Generative Engine Optimization (GEO): What the Research Found and How to Get Cited

Generative engine optimization (GEO) is the practice of editing a page so that AI search tools are more likely to use it and cite it in their answers. The most-cited study of it found that adding sources, quotations and statistics raised a page's share of the answer, while repeating keywords did not. That is a narrower claim than most GEO articles make, so this post goes through what the paper actually measured, where it stops being useful, and what to change on a page this week.

I read the paper itself (arXiv:2311.09735, "GEO: Generative Engine Optimization", Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, published at KDD 2024) and its tables on 2026-10-08. Everything attributed to it below comes from that text. Claims about Google come from Google's own documentation, read the same day.

What did the GEO paper actually test?

The authors built a small stand-in for an AI search engine and measured how much of its answer a source received.

• The engine: for each query, fetch the top 5 Google results, give the full text to gpt-3.5-turbo, and have it write a cited answer. Five answers were sampled per query at temperature 0.7.
• The benchmark (GEO-bench): 10,000 queries across 25 domains, split 8,000 / 1,000 / 1,000 for train, validation and test. About 80% were informational, with 10% each transactional and navigational.
• The intervention: for each query, one of the five sources was picked at random and rewritten by an LLM according to one of nine methods. The other four were left alone.
• The outcome: "impression", meaning how much of the answer is attributed to your source and how early it is cited. Two versions: Position-Adjusted Word Count (word share weighted by citation position) and Subjective Impression (an LLM judge scoring seven aspects such as relevance and likelihood of a click).

So the question was "if this one source is rewritten this way, does it take a bigger share of the answer?" It was not "does this drive traffic" or "does this rank".

Which edits helped and which did not?

Here are the overall Position-Adjusted Word Count scores from the paper's Table 1 (higher is better, unmodified baseline is 19.3):

The authors summarise the best three (cite sources, quotations, statistics) as a 30 to 40% relative improvement on that metric and 15 to 30% on Subjective Impression. The "up to 40%" in the abstract is that range's top end.

Two details that most summaries drop:

1. Keyword stuffing is not cleanly negative everywhere. On the Subjective Impression average in Table 1 it scored 20.2 against a 19.3 baseline. On the Perplexity.ai run (Table 5) it scored 21.9 against 24.1 on Position-Adjusted Word Count, but 28.1 against 24.7 on Subjective Impression. The consistent finding is "no reliable gain", and the word-count metric, which is the more objective one, went down in both.
2. Tone did little. The Authoritative rewrite, the "sound more confident" edit that most marketing copy already does, showed no significant improvement in the text's own words. Presentation and verifiable substance moved the numbers, persuasion did not.

Does it matter which domain you write in?

A bit. Table 3 lists the query categories where each method did best. Cite Sources led on "Statement" and "Facts" queries, Statistics Addition on "Law & Government", "Debate" and "Opinion", and Quotation Addition on "People & Society", "Explanation" and "History". The authors' reading is that quotes suit narrative and historical topics and statistics suit policy and opinion topics. Those are their interpretations of one benchmark, so use them as a prompt to test, not a rule.

Combining methods helped on a 200-query subset: Fluency Optimization plus Statistics Addition was the best pair, beating any single method by more than 5.5%.

Where does the evidence stop?

This is the part to read before you rebuild your site around the paper.

• Old, small engine. The main results use a GPT-3.5-turbo pipeline fed five sources. Production engines retrieve differently, use newer models and may filter or verify sources. The paper's own limitations section says methods "may need to adapt over time as GEs evolve".
• One real-world check. The Perplexity.ai experiment (Table 5) is the only test on a deployed product. It points the same way for quotations (29.1 vs 24.1) and statistics (26.2 vs 24.1) on word count. Cite Sources is mixed there: 26.8 on word count, but 19.0 against a 24.7 baseline on the LLM-judged Subjective Impression. It is also one run from the paper's 2023 to 2024 window, and I could not check how closely today's Perplexity resembles it.
• A visibility proxy, not a business outcome. Impression is share of the answer. Clicks, brand mentions and conversions were not measured.
• LLM-written edits and LLM-judged scores. The Subjective Impression metric is scored by a model with calibration adjustments, which the authors themselves note.
• Ranking interacts with the result. When all five sources were optimized together (Table 2), the Cite Sources method raised the fifth-ranked source's visibility by 115.1% and reduced the top-ranked source's by 30.3%. Good news if you rank low today, a warning if you rank first: once everyone adds citations, the gain shrinks.
• The example in Table 4 is a trap. One illustrative rewrite adds "a staggering 70% increase" to a sentence with no source shown. Do not copy that. A fabricated number helps no one once a model or reader checks it.

My read: the direction (verifiable specifics beat keyword repetition) is plausible and consistent with the tables. The effect sizes are not transferable.

How does GEO differ from classic SEO?

Google's page on AI features (last updated 10 December 2025 at the time I read it) states there are "no additional requirements" for AI Overviews or AI Mode and no "special optimizations necessary". That cuts against the idea that GEO is a separate technical discipline for Google. It is silent about ChatGPT, Claude and Perplexity, which publish their own crawler rules (see the robots.txt and llms.txt guide).

For the content-production side of SEO, the AI SEO content strategy prompts post covers briefs and drafting. This one is about the edits that make an existing page easier to quote.

What should I change on a page this week?

Pick your ten most useful pages, not the whole site. For each, run this checklist.

1. Put the answer in the first two sentences. An engine pulling a passage should find a complete, quotable claim without scrolling. (This is standard writing advice, not a paper finding.)
2. Attach a source to every factual claim a reader might doubt. Link the primary source: a vendor doc, a regulator, a paper. This is the Cite Sources method with the part the paper's example skipped, a real source.
3. Replace vague adjectives with a number you can defend. "Much faster" becomes "dropped from X ms to Y ms in my test on DATE", only if you really measured it, and say how you measured. If you cannot defend the number, delete the sentence.
4. Quote a named expert or document, briefly and accurately. Use real quotes with attribution. Short is better, and check the wording against the source.
5. Fix fluency. Read the page aloud. Cut padding, repeated phrases and hedges. The paper's Fluency Optimization was the third strongest single edit.
6. Delete keyword repetition. If a phrase appears in every other sentence, it is not helping.
7. Keep key facts in text. Not in images, accordions that need a click, or PDFs. Google's AI features page lists "important content in text form" as standard practice.
8. Check that bots can fetch the page at all. A blocked retrieval bot cannot cite you. See the crawler allow/block decision table.

A prompt for step 2 and 3, labelled an untested template:

Here is a page from my site. List every factual claim in it that a reader
could reasonably doubt. For each: quote the sentence, say what kind of
source would settle it (vendor doc, regulator, paper, my own measurement),
and mark it "has source", "needs source" or "cannot be sourced, cut it".
Do not invent sources or numbers. If you do not know a source, say so.

Use the output as an edit list. Look up each source yourself before you link it.

How do I know whether any of this worked?

Do not rely on the paper's percentages or your instinct. Measure before and after with the same questions, on the same engines, with enough repeats to see past randomness. The AI citation test protocol gives a logging sheet and scoring method for that. Also keep a plain record of what you changed and when, so a rise in citations can be tied to an edit rather than a model update.

If you are checking AI answers about your own topic, the fact-checking checklist for AI answers covers how to verify what an engine says about you.

What I would not do

• Rewrite whole pages with an LLM and publish them unread. The paper's edits were machine-made in a lab; yours will carry your name.
• Pay for a "GEO score" that cannot say how it was validated against a real engine.
• Chase one engine's quirks. They change faster than a content calendar.
• Treat GEO as a replacement for being the original source. Pages that publish original data, tests or documentation give engines something to cite that copies do not.

Related reading

• AEO checklist: what Google and OpenAI actually say
• Track AI assistant referral traffic in GA4
• Answer-first rewrite prompts for featured snippets and AI answers