Definition
A generative engine writes its answer with a language model instead of returning documents, using passages it just retrieved from the web. Unlike an answer engine, which may return a stored fact or a quoted snippet, a generative engine composes new text every time. Being cited by one means being named inside that text.
At a glance
- CategoriesCore Concept
- Related fieldsAI Search
- DifficultyBeginner
At a glance
- A generative engine always does two jobs: it retrieves, then it writes.
- The written output is new text, so your sentences inform it rather than appear in it.
- ChatGPT, Perplexity, Claude and Google’s AI surfaces are all generative engines with different retrieval behind them.
- Optimizing for one is generative engine optimization, which targets citation rather than rank.
The pipeline behind a written answer
Four steps, always in this order.
The question gets interpreted and usually split, so one typed sentence becomes several searches aimed at different parts of it. Retrieval runs those searches and returns candidate passages, cut down to a set that fits the model’s context window.
The model then writes, composing a reply in its own words from the surviving passages, guided by product instructions about length, tone and how confidently to speak.
Attribution comes last and is looser than people assume. Citations mark where the model believes it drew on a source. A sentence may blend two sources and carry one link, or restate something the model already knew and carry none.
Your page becomes raw material
The old deal was that search sent people to your page and your page did the persuading. A generative engine reads the page on the reader’s behalf and does the persuading itself.
Take a shopper asking which mattress firmness suits a side sleeper. The engine reads your guide, a competitor’s guide and a couple of review sites, then writes a short reply drawing on all of them. Your framing, your photography and your call to action never reach the person. A paraphrase does.
So the work moves. You stop competing on presentation and start competing on whether your facts are clear, checkable and consistent with what other sources say. An AI-generated answer is only as good as the passages behind it, and being one of those passages is the job.
Feeding a generative engine cleanly
Everything here is doable with the pages you already have.
- Put one clear claim at the top of each section. The model quotes claims, not build-ups. If the first sentence hedges, the section is hard to use.
- Make your facts checkable. State the number, the date and where it came from. A fact the model can attribute is a fact it will attribute.
- Keep your claims consistent across the site. Contradictory descriptions on your product page and your about page make you a source the model will hedge around.
- Serve it as plain HTML. Content that needs JavaScript to appear is content that was never retrieved, and a passage that is never retrieved is never rewritten.
Frequently asked questions
-
Is a generative engine the same as a search engine?
No. A search engine finds documents and ranks them for you to read. A generative engine finds documents, reads them itself, and writes you an answer. The retrieval half is similar enough that most of your technical work carries over. The output half is completely different.
-
How do I get cited by a generative engine?
Be retrievable, then be quotable. Retrievable means crawlable HTML, indexed pages, facts in text rather than images. Quotable means short self-contained passages that answer one question outright. Most brands fail the first test and spend their time on the second.
-
Will a generative engine use my exact wording?
Sometimes, more often not. The model paraphrases by default, so plan to influence the substance rather than the phrasing. Consistent phrasing across your own pages and third-party mentions raises the odds of your framing surviving, but nothing guarantees it.
-
Can I stop a generative engine from using my content?
Partly. Robots directives block the crawlers that respect them, and snippet controls limit how much gets shown. Neither is complete, since models hold knowledge from training and other sites quote you. Blocking also removes you from answers where competitors appear, so treat it as a business decision.
-
Is there a metric for generative engine performance?
Not a standardized one, and be wary of anyone presenting theirs as an industry figure. What you can measure honestly is how often you are named across a fixed list of prompts, checked on a schedule. Anything expressed as a share of a market nobody can size is an estimate wearing a decimal point.