Lab GEO Studies vs Live AI Engines: What Transfers

A live AI answer passes through four stages:

1. Parametric: what the model knows with no web access, frozen at a training cutoff.
2. Access: whether a crawler may fetch and store your page.
3. Retrieval: whether your page enters the model's context for a given query.
4. Generation: which retrieved documents the model uses and cites.

Nearly every published GEO effect size sits in stage 4. Researchers hand the model a fixed set of documents, edit one, and count how often it gets cited. No index is queried. No crawler visits. No ranking is earned.

A critical survey of 45 GEO studies (arXiv 2607.14035, July 2026) concluded that "already-retrieved content can causally alter its citation or use, but no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior." Our digest of that survey covers the full corpus. The practical read: a lab effect only applies once your page has already been found.

The "GEO gets you up to 40% more visibility" line traces to the original GEO paper (Aggarwal et al., arXiv 2311.09735, KDD 2024). Three details change how you should read it.

• The engine was built by the authors. It was a two-step pipeline with sources supplied directly in context.
• The one "in the wild" test was not live retrieval. The Perplexity test uploaded source text as files, because, in the paper's words, "perplexity.ai does not allow the user to specify source URLs."
• The metric is partly circular. Position-Adjusted Word Count counts a source's words in the answer, including words the intervention itself added.

The paper reports a 30 to 40% band across three methods. It never reports the itemised per-method figures that circulate in vendor decks. The same paper found keyword stuffing lowered its own metric (17.8 against a 19.5 baseline). Read our digest of the original GEO paper for the method in detail. Nothing in it measures whether a page gets discovered.

Later work tightened the method. The results got less flattering for content tactics.

• C-SEO Bench (arXiv 2506.11097, NeurIPS 2025) found that "most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking." It also found gains shrink as more publishers adopt the same tactics. Its "traditional SEO" arm worked better, but it was implemented by moving the document to position 1 in the context window. No practitioner can do that.
• The 252,000-trial factorial (arXiv 2605.25517) found "formatting-only edits have little impact," and that topical relevance and list position drive who gets cited first. It never called a search engine and used two candidate documents per trial on a synthetic corpus. An industry team wrote it.
• Ahrefs' schema study came closest to live conditions, with real engines and matched controls. It found AI Overview citations fell 4.6% after schema was added. Every treated page already had 100+ AI Overview citations, so it cannot say whether schema helps a page enter the set at all.

These designs explain much of why GEO studies disagree.

The gap is not a detail. OpenAI says ChatGPT search rewrites a prompt into one or more targeted queries. Google says AI Overviews and AI Mode "may use different models and techniques." A lab holds all of that still.

What it can tell you

• How models choose among documents they already hold.
• That position and topical relevance dominate citation once you're in the pool.
• That citation-oriented rewrites can backfire, so treat them as risk, not as wins.

What it cannot tell you

• Whether AI crawlers can read your content. Vercel's crawler study found ChatGPT and Claude crawlers fetch JavaScript files but don't execute them, so client-rendered content may be invisible.
• Whether your robots.txt allows the right bot. OAI-SearchBot controls ChatGPT search. GPTBot governs training only.
• Whether a firewall or CDN returns a 403 to a crawler your robots.txt allows.
• Whether Google can show you. Google says a page must be "indexed and eligible to be shown in Google Search with a snippet" to appear as a supporting link in AI Overviews or AI Mode, and "There are no additional technical requirements."
• What the model says with retrieval off. Nothing you publish this quarter reaches that answer.

These are discovery questions, and they live in stages 1 and 2. Lab studies rarely touch them. For more on reading study designs, see correlation vs causation in AI search studies.

Use lab studies to understand why content edits underperform, and to vet a claim. Ask four questions of any number: Was the content already in context? Which engine, and was it real? What does the metric count? Who ran it, and what do they sell?

Use live access checks for the question you actually have: can engines find you? The levers with real evidence behind them remove blocks:

1. Server-render prices, product names and key facts so they appear in raw HTML.
2. Set robots.txt rules per bot, and allow the retrieval bots you want.
3. Request a page as each AI crawler and confirm it gets a 200, not a 403 or a challenge. 99Visibility's AI Crawler Checker runs that request for you.
4. Check for stray noindex, nosnippet, max-snippet:0 or data-nosnippet.

A free GEO audit covers access first, then accuracy.

Use field experiments with care. Otterly's 2,000 AI blogs test ran on live engines: Google deindexed both sites, Bing kept them, and Microsoft Copilot kept citing them. It used one site per arm, so read it as a signal about index dependence, not proof. For the wider picture, see what AI SEO research actually shows. Nobody controls what models say. You do control whether they can reach you.