The Princeton GEO paper, explained for Australian business owners

In 2024, researchers from Princeton University, Georgia Tech, IIT Delhi, and the Allen Institute for AI published the first peer-reviewed study on how to get your website cited by AI tools like ChatGPT, Perplexity, and Google's AI Overviews. They tested nine tactics across 10,000 real queries. Three of them — adding quotations, including statistics, and citing credible sources — boosted visibility in AI answers by 28–41%, with quotations as the standout performer. One widely used tactic, keyword stuffing, actively hurt. And the finding almost nobody talks about: the smaller your website, the bigger the gain.

What is the Princeton GEO paper?

GEO: Generative Engine Optimization (Aggarwal et al., 2024) was presented at KDD, one of the world's top data-science conferences. The researchers coined the term "generative engine" for tools like ChatGPT and Perplexity that answer questions by synthesising multiple sources — rather than handing you ten blue links.

Their problem statement is the one every business owner now faces: when an AI gives one answer instead of ten links, who gets seen? The paper built a testing framework called GEO-bench, comprising 10,000 queries across 8 domains, and measured how content changes affect whether — and how prominently — a site gets cited in the AI's answer.

This matters because it is the only foundation in this industry that is not vendor marketing. Every credible AEO tool and agency anchors to this study. It introduced two metrics that have become the standard for measuring AI citations: Position Adjusted Word Count (how many words from your source appear in the AI's answer, weighted by position) and Subjective Impression (overall visibility within the response).

What they tested: 9 tactics, 3 clear winners

The researchers rewrote website content nine different ways and measured the change in AI visibility on a generative engine designed to closely mimic Bing Chat. They then validated the strongest findings on live Perplexity.ai, where top tactics produced a 37% real-world improvement in citations.

Two things stand out from the results.

First, the winners are all credibility signals. AI engines are trained to prefer content that demonstrates its claims — a hard number, a named source, a direct quote. The single biggest lever is showing your evidence, not repeating your keywords.

Second, the classic SEO playbook can backfire. Keyword stuffing — still standard practice at plenty of agencies — scored 17.8 against a baseline of 19.5: worse than no optimisation at all. If your current provider is optimising your site the 2015 way, you may be paying to go backwards in AI search.

"Cite Sources, Quotation Addition, and Statistics Addition emerge as consistently effective strategies, achieving a relative improvement of 30–40% on the Position-Adjusted Word Count metric."— Aggarwal, P., Murahari, V., et al., GEO: Generative Engine Optimization, KDD 2024 (arXiv:2311.09735)

Why this matters for Australian SMBs specifically

In traditional Google search, position five might as well be invisible — the top result captures most clicks. The Princeton team's finding is that this logic does not carry over to AI search. AI engines synthesise a single answer from multiple sources. A Parramatta accountant or a Geelong manufacturer with well-evidenced content can out-cite a national brand on the queries that matter to them.

That window will not stay open forever. In the US, enterprise brands are already paying US$499+ per month for AI visibility tooling. Globally, 70% of organisations say AI search will reshape their strategy within three years, but only 20% have started acting on it (HubSpot, 2026). Most Australian businesses have not started.

The gap is widest right now. The businesses that move first — with evidence-backed content, external citations, and real quotations — will be the ones AI engines reach for when your customers ask for a recommendation.

What to actually do: a 5-step starting plan

1. Audit one money page. Pick the page that wins you business. Count the verifiable claims on it. If it says "trusted by hundreds" instead of "trusted by 214 Australian builders since 2019", you have found your first fix.
2. Cite credible sources. Link your factual claims to the Australian Bureau of Statistics, industry bodies, standards bodies (AS/NZS), or published research. The study recorded a +27.5% lift — solid third of the three winning tactics.
3. Add statistics. Replace every vague claim with a real number — project counts, years in business, percentages, dollar figures. The study recorded a +30.6% lift, second strongest of the three.
4. Add quotations. Customer quotes with names and businesses, or expert commentary. This was the single strongest tactic in the study — a 40.9% lift. AI engines lift well-attributed quotes into answers — make sure they are yours, not your competitor's.
5. Stop keyword stuffing. If a paragraph reads like it was written for a robot, rewrite it for a person. The research says the robot prefers it too — keyword-stuffed content scored 17.8 against a baseline of 19.5.

Do this on one page, wait three to four weeks, then ask ChatGPT and Perplexity the questions your customers ask. Check whether you appear. That is a measurable, evidence-based start — no subscription required.

GEO tactics are quick wins — not a complete strategy

The Princeton study measures content signals in isolation: what you put on individual pages. Apply the three winning tactics and you will see real improvement in 3–8 weeks. That is exactly what the research says, and it is where every business should start.

What the study cannot measure is structural AI visibility — whether AI engines know your brand well enough to recommend you consistently, how clearly they understand your entity (who you are, what you do, where you operate), and whether your schema and citation signals back that up across hundreds of different queries. Those are not content fixes. They are architecture.

Think of GEO tactics as the quick coat of paint that makes your house presentable today. The AEO Traction Stack is building a house AI engines will keep returning to for years.

Frequently asked questions

Sources

• Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). GEO: Generative Engine Optimization. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. arXiv:2311.09735
• HubSpot (2026). Answer engine optimization trends in 2026. blog.hubspot.com
• Semrush (2026). AI SEO statistics. semrush.com/blog/ai-seo-statistics