Generative engine optimization
Generative Engine Optimization (GEO): What the Evidence Supports in 2026
GEO is the practice of improving the likelihood that a brand or source is surfaced, used, cited or recommended by generative answer systems. The field is real. The evidence is narrower than the marketing around it, and the gap between those two sentences is where budgets get wasted.
What the original GEO research actually showed
The term was formalised in a 2023 paper that gave the field a measurable research frame — and produced the most misquoted number in the category.
The term was formalised in “GEO: Generative Engine Optimization” by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, later accepted at KDD 2024.
The paper reports visibility gains of up to roughly 40% — and the qualifier is not optional: that result was measured inside GEO-bench, the paper's own controlled benchmark of queries and relevant web sources, by modifying content that had already been selected as a source.
What that does not establish
It does not show that adding a tactic to a website produces organic retrieval, real-world inclusion in ChatGPT, durable cross-engine visibility, or downstream revenue. Anyone quoting “GEO gives you 40% more visibility” without the benchmark qualifier is quoting a controlled experiment as if it were a market outcome.
What later research says
A 2026 critical survey reviewed 45 studies and described generative visibility not as a ranking task but as a stochastic, partially observable pipeline.
“Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023–2026)” describes a pipeline spanning search activation, crawling and indexing, retrieval, reranking and context allocation, citation, prominence, factual absorption, fidelity and user behaviour.
Survey conclusion, not Google guidance
Its central caution: no single optimisation trick has been shown to produce a stable, longitudinal, cross-platform causal improvement in organic discoverability and downstream outcomes. That does not mean GEO is useless — it means the job looks like SEO rather than magic. Improve what can be improved, measure the outcome, and treat black-box claims carefully.
What Google documents
For Google Search’s own generative features, the guidance is unusually clear: SEO remains the foundation.
Google's AI features and your website documentation states that “the best practices for SEO remain relevant”, that there are “no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary”, and that a page must be indexed and eligible to appear in Google Search. It also states plainly that “You don't need to create new machine readable files, AI text files, or markup to appear in these features”.
Myth 1 — Google needs an llms.txt file
It does not, per the sentence quoted directly above. Publish one for other reasons if you like; do not sell it as a Google requirement.
Myth 2 — you need a page for every fan-out query
Generative systems can fan out into related queries. That is not an instruction to create one page per variation. Google's spam policies define scaled content abuse as “when many pages are generated for the primary purpose of manipulating search rankings and not helping users”. Build one strong page when one strong page is enough.
Myth 3 — there is a universal generative ranking formula
Different systems use different retrieval methods, sources, models, grounding behaviour, context and freshness mechanisms, and the same engine can vary between runs. There is no defensible equation that predicts “GEO rank”. Measure each target environment separately — which is the whole argument for per-engine reporting on AI search visibility.
GEO and SEO overlap at six layers
SEO is the foundation; GEO is an additional answer-layer objective and measurement discipline.
| Layer | Why both disciplines care |
|---|---|
| Crawl and index | A search-grounded system cannot retrieve a page it cannot access. |
| Relevance | The content still has to address the question being asked. |
| Quality and trust | Evidence, authority and original value matter to both. |
| Entity clarity | A system needs consistent information to understand what the brand and product are. |
| External evidence | Third-party sources influence how a brand is represented in both surfaces. |
| Measurement | This is where they diverge — rank tracking cannot observe a generated answer. |
GEO and AEO overlap so substantially that running them as two disconnected programmes is a budgeting mistake rather than a strategy.
The multi-stage funnel — and why it prevents bad diagnosis
A failure at stage 6 should not be diagnosed as a stage 1 crawl problem, which is exactly what happens when the only tool in the room is a technical audit.
Stage 1 · Retrieval
Can the system find a page of yours at all? If not, nothing downstream can be fixed with better copy.
Stage 2 · Selection
Of what was retrieved, is yours chosen as a source for this answer?
Stage 3 · Use
Is your page actually cited in the output, or read and discarded?
Stage 4 · Framing
When you appear, are you recommended — or named while a competitor is put forward?
Diagnosing the wrong stage is the expensive mistake. Rewriting a page for persuasiveness cannot fix a retrieval problem, and adding more content cannot fix a framing problem. Each stage fails for different reasons and is measured differently.
- Discoverability — can the system or index find the information?
- Retrieval — is the source retrieved for this question?
- Selection and context — does it get enough attention allocated to it?
- Use — is the information actually used in the answer?
- Citation — is the source credited or linked?
- Recommendation — does the brand make the shortlist?
- Business outcome — does the answer lead to qualified visits, signups or revenue?
Inference, not measurement
Being retrievable and being chosen are different failures with different fixes. Most “GEO audits” only examine stages 1 and 2 because those are the only ones a crawler can see.
A GEO strategy you can actually check
Nine steps, each of which produces something a client can inspect.
- Define the buyer questions. Category selection, alternatives, comparisons, use cases, strategic capabilities.
- Measure a repeated baseline. Across your target engines, storing the answers.
- Identify the high-value losses. Prioritising commercial importance separately from fixability, so a hard-to-fix high-value gap stays visible.
- Inspect winning evidence. Pricing model, capability, proof, independent comparison, local data, expert sources — observed answer themes, not hidden ranking factors.
- Compare owned evidence. Is it present, partial, not found in the inspected set, or unknown?
- Verify product truth. Before telling anyone to publish a claim.
- Evaluate third-party sources. Separating source relevance from source addressability.
- Implement one defensible action. The smallest change the evidence supports.
- Remeasure. Under a compatible protocol, and record the measured movement without claiming it was caused.
What kind of content has GEO value
The strongest candidates add evidence the web can reuse. Commodity summaries are easy to reproduce, which is exactly why they are easy to replace.
Original data
Benchmarks, surveys, experiments, product datasets.
First-hand reviews
Specifics that cannot be generated from summaries.
Clear product facts
Pricing, capabilities, limitations, eligibility.
Decision content
Who a product is for, and when not to choose it.
Expert explanations
Precise, sourced, with stated limitations.
Local or vertical evidence
What a generic national page does not contain.
If your page is the original source of a benchmark, a study, a dataset or a clear methodology, other publishers may cite it — and answer systems can then meet the same evidence through several independent paths. That is more defensible than rewriting the definition everyone else has already written.
GEO is not the same job in every business
A system that forces all four of these into the same score is optimising for a chart rather than a decision.
SaaS
Recommendation and positioning are usually central. Which claim do the engines associate with you?
Ecommerce
Product data, merchant information, reviews, availability and shopping surfaces carry more weight.
Local
Location-specific evidence and consistent business and entity information matter most.
Publishers
Citation and source visibility can matter far more than vendor recommendation.
What not to use as a GEO north star
A better north star: how many commercially important AI answer gaps can we identify, act on and remeasure?
Avoid organising the company around
- the number of prompts tracked
- the number of GEO pages published
- a single proprietary score
- the number of schema types added
- the length of an llms.txt file
- raw citation count without question importance
Can GEO drive revenue?
It can contribute. The cleanest evidence chain runs answer visibility → AI referral sessions → first-party conversions and revenue.
Attribution is hard, and two shortcuts are worth refusing explicitly: do not convert a visibility gap directly into “this costs $X” without a real economic model, and do not claim a page edit caused revenue because visibility later changed. Build the chain and label each inference for what it is.
What GEO cannot guarantee
Not a ranking, not a citation, not a recommendation, not permanent model behaviour, and not a causal explanation of a black-box answer. Google states directly that no one can guarantee a #1 ranking and that third-party tools have no access to its internal ranking data. That constraint applies to every vendor in this category, ours included.
Common questions
Who coined generative engine optimization?
The term was formalised in the 2023 research paper “GEO: Generative Engine Optimization” by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, later accepted at KDD 2024.
Does GEO replace SEO?
No. For Google’s AI features, Google states SEO best practices remain relevant and that a page must be indexed and eligible to appear in Google Search. GEO is an additional answer-layer objective and measurement discipline, not a replacement.
Does llms.txt improve Google AI visibility?
Google states you do not need to create new machine-readable files, AI text files or markup to appear in its AI features, and that no special structured data is required either.
Does GEO work the same way for SaaS, ecommerce and publishers?
No. For SaaS, recommendation and positioning are usually what matter. For ecommerce it is product data, merchant information, reviews and availability. For local businesses it is location-specific evidence and consistent entity information. For publishers, citation and source visibility often matter more than vendor recommendation. A system that forces all four into one score is optimising for a chart rather than a decision.
Can GEO guarantee ChatGPT recommendations?
No. AI answers are stochastic and the pipeline behind them is multi-stage and only partly observable. GEO can improve evidence and measure outcomes; it cannot guarantee a recommendation, and neither can any tool selling one.
What did the original GEO paper actually prove?
It reported visibility gains of up to about 40% for some content-optimisation strategies inside GEO-bench, its own controlled benchmark of queries and relevant web sources. That is a demonstration in an experimental setting, not evidence of guaranteed organic discoverability or durable business outcomes in the open web.
Methodology
This page separates three classes of claim and labels each one. Experimental results are attributed to the paper that produced them together with the setting they were measured in. Survey conclusions are attributed to the survey and are not presented as vendor guidance. Platform guidance is quoted from the platform's own documentation rather than from commentary about it.
Where a claim could not be traced to one of those three, it is not on the page.
Sources
Official documentation and first-party measurement, in that order. Third-party facts carry the date we last read them from the vendor's own page — verify before you act on a number.
- Aggarwal et al., “GEO: Generative Engine Optimization” (2023) Origin of the term; the “up to 40%” figure and the GEO-bench controlled setting it applies to.Checked 2026-08-23
- Critical survey of generative engine optimization, 2023–2026 (arXiv 2607.14035) 45 studies reviewed; the multi-stage partially observable pipeline; the absence of a stable cross-platform causal result.Checked 2026-08-23
- Google Search Central — AI features and your website SEO fundamentals remain relevant; no AI text files or special structured data required; indexing prerequisite. Page last updated 2025-12-10.Checked 2026-08-23
- Google Search Central — spam policies Scaled content abuse, quoted directly.Checked 2026-08-23
- Google Search Central — do you need an SEO? No guaranteed rankings; third-party tools have no access to internal ranking data.Checked 2026-08-23
How we measure, and how this page is sourced
What we actually measure
Every result comes from the official OpenAI, Google Gemini
and Perplexity APIs with web search or grounding enabled. In our data those
surfaces are recorded as openai_api_web_search,
gemini_api_google_search and perplexity_api_sonar, and reports name
them explicitly.
That is not the same thing as a logged-in person using the ChatGPT app. These are official interfaces with web search on, but they are measurement surfaces rather than consumer sessions: no personalisation, no memory, no app-only features. Results can differ from what any individual user sees, and we do not claim the two are identical. We would rather put that on the pricing page than in the small print.
How this page is sourced
- Vendor pricing and features come from each vendor's own public pricing page, with the date we checked it. We do not copy figures out of other people's comparison posts.
- Google-specific claims come from Google Search Central, not from SEO commentary about Google.
- Research claims name the paper and separate what was demonstrated in a benchmark from what holds in the open web.
- First-hand numbers come from our own measurement runs. Where the brand measured is founder-owned, the page says so in the same sentence as the finding.
This page is maintained by the person who built the measurement engine. There is no editorial team and no external reviewer behind it, and we are not going to invent one for a byline. How the measurements are collected.