The 40% GEO Visibility Claim, Examined

The number traces back to one paper: "GEO: Generative Engine Optimization" (Aggarwal et al., arXiv 2311.09735, published at KDD 2024). It tested whether editing a source document changes how much of it a generative engine uses in its answer.

The paper reports a 30 to 40% improvement band across three methods. It does not report the itemised +41%, +40% and +30% figures that circulate in vendor decks and blog posts. Those precise numbers are a retelling, not a quotation.

When a pitch says "GEO gets you up to 40% more visibility", it takes the top of a lab band and drops the conditions the band was measured under. Those conditions are the whole story. If you want the paper section by section, read our digest of the original GEO paper.

What this means for you: the 40% is a real number from a real study. It answers a narrower question than the one it is usually used to answer.

Three design choices limit what the result can tell a marketer.

The engine was built by the authors

The experiments ran on a two-step engine the researchers built themselves. It was not ChatGPT, Perplexity or Google. No live search index was queried to decide which sources the engine would see.

The sources were supplied in context

Every candidate document was handed to the model directly. The question "would a real engine ever retrieve this page?" never came up, because retrieval was skipped. Even the paper's one "in the wild" test on Perplexity uploaded the source text as file attachments, because, in the authors' words, "perplexity.ai does not allow the user to specify source URLs."

The metric counts the words the edit added

The primary metric is Position-Adjusted Word Count. It measures how many words from a source show up in the answer, weighted by where they appear. Many GEO edits work by adding text to the source: a quotation, a statistic, a line of authority boilerplate. When the model then reuses those added words, the metric goes up. Part of the measured gain is the intervention counting itself.

One result in the same paper cuts the other way. Keyword stuffing lowered the score to 17.8 against a 19.5 baseline. So some tactics did worse than doing nothing, even in the friendliest possible setup.

The paper does show something real: once a document is inside a model's context, edits to it can change how much of it the answer reuses. That is a Stage 3 (generation) finding. Most buyers want a Stage 1 or 2 outcome: being fetched and retrieved for real questions.

The gap between the two columns is why lab gains rarely transfer. Our comparison of lab GEO studies and live engines goes through each row.

The next move: treat any number from this paper as evidence about reuse of already-retrieved text, not about traffic, citations or rankings.

Better-designed studies have since tested the same family of tactics.

• C-SEO Bench (arXiv 2506.11097, NeurIPS 2025) concluded that "most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking." It also found that gains shrink as more sites adopt the tactics, describing the problem as "congested and zero-sum."
• The same benchmark found that moving a document to position 1 in the model's context mattered far more. That arm placed the document there directly. No ranking was earned, so it is not a lever any publisher can pull.
• A 252,000-trial factorial study (arXiv 2605.25517) found "formatting-only edits have little impact" and that topical relevance and list position were the biggest drivers of being cited first. It also never queried a live search engine.
• A July 2026 critical survey of 45 GEO studies (arXiv 2607.14035) summarised the field: "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."

Read together, the evidence says rewriting for citation is at best neutral and sometimes harmful. Our page on whether rewriting content for AI can hurt visibility covers that risk, and the C-SEO Bench digest covers the benchmark's design.

It helps to split an AI answer into four stages.

1. Parametric. What the model knows with no web access, frozen at a training cutoff. Nobody changes this in a quarter.
2. Access. Whether a crawler may fetch and store your page. You and your CDN control this.
3. Retrieval. Whether your page enters the model's context for a given query. The engine controls this, and the real queries are usually rewritten and hidden from you.
4. Generation. Which retrieved documents the model uses and cites. The model controls this.

The 40% claim lives entirely in stage 4, and only for documents that already passed stages 2 and 3. If a crawler can't reach your page, no wording change matters.

Stage 2 is where the strongest evidence sits, and it is mostly about removing blocks. Server-side rendering, per-bot robots.txt permission, firewall reachability, index eligibility and snippet controls all have documented mechanisms. No markup, phrasing or page structure has credible evidence behind it as a way in.

One on-page habit does have a small positive signal: stating a specific fact (a price, a number, a one-line definition) on a page that can be fetched. That gives an engine a sentence it can quote. It is not a promise of citation.

Work the stages you control, in order. Each step removes a possible block.

1. Check robots.txt bot by bot. OpenAI's crawler page says each setting "is independent of the others". GPTBot governs training use only. OAI-SearchBot governs ChatGPT search. Look for PerplexityBot and Claude-SearchBot too. A blanket AI block often removes search eligibility by accident.
2. Check the firewall and CDN. A CDN can return 403 to a crawler that robots.txt allows, and nothing in your dashboards will flag it. From 15 September 2026, Cloudflare blocks Training and Agent crawlers by default on ad-displaying pages for new domains and existing free-tier customers. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.
3. Check that key content is in the raw HTML. Vercel's crawler study found ChatGPT and Claude crawlers fetch JavaScript files but don't execute them. If your pricing or product description only appears after JavaScript runs, those crawlers can't read it. Googlebot renders, so this check matters less for Google.
4. Check snippet controls. A stray noindex, nosnippet, max-snippet:0 or data-nosnippet from a template keeps content out of AI answers.
5. State your key facts plainly. Put your price, core numbers and a one-line definition of what you do in fetchable text. Make sure the same fact reads the same on every page.
6. Take a baseline, per engine. The free checker on this site gives a snapshot of whether AI answers name you for up to three buyer questions. Record it before you change anything.

If your roadmap already holds rewriting tasks justified by the 40% figure, the playbook for retiring unsupported GEO tasks shows how to clear them out without losing work that does have support.

• Treating a lab band as a forecast. The 30 to 40% describes reuse of pre-loaded text in one research engine. It says nothing about your traffic or citations.
• Rewriting the whole site. Citation-oriented rewrites can impair retrieval, according to the 2026 survey. A site-wide rewrite risks the pages that already work.
• Stacking tactics. Adding quotation blocks, injected statistics and "according to authoritative sources" boilerplate is the tactic family later studies found neutral or harmful. Treat it as a risk, not a win.
• Skipping access checks. Teams polish copy while a firewall returns 403 to OAI-SearchBot. Access comes first.
• Trusting a single blended score. Engines retrieve differently and decide per query whether to search at all. Measure each platform separately.
• Taking case studies at face value. Ask whether the study controlled for brand size, retrieval state and which prompts were tested. Our guide to trusting GEO case studies lists the questions, and the GEO claims red flags page covers other numbers that fail the same test.