The primary literature this field rests on, read properly and summarised honestly. Most GEO marketing cites one 2023 paper it has misread. This is the rest of the picture — including the papers that argue against us.
The paper that coined the term. It defines the “generative engine”, frames content creators as a third stakeholder with “little to no control over when and how their content is displayed”, introduces the GEO-bench benchmark, and reports visibility gains of up to 40% for tactics such as adding quotations, statistics, and authoritative phrasing.
Our reading
Cite it for vocabulary — this is where GEO as a term comes from, and that's a historical fact that won't decay. Do not cite the 40% as a promise.
What it doesn't establish
The 40% is a redistributive share metric across a fixed set of five sources that sums to one — not an absolute lift; the paper itself notes it “does not mean that 40% more readers will click”. Measured on GPT-3.5-turbo in a custom two-step harness, never on production ChatGPT/Gemini/AI Overviews, and conditional on the source already being in the model's context. No 2026 replication. The abstract concedes efficacy “varies across domains”.
Haritz Puerto, Martin Gubri, Tommaso Green, Seong Joon Oh, et al. · NeurIPS 2025 Datasets & Benchmarks (arXiv 2025-06-06)
Re-tests seven of the methods from the original GEO paper across multiple domains and finds that “most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking”. It reports that traditional SEO — getting the source retrieved and placed high in the model's context — is “significantly more effective”, and that gains are “congested and zero-sum”: as more competitors adopt a tactic, the advantage decays toward zero.
Our reading
The most important paper in this field for a buyer to know about, and the one no agency will send you. It reframes GEO as subordinate to retrieval: content rewriting is hygiene, not a moat. It is why we sell measurement and crawler access rather than promising that adding statistics to your pages will win citations.
What it doesn't establish
Tested on GPT-4o-mini and Claude-3.5-Haiku, English-only, white-hat tactics only, in a simulated RAG setup rather than live ChatGPT/Perplexity/Gemini. Its zero-sum finding is partly baked into the design (a fixed candidate pool and a ranking metric) and does not measure whether content can help you *enter* the pool. The authors conclude C-SEO “will not replace SEO, but will complement it” — so “GEO is worthless” overreads the paper; “GEO is subordinate to retrieval” is what it supports.
A survey of the GEO literature to date, assessing which techniques have durable, replicated evidence behind them.
Our reading
Useful as a map of the field's evidence base rather than for any single result. Its headline conclusion — that no reviewed technique shows a stable, longitudinal, cross-platform causal effect — matches what we found independently and is the honest state of this category in 2026.
What it doesn't establish
A survey, not new experimental evidence, and recent enough that it has not yet been widely engaged with. Treat as a literature map.
Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, et al. · NeurIPS 2020 (arXiv 2020-05-22)
Introduces Retrieval-Augmented Generation: combining a parametric language model with a non-parametric retrieval index, so answers are generated from documents fetched at query time rather than from training weights alone.
Our reading
The mechanism underneath almost every cited AI answer, and the reason GEO is fundamentally a retrieval problem. If your page is never fetched into the context window, nothing written on it can influence the answer. Read this before believing anyone who tells you AI visibility is about phrasing.
Google's own guidance on how its generative features relate to Search. It states that AI Overviews and AI Mode are rooted in core Search ranking and quality systems, that no structured data type is required for or specific to AI features, and that it does not use llms.txt.
Our reading
Adverse to GEO agency positioning, and worth reading precisely for that. Google's position is that optimizing for its AI features “is still SEO”. Our honest read: for Google's own surfaces that's largely right; for ChatGPT, Claude and Perplexity — which don't run on Google's ranking at all — it isn't. That gap is the part worth paying for.
What it doesn't establish
It is Google's account of Google. It says nothing about how non-Google engines retrieve or cite, and a platform describing its own ranking has obvious incentives.
Documents OpenAI's separate user agents — GPTBot (training), OAI-SearchBot (search indexing) and ChatGPT-User (live retrieval on a user's behalf) — and the robots.txt and IP-range controls that govern each.
Our reading
The single most actionable document in this list. Because the agents are separately blockable and do different jobs, a blanket “block AI bots” rule silently removes you from the citable surface — you stop being quotable in live answers, not merely excluded from training. It never mentions reading llms.txt from third-party sites.
What it doesn't establish
Compliance is voluntary and honour-based; Cloudflare has published findings of at least one AI company circumventing robots.txt.
Verified 2026-07 against the arXiv API and each vendor's live documentation. Missing something important — especially something that contradicts us? Email contact@wordofgpt.com and we'll add it.
Measurement built on the evidence, not the marketing
We'll run your domain through our tracker and send a free one-page snapshot — with the methodology attached.