Answer engine optimization, or AEO, is the practice of structuring content so that AI assistants cite it when answering a question. The target is not a ranking position. It is being the source an answer engine quotes, links, or names when someone asks ChatGPT, Gemini, Perplexity or Google's AI Mode about your subject. The discipline exists because a growing share of queries now end in a generated answer rather than a list of blue links, and a page that is never cited in that answer is invisible regardless of where it ranks.
That is the definition, and it is the easy part. The harder question is which AEO tactics actually work, where the honest answer is uncomfortable: most published AEO advice has never been tested against anything. This piece separates the small body of measured evidence from the much larger body of confident assertion, and names one widely-sold tactic whose justification Google switched off in May 2026.
AEO, GEO, LLMO: The Same Thing With Different Labels
The vocabulary proliferated faster than the practice.
- AEO (answer engine optimization) is the most common term in commercial use.
- GEO (generative engine optimization) is the term used in the academic literature, originating with the paper that produced most of what we actually know.
- LLMO and "AI SEO" are marketing coinages for the same activity.
There is no meaningful methodological difference between them. Anyone selling GEO as a distinct discipline from AEO is selling a label. This piece uses AEO throughout and cites the GEO literature, because that is where the measurement happened.
What Is Actually Measured
Four sources carry nearly all of the reliable evidence. Everything else in circulation is inference from these, or assertion with nothing behind it.
| Finding | Figure | Source | What it establishes |
|---|---|---|---|
| Adding statistics and authoritative quotations lifts citation rate | Up to 40% relative improvement | Princeton and IIT Delhi GEO study, ACM SIGKDD 2024, 10,000 queries across 8 domains | The only controlled experiment on citation tactics |
| ChatGPT now cites sources far more often than a year ago | US citation rate rose from ~1.6% (June 2025) to ~6.8% (May 2026) | Similarweb, from its own traffic panel | The citation surface is expanding, not closing |
| Citation rates vary enormously by sector | Travel and hospitality ~23%; professional services under 4% | Similarweb, from its own traffic panel | Sector, not tactics, may dominate outcomes |
| Citations mostly pay off without a click | AI recommendations make users ~2.5x more likely to reach a brand via branded search | Similarweb, from its own traffic panel | The payoff is largely invisible to referral analytics |
| AI-sourced traffic converts better than other channels | 42% better than non-AI traffic by March 2026, reversed from 38% worse a year earlier | Adobe Analytics Q2 2026 AI Traffic Report, over one trillion US retail visits | Citations are commercially worth pursuing |
The Princeton study is the load-bearing item and deserves care. Aggarwal and colleagues tested content modifications across 10,000 queries and found that the strategies producing the largest measured lift were adding relevant statistics and adding quotations from authoritative sources. The headline "40%" is a maximum on a position-adjusted word count metric, not an average, and not a promise. It remains the strongest evidence in the field, which says as much about the field as it does about the study.
The Tactic Google Killed, That the Industry Still Sells
The clearest illustration of how much AEO advice runs on folklore is FAQ schema.
For years, adding FAQPage structured data was standard advice, justified by the expandable FAQ rich results it produced in Google's search listings. Those rich results stopped appearing on 7 May 2026. Google removed the FAQ search appearance and rich result report, dropped support from the Rich Results Test in June 2026, and removed Search Console API support in August 2026, adding a line to its own documentation stating that FAQ rich results no longer appear in Search. No blog post, no explanation.
The advice did not change. A great deal of AEO guidance published since May 2026 still recommends FAQ schema, and still justifies it by the rich results that no longer exist.
Here is the nuance that matters, because the conclusion is not "stop writing FAQs". The evidence points to the content shape doing the work, not the markup. Answer engines extract clean, self-contained question-and-answer pairs whether or not schema wraps them. So FAQ sections remain worth writing, for a completely different reason than the one everyone was given. Keep the FAQPage markup if you have it, since it is a valid Schema.org type and Google confirms unused structured data causes no harm. Just stop justifying the work with a lever that was switched off.
The general rule this illustrates: if you cannot name the mechanism by which a tactic works, you are following folklore. Applied honestly, that test removes a large fraction of published AEO checklists.
What The Evidence Supports Doing
Five practices survive the test. Each has either measured evidence or a mechanism you can state plainly.
Put the answer first, in the first paragraph. An answer engine extracting a passage needs one that stands alone. A paragraph that begins with scene-setting gives it nothing liftable. This follows directly from how retrieval works: content is chunked and embedded, and a chunk that depends on the paragraph above it loses its meaning when returned by itself.
Include statistics with named sources. This is the single strongest finding in the Princeton study, and it is cheap to do. An unattributed number is weaker than an attributed one, and an invented one is a liability.
Quote authoritative sources directly. The second strongest finding. Quotation gives the answer engine something to carry verbatim with attribution already attached.
Write definitions in "X is Y" form. Explicit definitional sentences get extracted as answers because they map exactly onto the question shape. "Answer engine optimization is the practice of structuring content so that AI assistants cite it" is retrievable; "in today's evolving search landscape, brands must consider..." is not.
Make each section self-contained. Re-state the entity rather than relying on a pronoun that refers back three paragraphs. This is the same answer-first discipline applied at section level rather than only at the top of the page, and it is the most commonly skipped step.
Before any of that, check the engines can read the page at all. Extraction depends on ingestion, and this is the step most AEO checklists assume rather than verify. If robots.txt blocks GPTBot, Google-Extended, ClaudeBot or PerplexityBot, none of the tactics above can help, because the content is never retrieved. The same applies to content that only appears after client-side JavaScript execution, which some crawlers will not wait for. Bot access is not an AEO tactic, it is the precondition for every AEO tactic, and it takes two minutes to check.
What The Evidence Does Not Support
Schema volume. There is no measured evidence that adding more structured data types increases citation. Accuracy matters; quantity is unevidenced.
Keyword density in any form. Retrieval operates on embeddings, not term frequency. The mechanism that made keyword density matter for classic search does not exist here.
"Optimising for each engine separately." The engines draw on overlapping corpora and similar retrieval architectures. There is no published evidence of engine-specific tactics that outperform writing well-structured, well-sourced content once.
Any promised citation rate. Note the Similarweb sector spread: travel and hospitality at roughly 23%, professional services under 4%. If sector variance is that wide, an agency quoting you a target citation rate without knowing your sector is guessing.
The Uncomfortable Structural Point
AEO has a measurement problem that SEO does not. In search, rank is observable: you can look it up. In AEO, whether you were cited depends on the prompt, the engine, the day, and the user's context, and none of that is in your logs. Answer engines send comparatively little referral traffic, so analytics understate the effect badly.
Similarweb's finding that AI recommendations make users roughly 2.5 times more likely to reach a brand through branded search, rather than through a referral link, is the crux. The mechanism by which AEO pays off is largely invisible to the tools most organisations use to measure it. A brand can be cited constantly and see almost nothing in its referral reports. The realistic outcome of good AEO is not a traffic spike; it is being present in the answer, and then being searched for by name later. You are optimising for mindshare inside the chat interface, and accepting that most of it will never produce a click you can attribute.
The practical consequence is that measurement has to be built rather than read off a dashboard. What it requires: a fixed set of prompts representing how buyers actually ask about your category, run against each engine on a repeating schedule, with the generated answers parsed for brand mentions and cited URLs, stored as a time series. The engines are non-deterministic, so a single check tells you almost nothing; the denominator is days sampled, not answers. Without that, you cannot distinguish work that is succeeding invisibly from work that is not succeeding at all, and you will be tempted to judge AEO by the one number it was never going to move.
Where To Start
Audit one page you want cited. Check three things in order. Does the first paragraph answer the question by itself, without the heading? Does each section stand alone if extracted? Does every factual claim carry a named source at the point it is made?
That sequence covers the tactics with actual evidence behind them, and it takes an afternoon. It is a better use of the time than another pass over structured data.
