How AI Assistants Choose What to Cite: The Evidence So Far

Ask ChatGPT or Perplexity a question about, say, the best CRM for a 20-person startup, and you'll get an answer with sources attached — three links, five links, sometimes none. Marketers want to know why their page didn't make the list. The honest answer: nobody outside the model labs has the full mechanism, and the labs aren't publishing it. But after two years of people probing these systems, patterns have emerged that are solid enough to act on.

What we actually know

Retrieval happens before generation, and it's not just a Google API call. Perplexity, ChatGPT with browsing, and Gemini all run some form of retrieval-augmented generation: a query goes out to a search index or live crawl, a set of candidate pages comes back, and the model reads those pages before writing an answer. This is confirmed by the labs themselves in product documentation, even if the ranking internals aren't. Perplexity has been the most transparent, describing a multi-stage pipeline of query expansion, retrieval, and re-ranking before the language model ever sees the content.

Authority signals carry over from traditional search. Domains that rank well in Google and Bing tend to show up in AI answers too, because several of these systems lean on the same underlying web indexes or on Bing's index specifically (this was confirmed for early ChatGPT browsing and is still plausible for current versions). That means backlinks, domain age, and established topical authority haven't become irrelevant — they've become one input among several rather than the whole game.

Structure helps retrieval, even if it doesn't decide citation on its own. Pages with clear headers, a direct answer near the top, tables, and defined terms are easier for a retrieval system to chunk and match against a query. This isn't a mystical "AI likes bullet points" claim — it's a byproduct of how retrieval systems work. Most of them break pages into chunks (paragraphs, sections) and embed those chunks for semantic search. A page where the answer to "what does X cost" is buried in paragraph nine of a rambling post is genuinely harder to retrieve accurately than one where it's stated plainly under a heading that says "Pricing."

Recency matters for some query types and not others. Ask about "AI marketing trends 2026" and recency is a heavy signal — stale content gets filtered out or down-weighted. Ask "what is a lookalike audience" and a page from 2019 might still get cited if it explains the concept clearly, because the query isn't time-sensitive. Confusing these two categories is a common mistake: teams update timeless glossary content on a schedule as if it were news, which wastes effort that would be better spent on genuinely time-sensitive pages.

Direct answers get quoted; hedge-y answers get skipped. Across informal testing by SEO practitioners (not lab-published research), pages that state a clear position — a number, a verdict, a named winner — get pulled into AI answers more often than pages that say "it depends" for six paragraphs before offering anything concrete. This tracks with how extraction likely works: a system trying to answer a direct question needs a directly quotable sentence, and vague content doesn't supply one.

Citation counts vary wildly by product and by query type, and this seems intentional. Perplexity often cites five or more sources per answer. ChatGPT's default answers sometimes cite zero, then switch to heavy citation mode when browsing is explicitly invoked. Gemini's citation behavior differs across its search-grounded and non-grounded modes. This isn't one universal algorithm — it's several different products with different citation UX philosophies, which is worth remembering before treating "AI SEO" as a single target to optimize for.

What's still genuinely unclear

Nobody outside these companies knows the exact weighting between authority, recency, semantic match, and structural clarity. Nobody knows how much duplicate or near-duplicate content gets filtered before it's even considered. Nobody has confirmed whether user engagement signals (like click-through on cited links) feed back into future citation likelihood, though it would be unsurprising if they did. Any content strategist who tells you they've "cracked the algorithm" is overselling a set of reasonable inferences as certainty.

What this means practically

Given the honest uncertainty, the practical response is to optimize for the things we're confident matter and treat everything else as a bet, not a certainty:

• Write direct, quotable answers near the top of the page for any question-shaped topic.
• Keep genuinely time-sensitive content current; don't waste cycles "refreshing" evergreen glossary pages that don't need it.
• Maintain the same authority-building habits that worked for traditional SEO — backlinks and topical depth are not obsolete.
• Structure pages so a machine parser can find the answer without reading the whole thing: headers, short paragraphs, defined terms, tables where relevant.
• Don't chase citation counts in one tool at the expense of doing the fundamentals across all of them — a page that's well-structured for Perplexity tends to also work for ChatGPT and Gemini, because the underlying retrieval mechanics rhyme even if they aren't identical.

The mechanism will keep evolving, and probably keep getting less public rather than more. Treat this space the way SEO practitioners eventually treated Google's algorithm: build on the load-bearing fundamentals, test cautiously, and don't over-invest in any single signal that could change with the next model update.