Generative Engine Optimization and Backlinks: What Gets Cited in AI Search

Generative engine optimization (GEO) is the work of getting a page cited in AI-generated answers: Google's AI Overviews and AI Mode, ChatGPT search, and other assistants that answer with sources. It overlaps with answer engine optimization (AEO), the older term for getting content used as the direct answer in search features and assistants. A lot of advice about it is confident and unsourced. This guide sticks to what the engines themselves document and what published research has measured, then looks at the part that matters to anyone buying links: what backlinks do and do not do for AI search.

What the engines actually say

Google. Google's guidance for site owners on AI features is short and direct: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." To be shown as a supporting link, "a page must be indexed and eligible to be shown in Google Search with a snippet." Google also says "all existing SEO fundamentals continue to be worthwhile", and that you don't need to create "new machine readable files, AI text files, or markup to appear in these features." You can limit what appears with the same controls that govern snippets: nosnippet, data-nosnippet, max-snippet and noindex.

In other words, for Google, getting cited in AI answers starts with being indexed and eligible for a snippet in ordinary search. Nothing Google publishes describes a separate AI ranking you can buy your way into.

ChatGPT search. OpenAI runs separate crawlers for separate purposes. OAI-SearchBot is "used to surface websites in search results in ChatGPT's search features", and OpenAI asks site owners who want to appear to allow it in robots.txt. GPTBot crawls content that may be used to train OpenAI's models. The two are independent: you can disallow GPTBot for training and still allow OAI-SearchBot for search.

The practical consequence is simple and often missed. A site that blocked "AI bots" in one sweeping robots.txt rule may have removed itself from ChatGPT search without meaning to. Check which crawler each rule names.

What the research measured

The term comes from a 2023 paper, "GEO: Generative Engine Optimization" by Aggarwal and colleagues, published at KDD 2024. The authors built a benchmark of real queries and tested how changes to a source page affected its visibility in a generative engine's answer. Their headline finding is that GEO methods "can boost visibility by up to 40%".

Which methods worked is the useful part:

Most experiments ran on a simulated engine (GPT-3.5 answering over Google results). The authors also tested on Perplexity, a deployed engine, where quotation addition improved visibility by 22% and statistics addition by 37% on one of their measures. The pattern is consistent: content that carries evidence (quotes, numbers, sources) gets used, and repeating a keyword does not.

That result should be familiar to anyone who has watched keyword-stuffed pages age badly in ordinary search.

Where backlinks fit, and where they don't

Nothing in Google's or OpenAI's documentation says that AI answers weigh backlinks differently from search. What the documentation does say is that the page must first be crawlable, indexed and eligible in search. Links remain part of how search engines discover and evaluate pages, and for AI features Google points site owners back to the same SEO fundamentals. So backlinks matter to AI search in the same indirect way they matter to search: they help a page get found, indexed and trusted. They are not a separate lever.

Three consequences follow for anyone buying links with AI visibility in mind.

A paid link does not buy a citation

An AI answer cites a page because the page answers the question with something worth citing. A placement bought on a site nobody reads does not change that, and Google's spam policies still treat paid links that pass ranking credit as link spam: they list "exchanging money for links, or posts that contain links" among their examples, unless the link is marked rel="sponsored" or rel="nofollow". Unless a vendor selling "AI citation links" can show otherwise, those are ordinary placements under a new name.

Syndicated copies are one source, not many

However much independent mentions count toward an AI answer, copies of one press release are not independent sources. bklink's public sample report audited 173 placements bought through a press-release distributor: 161 of them were the same two articles republished, none of the links was followed, and 70% of the publications shared server infrastructure. Read as mentions, that list is two sources repeated, not 173. The PR distribution agency audit shows how to count a placement list honestly.

Shared hosting is visible to anyone who looks

A catalogue of "independent" sites that resolve to the same server looks independent only on a sales sheet. The operator's footprint is one DNS lookup away. The backlink footprint audit covers what to check.

What to do instead

Most of this is ordinary good practice, which is the point: the engines say so.

1. Make sure the right crawlers can reach you. Googlebot for Search, OAI-SearchBot for ChatGPT search. Decide separately whether to allow training crawlers like GPTBot, because that choice does not affect search inclusion.
2. Be indexed and snippet-eligible. Check the page is indexed in Google, and that no nosnippet or max-snippet rule is hiding the passage you want quoted.
3. Put the evidence on the page. Quote the source you rely on, give the number with its origin, and link the primary source. These are the changes the GEO research found moved visibility most.
4. Answer the question in plain text near the top. Google says important content should be available in textual form. A clear answer in the first paragraph is easy to lift and easy to attribute.
5. Keep structured data honest. If you use it, Google asks that it match what is visible on the page.
6. Earn mentions from sources that stand on their own. One independent publication covering your data is a separate source; a hundred syndicated copies of your release are still one.
7. Check what you bought. If links are part of the plan, confirm each placement is live, followed where it should be, on a site that is not part of a network, and not a syndicated duplicate.

A note on llms.txt

llms.txt is a proposed file that summarises a site for language models. bklink publishes one at bklink.uk/llms.txt, because it is a cheap, readable summary of the product. But Google says outright that you do not need "AI text files" to appear in its AI features, and OpenAI's crawler documentation asks for robots.txt access, not an llms.txt file. Treat it as optional documentation, not a ranking factor.

The short version

• Google: no special optimisation for AI features; be indexed and snippet-eligible, and the SEO fundamentals still apply.
• ChatGPT search: allow OAI-SearchBot. GPTBot is a separate, training-only choice.
• The research: quotes, statistics and citations on the page improve visibility; keyword stuffing does not.
• Backlinks: they help a page get found and trusted, as in search. They do not buy a citation, and syndicated copies count as one source.

Related reading

• DR vs DA: What's the Difference, and Which One Should You Trust?: authority scores are vendors' own metrics, in AI search as anywhere else.
• How to Vet Link Building Companies Before You Hire One: what to check before buying placements for any reason.
• PR Distribution Agency Audit: The Placement List Truth: counting syndicated copies honestly.
• Backlinks for Finance Sites: how the same evidence standard applies to YMYL topics.