How to Get Mentioned in ChatGPT, Gemini, Perplexity & AI Overviews: The Complete GEO Framework for 2026
How to Get Mentioned in ChatGPT, Gemini, Perplexity & AI Overviews: The Complete GEO Framework for 2026 I spent the better part of last year staring at something that didn’t make sense: a client ranking #1 for their money keyword on Google, watching a competitor ranked #6 get quoted by name in ChatGPT for the exact same query, over and over.That gap is the entire subject of this article. Section 1: The Biggest Misconception About AI SEO Most SEOs are still treating generative engines like a new SERP. They’re optimizing title tags, chasing featured snippets, and hoping the same playbook that got them to position one on Google will get them cited in ChatGPT. It won’t, and the reason is structural, not tactical.Google’s ranking algorithm answers one question: which page best satisfies this query, given the click behaviour of millions of past searchers? It’s a relevance-and-popularity contest, refined over two decades, and it’s fundamentally comparative — page A vs page B vs page C.An LLM answering a question isn’t running a comparative contest. It’s doing something closer to evidence synthesis. It has already generated a draft answer from its training weights, and it’s now deciding: do I trust this draft enough to state it plainly, or do I need to retrieve a source to back it up — and if I retrieve, which source do I lean on?That’s a confidence decision, not a ranking decision. And confidence is built from a different set of signals than PageRank-style authority.Here’s the part that trips up experienced SEOs specifically: a page can rank #1 because it’s the best clickable answer to a broad query, while being a terrible citation candidate because it hedges, buries the actual number three paragraphs down, or reads like eight other top-10 pages that said the same thing in the same order. Google’s ranking rewarded it for comprehensiveness and UX. The LLM doesn’t care about UX. It cares about whether the specific sentence it needs is stated cleanly enough to lift.I saw this literally with a fintech client (a money-transfer product). They ranked top 3 for “send money to India from the UK” — solid position, strong domain. But when I ran the same query family through ChatGPT and Perplexity, they were absent. The site that kept showing up was a much smaller comparison blog, DR nowhere near ours, that had one thing we didn’t: a table with exact transfer fees, updated dates, and a named author with a finance byline. The LLM wasn’t rewarding their domain authority. It was rewarding the fact that their number was retrievable and attributable in one clean lift.That’s the misconception in one sentence: ranking optimizes for “best page for a human to click.” GEO optimizes for “safest sentence for a machine to repeat.” Those overlap less than the industry assumes. Section 2: Reverse Engineering How AI Systems Find Sources Skip the academic definitions. Here’s what’s actually happening when someone asks ChatGPT or Perplexity a question, from the seat of someone who’s been watching citation patterns across dozens of client verticals.The retrieval step (RAG) isn’t a black box, it’s a filter cascade. The system takes the query, expands it into related sub-queries (this is query expansion — “best CRM for small business” quietly becomes searches for “CRM pricing comparison,” “CRM features small business,” “CRM reviews 2026”), then runs retrieval against an index (its own crawled corpus, a live search API, or both, depending on the product). What comes back isn’t one page — it’s dozens of candidate passages. The generation model then picks which passages to actually cite.This is the step most SEOs never think about: you’re not competing to be found, you’re competing to be selected after you’re found. Plenty of pages get retrieved into the candidate pool and never get cited because the passage that got pulled was vague, unattributed, or contradicted by three other candidate passages.Why does ChatGPT cite one website and ignore another that covers the same topic? In practice, from pattern-watching across client sites, it comes down to a handful of things:Passage isolability. Can a single paragraph or sentence stand alone and answer the question without needing the rest of the page for context? Pages that build an argument across six paragraphs before stating the conclusion get skipped in favour of pages that state the conclusion first and support it after.Entity clarity. Does the page clearly say who is speaking? A stat with no named source, no organization, no date reads as unverifiable. LLMs are conservative about repeating unattributed claims because hallucination risk cuts both ways — the model doesn’t want to launder an unreliable number.Corroboration. If three sources say the same number and one says something different, the outlier gets dropped even if the outlier is your page. This is the single most underrated factor — being right and alone is worse, from a citation standpoint, than being repeated across a small cluster of trusted sources.Freshness signals that are actually verifiable. Not just a “last updated” stamp (those get gamed and the models increasingly discount them) but internal evidence of recency — a dated example, a specific version number, a stated “as of [month, year].”Entity recognition and knowledge graph relationships matter less at the sentence-retrieval level and more at the “should I trust this domain at all” level. If your brand entity is inconsistently represented — different name variants, no clear Wikipedia/Wikidata presence, no consistent author identity across content — the model has a harder time resolving “who is this” before it even gets to “what are they saying.” Weak entity resolution doesn’t block retrieval, but it lowers the confidence weighting on anything retrieved from you. Section 3: The GEO Framework I built this from the recurring pattern across every “why did they get cited and we didn’t” audit I’ve run. It’s not clever for the sake of being clever — it’s five things that consistently mattered, in the order they need to be tackled.G — Ground every claim in a retrievable unit A retrievable unit is