stability figure on the methodology page is reported (measured over thousands of back-to-back audit pairs) and why single-audit deltas are described as directional.

Finding

Learns generative-engine preferences automatically and rewrites content to match them, reporting an average improvement of about 36% over the strongest Princeton method (about 21% for the smaller AutoGEO-Mini) while preserving answer utility. Cooperative, semantically explicit content earns a 35–60% higher citation rate than adversarial or terse equivalents.

How GEO Optimizer uses it

Validates the "cooperative, explicit, machine-readable" direction of the fixer layer (llms.txt, schema, meta, AI discovery generation) over adversarial rewriting, and the choice to generate structure rather than manipulate phrasing.

Finding

Of 15 common GEO heuristics tested on e-commerce content, 10 were neutral or negative. Systematic optimization converges on domain-agnostic structural improvements rather than any single copy trick.

How GEO Optimizer uses it

Reinforces scoring structure (headings, lists, answer-first blocks, schema) over tactic-of-the-month advice, and the e-commerce checks that focus on machine-readable product and offer data.

Finding

First benchmark to test conversational-SEO methods across multiple tasks, domains, and competing actors. Most current C-SEO methods are ineffective or actively hurt document ranking, and the few positive effects erode as more documents adopt the same method.

How GEO Optimizer uses it

Used to validate the citability score and to keep the engine focused on durable structural signals. The "gains erode under adoption" result is why GEO Optimizer scores readiness rather than a competitive edge.

Finding

Prompt injection hidden in a document can move a target source up roughly 3 ranks in a conversational search engine (tested on Perplexity Sonar). Retrieved content is an attack surface: what a page says can influence the ranking of the answer it appears in.

How GEO Optimizer uses it

Drives the prompt-injection pattern detection in the audit (instructions aimed at AI crawlers, hidden directives) and the negative-signal checks that flag manipulative content.

Finding

Preference Manipulation Attacks: crafted website or plugin content tricks an LLM into promoting the attacker and discrediting competitors. Demonstrated on production Bing and Perplexity and on GPT-4 / Claude plugins. Everyone is incentivised to attack, which degrades answers for all users.

How GEO Optimizer uses it

Sets the boundary GEO Optimizer stays inside: it audits and fixes infrastructure and structure, and explicitly does not generate injection payloads or manipulative copy. Detects these patterns as negative signals instead.

Finding

In GeoReady’s audited sample, sites with valid JSON-LD (Organization, WebSite, Article) score higher on the GEO rubric on average than sites with none. That is an association inside a self-selected sample, and the score includes the schema category itself, so it does not show that structured data causes citations. The size of the gap is not published here.

How GEO Optimizer uses it

Drives the Schema JSON-LD scoring category (max 14 points in rubric v2, 16 in v1) and the structured-data fixer that generates complete @context + @type + sameAs blocks.

Finding

GeoReady’s own monthly benchmark of audited domains: average and median GEO score, llms.txt and schema adoption, band distribution, and the biggest readiness gaps, tracked month over month.

How GEO Optimizer uses it

Provides the empirical baseline for the readiness bands, the negative-signal thresholds, and the "what to fix first" ordering in every audit.

Finding

The on-page and technical factors the audit checks because they decide whether an AI system can fetch, parse and quote a page: crawlability, passage-level structure, entity signals, and freshness. Whether each one changes which sources an engine cites is not established; the audit measures readiness, not citation outcomes.

How GEO Optimizer uses it

Mapped into the scoring categories and the technical-signal checks (X-Robots-Tag, noai directives, crawl-delay, canonical, HTTPS).