GEO Schema Does Not Predict AI Citations

Definition blocks, FAQs, self-referencing canonicals and a references section are mandatory in our publishing pipeline. All 288 English articles carry them, which makes them useless for explaining which article an AI engine actually cites. Treating a constant as a variable is the most expensive mistake in GEO procurement today.

A controlled comparison we did not design

The scaffold was not added for an experiment. It exists so engines can extract our content cleanly: one to three declarative definitions, five FAQ pairs, a self-referencing canonical and a references section on every page. The result is that the features the industry calls GEO scaffolding became a constant in our own corpus.

Over a 30-day window, 17 sessions arrived from AI engine domains such as chatgpt.com, copilot.com and gemini, landing on 11 English articles. Comparing the cited 11 against the other 277 by median:

• FAQ pairs: 5 vs 5
• Definition entries: 3 vs 2
• H2 sections: 6 vs 6
• Paragraphs containing numbers: 13 vs 11
• Numbers per 1,000 words: 12.69 vs 10.67
• Self-referencing canonical and references: 100% vs 100%

The gaps are too small to be a signal. We tested the other direction as well: ranking every article by numbers per 1,000 words and taking the top 30 caught 3.3% of cited articles against a 3.8% site baseline, a lift of 0.87x. Numeric density had no predictive power in this data.

Why published research looks the opposite

Most optimistic GEO numbers come from one experimental design: add a feature to a set of content, then measure visibility before and after. The KDD 2024 GEO paper is the origin of that paradigm, reporting that statistics, citations and quotations can lift visibility by up to 40%. The measurement matters: the paper tracks share of visibility inside AI answers, not referral sessions or clicks.

Both findings can hold at once, provided one distinction survives: a feature works when it is missing and stops discriminating once everyone has it. When every page on a site carries FAQs, definitions and references, those features describe the baseline rather than the winners.

Independent crawl data points the same way. Ahrefs analysed server logs across 137,000 domains and found 97% of llms.txt files received zero requests in May 2026, with AI retrieval bots accounting for roughly 1.1% of the requests that did arrive. Google stated in June 2026 that llms.txt has no effect on Search rankings or AI Overviews.

For buyers the usable reading is narrow: GEO scaffolding solves extraction, not selection. The first is a compliance item that keeps definitions and figures parseable. The second is decided elsewhere.

Where citations actually land

41% of our 17 sessions went to two pages: one on total cost of ownership for agent deployment, one comparing enterprise agent platforms. The rest spread across news and governance explainers at one session or less per page.

External data shows the same pattern. Previsible analysed 6.77 million LLM referral sessions and found ChatGPT accounts for 92.4% of trackable AI referral traffic, sending 28.8% of it to internal search pages because it recognises the domain but not the page. Similarweb reports citations appear in roughly 6.8% of US ChatGPT prompts as of May 2026, up from about 1.6% in June 2025, but professional services sits under 4% while travel and automotive exceed 20%.

Both point at the same variable: engines cite answers that can be compared and checked. Those are the questions buyers bring to an assistant directly.

Our judgement: keep the scaffold as compliance, put the budget on comparable questions

Three moves we would make:

1. Maintain the scaffold as a compliance item. Its job is to keep definitions and figures extractable. That job is done, and doing more of it buys nothing. It does not deserve its own budget line.
2. Concentrate content spend on comparable, quantified questions. Cost, TCO, platform comparisons, migration paths: the questions buyers ask assistants directly. Judge the cluster by new AI-referred sessions over 30 and 90 days, not by article count.
3. Build the measurement before buying the optimisation. Our demand-side baseline is 17 AI sessions in 30 days. At that sample size, per-page invest-or-drop decisions are statistically meaningless and a 90-day window is the minimum. Without per-page AI session counting, no GEO vendor’s promise can be falsified.

A question for buyers

Ask what you are paying for: structural compliance, or incremental citations. If it is the second, answer this first: can you count AI-referred sessions per page today? If you cannot, the only evidence you will ever see is the number your vendor reports.

References: our own site analytics (Umami, read-only, 30-day window, session-level deduplication; the 17 AI-referred sessions, 11 landing articles and the per-feature median comparison across 288 articles all come from internal scripts), analytics.oometa.ai/websites/63b11a98-578c-4e9a-abfe-199fdd684dde | cited sample pages, oometa.ai/en/insights/ai-agent-tco-enterprise-deployment-2026 · oometa.ai/en/insights/enterprise-ai-agent-platform-evaluation-2026 | Aggarwal et al., GEO: Generative Engine Optimization (KDD 2024; the 40% lift is the paper’s metric for share of visibility inside AI answers), arxiv.org/abs/2311.09735 (curl HTTP 200) | Search Engine Journal reporting the Ahrefs log analysis (137,000 domains; 97% of llms.txt files received zero requests; AI retrieval bots about 1.1% of requests; May 2026), searchenginejournal.com/97-of-llms-txt-files-got-no-requests-ahrefs-data-shows/579478 (curl HTTP 200) | report on Google’s June 15, 2026 AI optimisation guide update (llms.txt has no ranking or AI Overviews effect; secondary report, not Google’s own page), digitalapplied.com/blog/google-llms-txt-no-seo-value-lighthouse-audit-2026 (curl HTTP 200) | Previsible 2026 AI Traffic Report (6.77 million LLM referral sessions; ChatGPT 92.4% share, 28.8% landing on internal search), previsible.com/seo-strategy/ai-traffic-report-july-2026 (curl HTTP 200) | Similarweb AI Search Stats 2026 (US ChatGPT citation presence about 1.6% to 6.8%; professional services under 4%), aisearch.similarweb.com/blog/gen-ai-stats (curl HTTP 200). Our own sample is small (17 sessions in 30 days): the finding is about discrimination power, not about per-page decisions.