When ChatGPT names three tools in an answer and leaves yours out, it feels arbitrary. It isn't. There's a specific, reverse-engineerable pipeline behind every citation — and once you see the seven stages, it becomes obvious why some brands get mentioned and others don't.
ChatGPT picks brands through a seven-stage retrieval pipeline: it decides whether to search, expands your question into multiple queries, filters candidate pages by metadata, fetches the survivors, splits them into ~128-token chunks, scores each chunk for relevance, and cites the 3–5 whose passages best answer the question. Your brand competes at the passage level, not the page level.
I'll walk through all seven stages, because each one is a place you can win or lose. Most brands that are invisible to ChatGPT are failing at one specific stage — and they usually have no idea which. Understanding the pipeline is how you diagnose it.
A lightweight classifier decides whether the question even needs a web search. Timeless facts get answered from training data with no retrieval at all — no source, no citation, no chance for your brand.
ChatGPT expands one question into multiple keyword and semantic sub-queries. "Best CRM for agencies" might fan into sub-searches for pricing, integrations, agency-specific features, and alternatives.
From 40–50 URLs, ChatGPT narrows to 10–20 using only title tags, meta descriptions, and domain signals. No page content is read yet. If your title and meta don't signal relevance, you're cut before your content is ever seen.
The survivors get fetched — but ChatGPT's fetcher gives up after roughly 2 seconds. Slow servers get skipped entirely. Content hidden behind JavaScript often isn't seen at all.
Fetched pages are split into small chunks of about 128 tokens — roughly a paragraph. From here on, your page doesn't compete. Your individual chunks do.
Each chunk is converted to a vector and scored against the question by similarity. The highest-scoring chunks advance. This is why front-loading matters — 44.2% of citations come from the first 30% of a page.
ChatGPT selects 3–5 sources to build the answer, weighing relevance, authority, freshness, and diversity — then writes the response and cites them. Your brand is named, or it isn't.
Your page doesn't get cited. One paragraph of it does. Everything about optimizing for ChatGPT follows from that single fact.
Notice how many stages are just search engine mechanics: classification, metadata filtering, fetch speed. That's not a coincidence — 87% of ChatGPT's search citations overlap with Bing's top results. The retrieval pool is the search index. If you're not indexed and competitive, stages 5 through 7 never happen for you.
This is the point I keep making: getting cited by ChatGPT is a two-stage game. Classic SEO gets you into the candidate pool (stages 1–4). Passage-level structure wins the citation once you're in (stages 5–7). I break the full model down in GEO vs SEO, and if you're a SaaS founder, the SaaS-specific version is here.
Map your symptom to the stage:
The catch is that you can't diagnose any of this without seeing whether ChatGPT actually mentions you — and that data lives nowhere in your analytics. It's exactly the gap I built Citevis to close.
Citevis shows whether ChatGPT mentions your brand, on which prompts, and where competitors win instead — so you can fix the exact stage you're losing.
Track Your AI Visibility →