Last week a client asked the question we now hear most. “Why did Perplexity quote our competitor by name when we answer the same query better?” Fair question. The honest answer is that getting cited by an answer engine is not the same job as ranking on Google, and most teams are still trying to win the new game with the old playbook.
Here is what we have learned from the last 18 months of pulling apart citations across ChatGPT, Perplexity, Claude and Google AI Overviews. Some of it overlaps with classic SEO. A lot of it does not.
How an answer engine actually picks a source
Most LLM products do not browse the live web for every query. They use a retrieval layer (sometimes a search index they built, sometimes Bing, sometimes Brave, sometimes their own crawler) to fetch a small handful of candidate documents, then they ground the generated answer in those documents. The citation you see is one of the documents the retrieval layer surfaced. That means two things matter independently. The retrieval layer has to pick your page. The generator has to find your page useful enough to quote. Classic SEO mostly handles the first half. The second half is new work.What the retrieval layer is looking for
The retrieval layer cares about the same things classic search has always cared about. Topical authority, freshness, structured data, credible inbound links, and a clean technical profile. If you have done good SEO over the last five years, you are in the candidate pool. What it cares about that you might be missing:- Entity coverage. Does your page mention the entities (people, products, companies, standards) the query is about, with the right names and the right relationships? Wikipedia and Wikidata are still the implicit dictionary the retrievers use.
- Crawl access. Have you accidentally blocked GPTBot, Google-Extended, ClaudeBot, or PerplexityBot? Check your robots.txt today. Half the sites we audit are blocking at least one.
- Source diversity. Retrievers pull a mix. If your page is the third article on a topic written from the same angle as 200 others, it does not get picked.
- Schema markup. FAQ, HowTo, Article, Product. Schema is no longer just a SERP feature. It is a parser shortcut for the LLM.
What the generator is looking for
Once your page is in the candidate pool, the generator decides whether to actually quote you. This is where most pages fail. LLMs preferentially quote text that is short, factual, self contained, and clearly attributable. A page that buries the answer 600 words in, behind three rhetorical questions and a story about the founder, will not get quoted even if it ranks. A page that opens with a clean two sentence definition will.Practical changes that have moved the needle for our clients:If a smart intern can copy three sentences from your page and paste them into a client email as the answer, the LLM can quote them too.
- Lead with the answer. The first 200 words of every page should contain a clean, quotable version of the core claim.
- Use short, declarative sentences for the parts you want quoted. Reserve flourish for everywhere else.
- Cite your own sources. LLMs treat pages that cite primary research as more trustworthy than pages that do not.
- Use stable definitions. If you call a thing “Answer Engine Optimisation” on one page and “generative engine optimisation” on another, you split your own entity.