What Is Generative Engine Optimization, and Where Did GEO Come From?

The paper calls these systems “generative engines”: tools that fetch sources and compose one reply instead of listing ten links. GEO is the craft of being one of the sources they choose.

GEO: where the term comes from and what the research tested

GEO was introduced in “GEO: Generative Engine Optimization”, first posted to arXiv on 16 November 2023. Version 3 followed on 28 June 2024, and the paper was accepted to KDD 2024.

The authors are Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande. The paper lists Princeton University, IIT Delhi and two independent researchers in Seattle.

What they built

They created GEO-bench, a set of 10,000 queries from nine sources across 25 domains. The sources include real search queries, Perplexity’s Discover feed and questions written by GPT-4.

Their test engine fetched the top five Google results for each query and passed them to GPT-3.5 Turbo, which wrote a cited answer. They repeated key tests on Perplexity.ai. That one is a live engine.

What they measured

Ranking positions do not exist inside a written answer, so the team defined new visibility scores. Position-adjusted word count measures how many words of the answer cite your page, weighted towards the top.

Subjective impression asks a model to judge relevance, influence, uniqueness and how likely a user is to click the citation.

The three leaders improved position-adjusted word count by 30 to 40% over the unedited baseline, according to the authors.

The abstract sums it up as visibility gains of “up to 40%”. On Perplexity, the best gains reached 37%.

GEO: how it works inside a generative engine

GEO works on the step where an engine picks which retrieved text to repeat. The paper describes engines that retrieve documents, then use a large model to write a response grounded on those sources.

Real products add a step before retrieval. Google’s July 2026 guide describes query fan-out: the model issues several related searches at once to gather more material.

Google also names retrieval-augmented generation (RAG), which it says is also known as grounding. Its core ranking systems choose which pages get retrieved in the first place.

That sets the order of work. A page must first be retrievable and rank well enough to be fetched.

Only then do GEO edits, such as clearer claims and real figures, change how much of it the model repeats.

One finding cheers smaller sites. When the authors applied Cite Sources, pages ranked fifth on Google gained 115.1% in visibility. On average the top-ranked page lost 30.3% in the same test.

GEO vs SEO vs AEO: what actually differs

GEO, SEO and AEO share most of their methods and differ mainly in what they count as success.

Google and Microsoft, side by side

Google’s July 2026 guide Optimizing your website for generative AI features says optimizing for generative AI search is “still SEO” from Google Search’s view. It asks owners to review third-party GEO advice with care.

Microsoft takes a warmer tone. Its Bing Webmaster blog calls the AI Performance report “an early step toward Generative Engine Optimization (GEO) tooling”.

Digital Magma reads the two statements as compatible. Google says the work is the same. Microsoft adds a scoreboard.

GEO for a business: why it matters

GEO matters to a business because a generated answer often names only a handful of sources, and the rest go unseen.

Ten links leave room for many winners. A written answer may cite three.

Being one of those three puts your brand inside the reply the buyer actually reads, which is why Digital Magma tracks citations as closely as rankings.

Microsoft now reports these citations directly. Its AI Performance dashboard shows how often your pages are cited across Copilot and Bing’s AI summaries, page by page.

The research also suggests a fairer field. In the GEO paper, lower-ranked pages gained the most from simple edits like added sources. That was a lab result, so test it on your own pages.

GEO: what to do (6 steps)

GEO work follows the paper’s evidence and Google’s limits. Digital Magma applies it in this order:

1. Make sure engines can fetch you. Allow Googlebot, Bingbot, OAI-SearchBot and PerplexityBot in robots.txt, and check the CDN. An engine cannot cite what it cannot retrieve.
2. Cite your sources in the text. Name the publisher and link the original. This mirrors the paper’s Cite Sources method.
3. Replace vague claims with real numbers. Use your own data where you have it, with a date and a source. Never invent figures to look citable.
4. Quote named people. A short quote from a credible expert, including your own specialists, mirrors Quotation Addition.
5. Write plainly and skip keyword stuffing. Stuffing did not help in the paper. Google adds that rewriting purely for AI is unnecessary.
6. Track citations every month. Use Bing AI Performance, Search Console and a fixed prompt set, then revise the pages that get fetched but not cited.

How to measure GEO: citations, mentions and share of voice

Measuring GEO needs three numbers, tracked on the same prompts every month.

• Citation rate. The share of prompts where an engine links to your page.
• Mention rate. The share of prompts where your brand is named, with or without a link.
• AI share of voice. Your mentions divided by all mentions of you and named competitors across the same prompts.

For Microsoft surfaces, Bing AI Performance gives first-party data. It counts total citations, cited pages, citations per URL and grounding queries.

For Google, Search Console’s Generative AI performance report counts impressions in AI Overviews and AI Mode. It reached all sites on 31 August 2026.

ChatGPT and Perplexity publish no such report. Fill the gap by hand. Keep 20 to 50 buyer questions, run them monthly, and log every brand named.

GEO myths and research limits

GEO claims often stretch the paper further than its authors did. This Digital Magma entry repeats only what the authors reported.

• “GEO adds 40% visibility.” The figure is a benchmark result on GEO-bench, measured with the authors’ own metrics. It says nothing about traffic or sales on a live site.
• “GEO replaces SEO.” The test engine read only the top five Google results. Pages that never rank never got the chance to be cited.
• “You need llms.txt or chunked pages.” Google’s July 2026 guide says Google Search ignores llms.txt. It also says pages need no chunking.
• “Mentions can be bought.” Google calls inauthentic mentions less helpful than they seem, as spam systems apply.

The authors list limits of their own. Engines change. Real queries drift.

The study did not test what GEO edits do to search rankings, and GPT-3.5 Turbo, a model that has since been superseded several times, ran every experiment.

Digital Magma treats the paper as a guide to what to try first. Your own citation counts decide what stays.

GEO: frequently asked questions

GEO questions that come up in client calls and training sessions.

GEO: related terms, sources and author

GEO links closely to these entries in the knowledge base:

• Answer engine optimization (AEO): winning the single answer passage
• AI SEO: the umbrella for all AI search visibility work
• AI citation: the linked source GEO aims to earn
• AI share of voice: your share of brand mentions across prompts
• Grounding: tying an AI answer to retrieved sources

Sources

Aggarwal et al., arXiv 2311.09735 v3 (KDD 2024); Google Search Central’s July 2026 generative AI guide; Microsoft’s Bing Webmaster blog, 10 February 2026; Google Search Console help.

Digital Magma checked each one on 21 October 2026.