What Is Answer Engine Optimisation, and How Does It Differ From SEO?

Answer engine optimisation

The practice of making a page retrievable by AI answer engines and writing its claims so that a single passage can be lifted out and attributed.

Also known as: AEO, Generative engine optimisation, GEO

The naming is unsettled and worth knowing about. Practitioners mostly say AEO, while the academic literature says GEO, from Aggarwal and colleagues, "GEO: Generative Engine Optimization", published at ACM SIGKDD 2024. Both names describe the same goal, so treat them as synonyms and ignore anyone selling a distinction between them.

A citation is a concrete, observable thing. When an assistant answers a question, it lists the pages it used, and in some cases the exact span of text it took. Being cited means your URL appears in that list. Being described without a citation is a different outcome, and it comes mostly from what other sites say about you rather than from your own pages.

The practical consequence is a change of unit, not a change of craft. A comprehensive page that buries its answer in paragraph six ranks fine and gets quoted rarely, because there is no sentence on it that survives extraction. A short page with one exact, dated, self-contained claim does the opposite.

Two of these are gates rather than opportunities. OpenAI documents that sites opted out of OAI-SearchBot "will not be shown in ChatGPT search answers", and Microsoft documents that Copilot Search is grounded on Bing results. So a robots.txt line and a Bing index entry decide two of the six surfaces before any writing happens.

Google is the only engine where the eligibility bar is published in full. A page has to be indexed and eligible to be shown in Google Search with a snippet, and Google states there are no additional requirements and no special optimisations necessary.

The strongest evidence against AEO as a separate craft comes from a benchmark built to test it. Puerto and colleagues, "C-SEO Bench: Does Conversational SEO Work?", published at NeurIPS Datasets and Benchmarks 2025, concluded that "most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking", while traditional SEO was "significantly more effective". The same paper showed that whatever gains exist decay as adoption rises.

The test is Vishwakarma, Kumar and Jamidar, "What Gets Cited: Competitive GEO in AI Answer Engines", ACM SIGIR 2026. It ran 252,000 trials across six large language models and 18 content factors, injecting paired documents into retrieval context with brands anonymised and source order counterbalanced. The four gatekeepers above had odds ratios far above 100.

What the same test found negligible is the interesting half: formatting and structure changes, paragraph organisation, and neutral versus promotional tone. That inverts most published AEO advice, which is largely about reformatting.

The honest caveat is the domain. The trials sat in a product-review setting with documents injected into context rather than retrieved live, so the ranking of factors is better evidence than the absolute effect sizes.

What deliberately is not on that list: publishing an llms.txt file, splitting pages into chunks, adding FAQ markup, and expanding pages to hit a word count. All four are common first moves and none has measurable support.

Recommendation questions are the exception to everything above, because they are decided off your site. Oppira tracks which brands and domains get named across the public sources those answers lean on, which turns "why do assistants recommend someone else" into a list you can act on rather than a guess.