Generative engine optimisation (GEO): getting cited by AI

Generative engine optimisation (GEO) is the work of making your pages the ones AI answer engines quote and link to. It is for any brand whose customers now ask ChatGPT, Perplexity, Copilot or Google's AI Overviews before they search. It matters because an AI answer names only a handful of sources, and being one of them is the new first page.

This is the deep dive. For how GEO relates to SEO, AEO, LLMO and the rest, start with my engine optimisation explainer. Here I go into what the research actually tested, how each engine picks sources, and how to measure the result.

What is generative engine optimisation, and where does the term come from?

Generative engine optimisation is a term from a research paper, not from an agency pitch. In November 2023, researchers from Princeton and IIT Delhi published GEO: Generative Engine Optimization (Aggarwal et al.), later accepted at KDD 2024. They defined a generative engine as a system that retrieves sources and uses a large language model to write one answer, and GEO as helping content creators be visible inside that answer.

The paper's real contribution is a way to measure that visibility. It scored a source by how many words of the answer were drawn from it, weighted by how early it was cited ("position-adjusted word count"), plus a second, model-judged score for how influential the citation felt ("subjective impression").

What did the GEO paper actually test and find?

The short answer: adding evidence worked, adding keywords did not. According to the full paper, the team built GEO-bench, 10,000 queries from nine sources across 25 domains, each paired with the top five Google results. They rewrote one source per query using nine methods and measured the change.

On the main measure, against a baseline score of 19.5:

Method	Score	Change
Quotation addition	27.8	about +42%
Statistics addition	25.9	about +33%
Fluency optimisation	25.1	about +29%
Citing sources	24.9	about +28%
Keyword stuffing	17.8	about −9%

A smaller test on the live Perplexity engine, with 200 queries, pointed the same way. Quotation addition raised the main score from 24.1 to 29.1, and keyword stuffing pulled it down to 21.9.

Where the GEO research falls short

The results are useful, but four limits matter before you build a strategy on them:

• The gains went to underdogs. Citing sources lifted visibility by 115.1% for pages ranked fifth, but cut it by 30.3% for pages ranked first. Quotations and statistics showed the same pattern. If you already rank first, the paper's tactics may not help you.
• The engine was simulated. The main tests used gpt-3.5-turbo summarising five sources in 2023. Today's engines fan out across many searches and much larger source pools.
• What works varies by subject. Statistics did best for law and government queries; quotations did best for people and history. The authors say methods must be tuned per domain.
• The authors flag their own limits. They tested only two engines, could not measure effects on search rankings, and note that their automated topic labels "can be noisy".

AI engines quote evidence, not adjectives. Give them a number, a source and a named expert, and you give them a reason to cite you.

Why does GEO matter now?

Because the audience for AI answers in India is now mass-market, and the measurement finally exists. In July 2025, Airtel offered a free year of Perplexity Pro to its 360 million customers, as Business Today reported. Google has also given site owners a way to see their own share of AI answers: on 3 June 2026 it launched generative AI performance reports in Search Console, and by 31 August 2026 had rolled them out to all websites worldwide.

When something can be measured, it gets budget. GEO has moved from theory to a line in the monthly report.

How do ChatGPT, Perplexity, Google and Copilot choose sources?

Each engine retrieves pages with its own crawler or index, then picks passages that answer the question with the strongest evidence. What you can control is access, and each company documents its own crawlers:

Engine	Crawler that decides search visibility	What the training or user bots do
ChatGPT search	OAI-SearchBot, "used to surface websites in search results in ChatGPT's search features"	GPTBot is for model training; ChatGPT-User fetches pages when a user asks
Perplexity	PerplexityBot, which surfaces and links sites and "is not used to crawl content for AI foundation models"	Perplexity-User fetches on a user's request and "generally ignores robots.txt rules"
Google AI Overviews and AI Mode	Googlebot, the normal Search crawler	Google-Extended controls Gemini training and grounding, and does not affect Search
Microsoft Copilot and Bing	Bingbot, the normal Bing index	Bing honours robots.txt and other supported controls

The sources are OpenAI's crawler page, which adds that robots.txt changes take about 24 hours to apply; Perplexity's crawler page; and Google's common crawlers page, which says Google-Extended "does not impact a site's inclusion in Google Search nor is it used as a ranking signal".

The practical lesson: you can block training bots (GPTBot, Google-Extended) and still be cited in ChatGPT search and AI Overviews. Many sites block everything by accident through a CDN's "block AI bots" switch.

Beyond access, the engines say what they reward. Google's AI optimisation guide (updated 10 July 2026) favours "non-commodity content" with "unique expert or experienced takes", and warns that making pages for every AI sub-query to manipulate answers breaks its scaled content abuse policy. Microsoft's Bing team wrote in May 2026 that "not all indexed content carries equal evidentiary weight" and "stale facts can directly produce wrong answers", so freshness and provenance count.

Is llms.txt worth adding?

llms.txt is a proposal, not a standard, and no major answer engine has said it uses the file. It was published on 3 September 2024 by Jeremy Howard as "a proposal to standardise on using an /llms.txt file": a Markdown file at your site root that lists your key pages for AI agents.

The evidence against relying on it is clear. Google's AI optimisation guide says Google Search "itself doesn't use" AI text files. In June 2025, Google's John Mueller wrote that "no AI system currently uses llms.txt". The picture is not entirely one-sided: Search Engine Journal reported in May 2026 that Chrome's Lighthouse 13.3 added an experimental llms.txt audit for agent readiness, separate from SEO. My advice: add one only after everything below is done, and expect nothing from it.

How do you measure AI visibility?

Combine platform reports, analytics and a fixed prompt test, because no single source shows the whole picture.

1. Google Search Console. The generative AI performance report shows impressions, pages, countries, devices and dates for AI Overviews, AI Mode and Discover, per Google's announcement. It does not list clicks or queries.
2. Bing Webmaster Tools. The AI Performance report (public preview since February 2026) shows total citations in Copilot and Bing's AI summaries, cited pages and sample "grounding queries".
3. Analytics. OpenAI's publisher FAQ says ChatGPT adds utm_source=chatgpt.com to referral links. Build a channel group for AI referrers, including perplexity.ai and copilot.microsoft.com.
4. A monthly prompt set. Write 20 to 30 questions your buyers ask. Each month, ask ChatGPT, Perplexity, Gemini and Copilot, and log whether you are named, linked, or replaced by a competitor. Answers vary run to run, so look at trends over three months, not single results.

How do you do GEO? A practical checklist

Run this in order. The first two weeks fix access; the rest is content work in monthly cycles.

Week one: access

1. Check robots.txt allows Googlebot, Bingbot, OAI-SearchBot and PerplexityBot. Decide separately on GPTBot and Google-Extended.
2. Check your CDN or firewall is not blocking AI crawlers by default.
3. Confirm key answers are in HTML text, not images, PDFs or closed tabs.

Weeks two to four: your ten most important pages

4. Open each section with a one- or two-sentence answer to its heading.
5. Add at least one dated statistic with a linked source per key claim.
6. Add a quotation from a named expert, ideally your own specialist, with their role.
7. Add a byline, an updated date and a line on how the page was made.
8. Add something non-commodity: your own data, a worked example, or an opinion others cannot copy.

From month two: the loop

9. Refresh facts quarterly, and ping Bing via IndexNow when pages change.
10. Review Search Console, Bing and your prompt log monthly; rewrite pages that are retrieved but not cited.

Where does GEO go wrong?

It goes wrong when teams treat it as a trick. Common failures: keyword-stuffing pages "for AI", which the paper found hurts; mass-producing pages for every prompt variant, which Google classes as spam; paying for fake "brand mentions", which Google's guide says help less than people think; and blocking every AI bot through the CDN, then wondering why ChatGPT never cites you.

In India there is a particular gap. Many brands outside the metros publish their best answers as WhatsApp forwards, Instagram carousels or scanned brochures. AI engines cannot cite what they cannot read, so put the answer on a web page, in the languages your customers ask in.

Who should own GEO?

The SEO or content lead should own GEO end to end, because it is a content and evidence job on SEO foundations. Developers own crawler access and speed. Subject experts own the statistics and quotations. PR and reputation teams own what third-party sites, reviews and forums say about you, because AI engines read those too. A small business does not need a GEO agency; a marketer with the checklist above can do most of it.

My view

My view is that GEO rewards the habits good marketers should already have: say something specific, prove it, and keep it current. In my years running online reputation management across Meta, Google, Glassdoor and MouthShut, I learnt that what others write about a brand often outweighs what the brand writes about itself. AI answers make that truer, because they blend your page with every review and forum thread they can find.

The trade-off is real. The GEO paper's own data shows that pages already ranked first can lose visibility when they add the same tactics, and a citation does not guarantee a click. Evidence-rich pages also take longer to write and must be kept fresh. I would still do it, but judge the work by enquiries and sales, not by citation counts alone.

Start with one prompt and one page

This week, ask ChatGPT, Perplexity and Copilot the single question your best customers ask before buying. Note who gets cited. Then take your page on that topic, put the answer in the first sentence, add one sourced statistic and one named quotation, and check that OAI-SearchBot and PerplexityBot can reach it. Ask again in a month.