Answer engine optimization is the practice of shaping content so that AI answers select it as a source, which happens in four separable stages: retrieval, selection, generation and citation.
Most guides on this topic define the acronym and stop. That is the wrong place to stop, because a page can pass one stage and fail the next, and the fix is different every time. This article walks the pipeline with the measurements that exist for each stage, then shows what we see across 8.94 million tracked AI answers.
Start with the least ambiguous source available. Google Search Central's own guide to generative AI features states the position plainly: "From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO."
The same document confirms two mechanisms by name. Retrieval augmented generation, also called grounding, uses Google's core ranking systems to fetch relevant, current pages from the Search index. Query fan-out is described as "a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results". Google's own example: for "how to fix a lawn that's full of weeds", fan-out queries might include "best herbicides for lawns" and "remove weeds without chemicals".
Google also says there are no special AI specific technical requirements beyond ordinary Search eligibility, and that AI Overviews and AI Mode may use different models, so their links can differ. We measured how far apart those two surfaces actually land in our work on query fan-out.
Take that at face value and one conclusion follows immediately: there is no separate AEO checklist that bypasses search fundamentals. What is genuinely new is the stage structure underneath, and that is where the measurements get interesting.
The single most useful thing to understand about answer engines is that these are not the same thing.
| Event | What it means | Where it is decided |
|---|---|---|
| Retrieved | Your page came back in the candidate set | Search ranking and query fan-out |
| Displayed | The interface shows your link | Interface rules, varies by engine |
| Used | Text from your page entered the model's context | Passage selection |
| Cited | The answer attributes a claim to you | Generation and citation rendering |
Venkit and colleagues at Salesforce AI Research and Penn State measured the gap between two of these in 2024. Across 303 curated questions run on You.com, Bing Chat and Perplexity, producing 909 engine question observations, the engines displayed a mean of 4.31 sources but cited a mean of 3.00 in the answer itself. Perplexity displayed 5.00 and cited 2.58. Bing Chat displayed 4.46 and cited 2.80. You.com displayed 3.57 and cited 3.57.
Being visible in the source strip is therefore not the same as being the source the answer was built from, and the difference is roughly a quarter of the sources on some engines.
Retrieval is ordinary search ranking, expanded by fan-out. This is the stage where classic SEO does almost all the work, and the stage where the ceiling is set. It is also the only stage a crawler can block outright, which makes AI crawler and bot traffic analytics the first thing to check when a page never appears at all.
The ALCE benchmark from Gao and colleagues at Princeton, published at EMNLP 2023, shows how tight that ceiling can be. The system retrieves the top 100 passages, splits corpora into 100 word passages, and passes only a small top-k subset into the model context. On the ASQA dataset, the top five retrieved passages covered only 56.8% of the answers. Nearly half the answerable material never reached the model at all.
That number reframes a lot of AEO advice. If your page is not in the candidate set, nothing you write on it matters. Retrieval eligibility is a prerequisite, not a tactic.
Once candidates exist, something has to choose which passages enter the context window. This is where the mental model most people carry, ranking pages one by one, is actively misleading.
SetR, published at ACL 2025, argues that scoring passages individually is insufficient for multi hop questions, and that selection should optimise relevance, diversity, completeness and redundancy at the set level. In its controlled experiments, SetR selects from the top 20 retrieved passages and puts an average of 2.91 passages into the final context, rather than the conventional five.
Read that as a competitive statement. If an answer assembles a set of roughly three passages that together cover a question, then a page that duplicates what another source already says is not a second option. It is redundant, and redundancy is a property the selector is explicitly penalising. The way to be picked is to cover something the rest of the set does not.
Even material that reaches the model does not reliably get attributed. Liu, Zhang and Liang at Stanford measured this in Evaluating Verifiability in Generative Search Engines, EMNLP 2023: 1,450 queries per engine across Bing Chat, NeevaAI, Perplexity and YouChat, with 34 human annotators, collected in February and March 2023.
| Engine | Citation recall | Citation precision |
|---|---|---|
| Perplexity | 68.7% | 72.7% |
| NeevaAI | 67.6% | 72.0% |
| Bing Chat | 58.7% | 89.5% |
| YouChat | 11.1% | 63.6% |
| Aggregate | 51.5% | 74.5% |
Citation recall is the share of generated sentences fully supported by their citations. Precision is the share of citations that support the sentence they hang on. In aggregate, roughly half of what these engines asserted was fully backed by what they cited.
These are 2023 figures from systems that have since changed, and they measure verifiability rather than ranking probability. They are still the cleanest published evidence that the attribution layer is its own failure point, and that appearing in an answer's reasoning does not guarantee appearing in its footnotes.
Gao's benchmark points at the same seam from the other direction: reranking four sampled responses lifted ChatGPT on ASQA from 73.6% citation recall and 72.5% precision to 84.8% and 81.6%. The generation step, not the source, moved the number.
Benchmarks use controlled engines. These figures come from tracking live ones: 8.94 million analysed AI answers and 8.85 million cited sources, of which 4.2 million answers across 11 models in the last 90 days.
Engines differ far more in how many sources they cite than in anything else. Across 731,000 answers in a 14 day window:
| Engine | Answers | Avg. citations | Median | Answers with any citation |
|---|---|---|---|---|
| Perplexity | 140,334 | 13.69 | 10 | 100.0% |
| Google AI Mode | 49,292 | 12.95 | 11 | 94.6% |
| AI Overviews | 219,495 | 8.73 | 8 | 96.2% |
| Copilot | 10,973 | 6.30 | 5 | 96.6% |
| Gemini | 65,778 | 5.24 | 4 | 82.2% |
| ChatGPT | 242,290 | 4.76 | 4 | 99.3% |
| Grok | 797 | 2.14 | 1 | 66.6% |
A Perplexity answer holds roughly three times the citation slots of a ChatGPT answer. That ratio decides more about where a given piece of AEO work pays off than any technique in any guide, and it is why Perplexity visibility and ChatGPT visibility deserve separate targets rather than one blended score.
The pool of cited domains is also far less concentrated than a search results page.
| Domains | Share of all citations |
|---|---|
| Top 10 | 6.2% |
| Top 100 | 16.6% |
| Top 1,000 | 37.1% |
| Top 5,000 | 57.9% |
It takes 1,000 domains to account for 37.1% of citations and 5,000 to reach 57.9%. The ten most cited domains together hold 6.2%. On a classic results page, ten results hold all of page one. The long tail in AI citations is not a consolation prize, it is where most of the volume is, and it is the strongest structural argument that a mid authority site can compete here at all.
| Stage | What decides it | What actually helps |
|---|---|---|
| Retrieval | Search ranking, query fan-out, crawlability | Rank in classic search, cover fan-out phrasings, keep the page fetchable |
| Selection | Set level relevance, diversity, completeness | Cover something the other candidates do not. Redundancy is penalised by design |
| Generation | Passage clarity, self contained paragraphs | One idea per paragraph, entities named explicitly, the answer in the first third |
| Citation | Attribution behaviour of the engine | Specific, checkable claims. Vague sentences are hard to attribute to anyone |
Reworking pages against those stages is what AI content optimization covers, and two habits follow from the whole pipeline rather than any single stage. Name entities explicitly instead of writing "the company" or "this tool", because a selected passage may start mid page with no subject in view. And make paragraphs stand alone, because you do not control the chunker and a paragraph that needs the one above it is a paragraph that can arrive without it.
The part no framework will tell you is which stage you are failing. That is an observation problem: you have to run the questions and look at what comes back, which is what AI citation tracking does by recording the cited sources per answer, per engine, over time. Pages that get retrieved but never cited are the largest and cheapest pool of upside most sites have, and you cannot find them without the answer level data.
The honest answer is that the distinction is mostly vocabulary. Google says optimising for its generative features is SEO. The Princeton GEO paper describes the same architecture the Google documentation does: a query reformulation model produces simpler queries, a search engine returns ranked sources, and downstream models summarise and synthesise them.
What is genuinely different from classic SEO is worth stating precisely, and it is three things. You are optimising for questions that often have no search volume, because fan-out generates them, which is what query fan-out analysis for AI Mode exists to surface. You are optimising for retrieval and selection rather than for clicks. And your result is a share of an answer rather than a position in a list, which means it is partly decided by who else is in the set.
If a term helps your team think, use it. Just do not buy a separate strategy on the strength of an acronym.
Our figures come from Finseo tracking data, aggregated across accounts, with no customer, project or private domain identifiable in any number. The engine comparison covers a 14 day window and states the answer count per engine, which is what makes Grok's 797 answers readable as the thin sample it is. The concentration figures cover the full source corpus.
Prompt sets are chosen by the customers who run them, so the corpus skews toward the categories those customers sell in. It is a large sample, not a random one, and any claim about AI search in general should be read with that caveat.
We tested nine of these techniques against the published evidence in a separate analysis of GEO techniques. Every external figure is attributed to its study with sample size and year. The 2023 verifiability figures come from systems that have since changed materially, which is stated rather than footnoted. Where studies use controlled or simulated engines rather than production endpoints, that is flagged inline.
What is answer engine optimization? AEO is the practice of shaping content so answer engines select it as a source. It covers four separable stages: being retrieved, being selected into the model's context, being used in the generated text, and being cited.
Is AEO different from SEO? Google's own documentation says optimising for generative AI features is optimising for search, and thus still SEO. What differs is that you are optimising for questions with no search volume, for retrieval rather than clicks, and for a share of an answer rather than a position.
What is the difference between AEO and GEO? Very little in mechanism. Answer engine optimization and generative engine optimization describe the same work from two vantage points. Both refer to the pipeline Google and the Princeton GEO paper describe in near identical terms.
Will AEO replace SEO? No, because retrieval runs on search ranking. In published measurements, pages ranking first in classic search are cited several times more often than pages outside the top 20. Classic performance is an input to the new surface, not a rival to it.
What is an example of AEO in practice? Rewriting a page so its first third answers the exact question it competes for, naming entities explicitly instead of using pronouns, and covering a fan-out phrasing that the other candidate pages miss. All three target a specific stage of the pipeline.
How do I know which stage my page fails? By running a fixed prompt set and recording, per answer, whether your page was cited and which sources were used instead. Retrieved but never cited is a different problem than never retrieved, and only answer level data separates them.
Ours
External
Disclosure: the tracking figures in this article come from Finseo, an AI visibility platform. External figures are attributed to their original source.
