A critical survey published in July 2026 reviewed every peer-reviewed GEO study from November 2023 through July 2026 — 45 papers in total. The conclusion is not what most marketers want to hear: no reviewed technique has demonstrated a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior. Not one.
This doesn’t mean GEO is fiction. It means the evidence base for most of what gets written about AI visibility is narrower than the advice built on top of it. The studies that do exist show real effects, but within tightly bounded conditions that don’t survive platform changes, model updates, or even re-runs of the same query. For website owners spending time and money on AI visibility, this gap between evidence and practice is the most important thing to understand right now.
What the Survey Actually Found
The arXiv paper (2607.14035) covers 45 studies across three years of GEO research. The core finding is a pipeline problem: the researchers draw a sharp distinction between two separate questions that most GEO advice conflates.
The first question is discoverability — whether a page appears in an AI system’s candidate retrieval pool at all. The second is citation selection — whether a retrieved page gets used in the final answer. The widely cited 2024 finding that GEO techniques boost visibility by 30–40% applies to the second question, not the first. The studies show that if your content is already in the context window, formatting and sourcing choices can influence whether it gets cited. They do not show that those same choices increase the likelihood your content enters the context window in the first place.
Commercial audits cited in the survey found “low source overlap, substantial run-to-run variability, and persistent fidelity gaps.” In practice, this means two things: AI systems draw from different source pools across platforms, and the same query on the same platform returns different citations on different runs. The optimization target is moving every time you look at it.
The Two Problems GEO Advice Usually Conflates
Most GEO content treats AI visibility as one problem. The survey makes clear it’s at least two, with different drivers and different evidence quality for each.
Discoverability — getting into an AI system’s candidate pool — is primarily determined by factors that look a lot like traditional web authority: crawlability, indexation by AI-specific crawlers (GPTBot, ClaudeBot, PerplexityBot), entity recognition in knowledge graphs, and cross-platform citation by sources that AI training pipelines trust. The GEO evidence here is thin. Earned media driving 84% of AI citations (Muck Rack, May 2026) is about discoverability — third-party mentions push content into training data and retrieval pools in ways that on-page changes cannot replicate.
Citation selection — getting chosen over other retrieved candidates — has more experimental evidence behind it. The factors that consistently show up: specificity over generality (pages that answer a narrow question outperform pages that cover everything), presence of verifiable claims with attribution, clear entity signals (schema markup that tells the AI exactly what an organization does and who it is), and structural markers that match the format of a question-and-answer response. These are real effects, but they’re conditional on your content already being in the retrieval pool.
What Actually Has Reproducible Evidence
The survey identifies two factors as “the most reproducible levers” across studies: topical relevance and context position. Topical relevance means content that precisely matches the semantic intent of a query — not broad category coverage, but narrow answer fit. Context position refers to placement within a document: content closer to the beginning of a page, and organized to resolve a question in the first few sentences, consistently outperforms content that buries the answer in paragraph seven.
Beyond those two, the commercial evidence (not just lab studies) points to a few additional signals that have held across model updates:
- Entity disambiguation via Organization schema with
sameAsproperties linking to Wikidata, Crunchbase, and LinkedIn. AI systems build an internal representation of what your brand is — fragmented entity signals produce fragmented citation behavior. - Third-party corroboration on sources that AI training pipelines prioritize: Wikipedia, G2, Capterra, TrustPilot, major editorial publications. Brandlight’s 2026 analysis shows the overlap between top Google results and AI-cited sources has dropped from 70% to below 20% — AI is pulling from different sources than Google, and many of them are review and reference sites.
- Freshness signals for time-sensitive queries. Ahrefs’ 2026 freshness study found AI systems prefer content published or substantially updated within 90 days for queries with implicit recency intent, even when older content is more authoritative overall.
What’s Not Working (Despite Widespread Adoption)
The survey is explicit that “generic heuristics transfer poorly.” Several widely recommended GEO tactics have weak or contradicted evidence:
llms.txt files — A 300,000-domain study (Search Engine Journal, 2026) found no measurable relationship between having an llms.txt file and AI citation rates. Google’s own May 2026 optimization guide explicitly states that llms.txt is not needed for AI Overviews or AI Mode. As of July 2026, no major AI provider — OpenAI, Google, Anthropic, Meta, or Mistral — has publicly committed to acting on llms.txt in production systems. The file is almost exclusively requested by SEO audit tools, not by the crawlers that drive actual citations.
Keyword density and NLP manipulation — Studies that showed gains from adding high-frequency phrases to content are not replicable across models. GPT-4-era findings don’t hold in GPT-4o or Claude 3.5. Model updates change the citation selection mechanism faster than optimization campaigns can adapt.
AI-specific meta tags — No published evidence shows that any proprietary AI meta tag (beyond standard Open Graph and Schema.org) influences citation rates on any major platform.
The Agent-Ready Gap No Study Has Addressed
The entire body of GEO research reviewed by the survey focuses on one AI modality: chat-style question answering. None of it addresses AI agents — systems that don’t answer questions but execute tasks. This is a structural blind spot, because agent-driven discovery operates on completely different mechanisms than retrieval-augmented chat.
When a user asks ChatGPT to “find me the best project management tool for a 10-person remote team and book a demo,” ChatGPT does not browse your website. It queries structured endpoints — and the emerging standard for those endpoints is the Model Context Protocol (MCP). By mid-2026, the MCP ecosystem has reached 97 million monthly SDK downloads and over 10,000 registered public servers. Anthropic, OpenAI, Google, Microsoft, AWS, and Cloudflare are all members of the Linux Foundation’s Agentic AI Foundation, which now governs the standard.
Brands that have published an MCP server appear as queryable entities in agentic workflows. Brands that haven’t are invisible to them — not because their content is bad, but because there’s no machine-readable endpoint for an agent to query. This is the same structural problem that llms.txt was supposed to solve, but MCP actually solves it: it’s not a hint file that crawlers may or may not read, it’s a live API that agents actively connect to.
The specific minimum viable MCP implementation for a business exposes three capabilities: a product catalog or service list with structured attributes, a booking or contact endpoint, and a FAQ or knowledge base. This mirrors what the GEO studies show works for chat-based AI — narrow, specific, structured answers — but in a form that agents can act on rather than just cite.
How to Prioritize Given What the Evidence Shows
Given 45 studies with limited transferable findings, a practical priority stack looks like this, ranked by evidence quality:
Tier 1 — High evidence, cross-platform: Earned media on AI-cited reference sources (G2, Wikipedia, editorial coverage). Entity disambiguation via schema and sameAs. Content freshness for time-sensitive verticals.
Tier 2 — Moderate evidence, conditional: Narrow topical focus per page (one question, one answer, near the top). Organization and FAQ schema. Crawl access verified for GPTBot, ClaudeBot, and PerplexityBot specifically.
Tier 3 — Low evidence or unproven, but structurally logical: MCP server for agent-driven discovery. Citation monitoring across platforms to identify where you’re absent. Cross-platform citation share tracking (not just Google).
Don’t bother: llms.txt files. AI-specific meta tags. Generic heuristics like “add more statistics” without specificity to your industry. Any optimization strategy derived from a single platform study and applied universally.
Measuring What Can Be Measured
One of the practical problems the survey surfaces is measurement: GEO results are volatile run-to-run, which makes it hard to know if an intervention actually worked. The commercial audits in the survey found substantial variability in which sources AI systems cite across identical queries on different days.
The most reliable measurement approach as of mid-2026 is citation share tracking — structured queries sent to ChatGPT, Perplexity, Claude, and Google AI Overviews on a fixed schedule, logging which sources appear and how frequently. This is different from rank tracking in traditional SEO: there’s no position 1–10, only cited or not cited, and the denominator changes with every model update.
For verticals where AI traffic is growing fast — B2B SaaS reached 4.5% of organic traffic from AI search by Q3 2025, with 127% growth in three months — the measurement gap is costing real revenue. You can’t optimize what you’re not tracking, and you can’t track AI citations with tools built for blue-link rankings.
The Practical Bottom Line
The July 2026 critical survey is a useful correction for an industry that has generated a lot of GEO advice faster than the evidence could support it. The takeaway isn’t that AI visibility optimization doesn’t work — it’s that the specific things most brands are doing have shaky evidence behind them, while the things with consistent evidence (earned media, entity authority, freshness, structured data) look a lot like the fundamentals that were already working before GEO became a category.
The one genuinely new requirement is the agent-ready endpoint: MCP for agentic workflows, structured endpoints that allow AI agents to query your data rather than browse your pages. That’s not an optimization of existing tactics — it’s a new layer of infrastructure that the GEO studies haven’t even begun to measure, because the agent traffic wasn’t there yet when those studies were designed.
The brands that will compound AI visibility over the next 18 months are the ones treating it as two problems: citation presence in chat-based AI (earned media, entity authority, structured content), and queryable presence in agentic AI (MCP endpoints, structured APIs). The first problem has evidence. The second problem has adoption pressure that makes the evidence question almost irrelevant.
To see where you actually stand in AI answers today — across ChatGPT, Perplexity, Google AI Overviews, and Claude — run a free visibility scan at ai-visibility.llmagnet.com. It shows your current citation footprint and the specific gaps your competitors are exploiting.