What Is an Answer Engine?
An answer engine reads the web on your behalf and returns a composed answer instead of a list of places to look. The sources are still there, but they have moved from being the result to being the evidence behind it.
What is an answer engine?
An answer engine is a system that takes a question, retrieves material from an index of documents, and writes a single response from what it retrieved, citing some of the documents it used. Google AI Overviews, ChatGPT search, Perplexity and Claude all work this way. The defining move is composition: the system does not hand back the ten best pages, it reads candidate passages and produces prose.
That shift matters because it changes what a publisher is competing for. On a results page, the unit of success is a position. In a composed answer, the unit of success is being the passage the system draws from and names. Those are related but not identical, and a page can win one while losing the other.
The term sits alongside two others that overlap with it. Answer engine optimization, or AEO, is the practice of writing and structuring content so it is retrieved and cited inside these answers. Generative engine optimization, or GEO, is used by many practitioners for the same work. Neither term has a standards body behind it, and in day to day use the distinction between them is mostly a matter of which word a given agency adopted first.
How is an answer engine different from a search engine?
The difference is not the index. In most cases an answer engine is reading the same web, often through infrastructure built by the same company, and in Google's case the AI features sit directly on top of Search. The difference is what happens after retrieval.
A search engine ranks. It produces an ordered list and leaves synthesis to the reader, who clicks, skims, and decides. The engine's commitment is thin: it says these pages seem relevant, in roughly this order.
An answer engine commits. It asserts something, in its own words, and then points at sources. That assertion is a far stronger claim than a ranking, and it creates a different failure mode. A badly ranked list wastes a click. A badly composed answer states something untrue in a confident voice, and the sources underneath it may or may not support the sentence they appear to back.
There is a second structural difference that publishers feel immediately. Responses are not deterministic. The same question asked twice, by two people, on two days, can return different sources. Model version, region, personalization and timing all move the result. A ranking can be checked once and reported. An answer engine can only be sampled.
How does an answer engine build a response?
Four stages, in order, each one a place where a publisher has a different amount of leverage.
Retrieval. The system turns the question into one or more queries and pulls candidate documents from an index. Google describes this stage for its AI features as query fan-out: the system issues multiple related searches across subtopics of the original question, which is why the links shown can be broader than what a single search would return. A publisher's leverage here is ordinary and familiar: be in the index, be retrievable for the subtopics around the question, not only for the question itself.
Selection. From the candidates, the system picks passages. This is passage level, not page level. A long page can contribute three sentences from its middle and nothing else. Leverage here comes from structure: a section that answers a question directly, near the top of that section, in self contained prose, is easier to lift than the same information spread across four paragraphs that depend on each other.
Synthesis. The system writes the answer. A publisher has no leverage at this stage. Nothing in a page controls how a model phrases a sentence or which of two conflicting sources it believes. This is the stage most AEO advice quietly pretends to influence.
Attribution. The system attaches links. Which sources get named, how many, and whether a named source genuinely supports the sentence next to it varies by engine and is not documented in detail by any of them.
Which answer engines are in use today?
Four are in common use, and each publishes its own crawler documentation, which is the most reliable public window into how it treats publishers.
Google AI Overviews appear inside Google Search results. Google states plainly that no new markup is required to appear in them, including no special structured data, and that the existing preview controls govern inclusion.
ChatGPT search is served by OpenAI, which runs separate user agents for search indexing and for model training, so a publisher can allow one and refuse the other.
Perplexity was built around cited answers from the start, and its documentation describes its indexing crawler as surfacing and linking websites rather than collecting training data.
Claude, from Anthropic, likewise separates a training crawler from a search crawler and from the agent that fetches a page because a user asked something.
Each has its own page here with what is documented and what is not.
What does this change for publishers?
Less than the loudest advice suggests, and more than the most dismissive advice admits.
What does not change: you still need to be crawlable, indexed and accurate. Every one of these systems reads the web with a crawler that obeys the same conventions publishers have worked with for twenty years. Google says directly that there is no AI specific markup and no AI specific schema to add.
What does change is the measurement, and this is the part most teams get wrong. If you report answer engine visibility the way you report rankings, with a single number from a single check, you are reporting noise. Visibility inside composed answers can only be established by sampling the same question set repeatedly across engines and dates, and then describing a distribution rather than a position.
The other real change is in control. Every vendor here now runs at least two crawlers with different purposes, and at least one of them, the agent that fetches a page because a user asked a question in that moment, is treated differently from an indexing crawler. Perplexity states that its user agent generally ignores robots.txt for this reason, and OpenAI states that its user triggered agent is not governed by robots.txt either. A publisher who blocks the indexing crawler has not necessarily stopped the system from reading the page.
Why do AEO and GEO both exist as terms?
Because two groups named the same practice at roughly the same time and neither name won. Answer engine optimization describes the work by the kind of system it targets. Generative engine optimization describes it by the technology inside that system. Both refer to writing and structuring content so it is retrieved and cited inside composed answers.
A few practitioners draw a finer distinction, usually that AEO covers the broader category including featured snippets and voice results, while GEO is specific to generative systems. The distinction is defensible but it is not consistently observed, and no vendor uses either term in its own documentation. Google's documentation calls them AI features and describes the path to appearing in them as ordinary SEO.
The practical advice is to use whichever term your audience uses and to be suspicious of anyone who insists the difference is important enough to charge for. A vocabulary dispute is not a methodology.
Where does this leave traffic?
Uncertain, and anyone who tells you otherwise with a percentage is reporting a third party estimate as if it were a measurement.
The mechanism is clear enough: an answer that satisfies a question reduces the reason to click through to the source of that answer. The magnitude is not clear, because no vendor publishes how often its feature is shown, for which queries, or what happens to clicks on the sources it names. Every figure in circulation on this comes from third party sampling with its own methodology, usually built on a query set chosen by the firm publishing the number.
What a publisher can do about the uncertainty is measure their own case rather than import someone else's. Search Console still reports impressions and clicks for pages that appear in Google results, including those appearing alongside AI features, and a publisher watching their own click-through rate on their own query set over time has a better number than any industry average. It is a narrower question than the one being debated publicly, and it is the one that affects their decisions.