Getting Found in ChatGPT Search

A few years ago, "getting found" meant ranking a blue link. Now a large share of the questions your customers ask never reach a list of links at all — they get typed into ChatGPT, which reads the web, writes the answer, and decides which two or three sources to name. If you're not one of those sources, you don't exist in that moment. This is the single biggest distribution shift I've watched as an operator, and ChatGPT is the front door for a huge slice of it.

I build the engine that does this for a living — Otto, the system behind RunOctopus — so what follows is the mechanics as I actually understand them from shipping against ChatGPT's real behavior, not a list of theories. By the end you'll know how ChatGPT's search surfaces sources, what it takes to be the page it quotes, and how to check whether any of it is working.

When ChatGPT searches versus when it doesn't

The first thing to internalize: ChatGPT does not search on every question. OpenAI's models — currently the GPT family powering ChatGPT — answer a lot of prompts straight from their own training knowledge with no live retrieval and no citations. That happens on timeless, general, or definitional questions where the model is confident and the answer doesn't change day to day.

Search fires when the question is time-sensitive, local, specific, product-shaped, or news-shaped — anything where the model "knows" its memory might be stale or thin. "What's the best way to do X right now," "compare these two tools," "who offers Y near me," "what happened with Z this week." Those route to a retrieval step that pulls live pages and attaches sources.

For AEO this distinction is everything, because you can only get cited on the queries that trigger search. So the first move isn't optimizing a page — it's identifying which of your customers' real questions actually fire retrieval. The way to know is to ask them in ChatGPT and watch whether a sources panel appears. Questions that surface live citations are your battlefield. Questions answered from memory are a different, slower game about long-term authority and being mentioned across the web enough that the model absorbs you.

How ChatGPT pulls and chooses sources

When search does fire, the rough pipeline looks like this. ChatGPT reformulates your question into one or more search queries, runs them against a live web index, pulls back a set of candidate pages, fetches and reads the most promising ones, and then writes an answer grounded in those passages with citations attached. The model is doing extraction and synthesis on top of retrieval — it's not just handing you a ranked list, it's reading the pages and composing a paragraph from what it found.

Three checkpoints decide whether your page makes it into the answer:

Retrievability. If OpenAI's fetcher can't read your page, nothing else matters. The page has to be crawlable, indexable, and served as real text — not locked behind aggressive JavaScript that returns an empty shell, not gated, not blocked in robots.txt for the relevant user agent. OpenAI runs distinct crawlers for training versus live search retrieval; the one that matters for citations is the user-facing fetch that happens when an answer is being written. Block that and you've opted out of the answer.

Relevance to the reformulated query. ChatGPT doesn't search your exact words — it rewrites the question into cleaner queries first. So your page needs to match the intent behind a question, phrased the way a search system would phrase it. A page titled around the actual question, with the question's natural language in the headings and body, lines up with that reformulation. A page that buries the answer under brand copy and jargon doesn't.

Quotability. This is the one most sites miss. The model has to be able to lift a clean, specific, self-contained claim out of your page and drop it into the answer without hedging. If your answer is spread across three paragraphs, wrapped in qualifiers, or hidden below a wall of intro, the model can't extract it cleanly and reaches for a competitor who stated it plainly. The pages that win are the ones that answer the question in the first sentence or two, in a way that survives being quoted out of context.

Be the answer, not just a result

Here's the operator's reframe: you're not writing for a reader who'll scroll your whole page. You're writing for a model that will read your page, find the one passage that answers the question, and quote it. Optimize for that extraction.

Lead with the answer. Answer-first structure beats build-up every time. State the direct answer in the opening, then expand with the why, the nuance, and the supporting detail underneath. The model gets its quotable claim immediately, and human readers who landed from a citation get what they came for.

Make claims specific and self-contained. "It depends on your situation" is unquotable. "For a business under X conditions, the answer is Y, because Z" is quotable. Specific, bounded, standalone claims are what get extracted. Vague hedges get skipped.

Match the real questions. Use the actual phrasing your customers use, including the long, natural-language form of the question. Put it in a heading. The closer your page maps to a reformulated search query, the more likely it's retrieved and the more confidently it's quoted.

Cover the question completely. ChatGPT favors sources that fully resolve the query so it doesn't have to stitch together fragments. A page that answers the main question and the obvious follow-ups in one place is more valuable to the model than a thin page it has to supplement.

Earn corroboration. ChatGPT leans toward claims it can see echoed across multiple credible sources. Being right and being referenced elsewhere makes the model more comfortable citing you. This is where classic authority work — real expertise, being mentioned and linked across the web — feeds directly into AEO. The two disciplines share a spine, and I dug into the structural side of getting cited in Get Cited by AI Search.

Keep it current and dated. Live retrieval favors fresh, clearly dated pages on anything time-sensitive. A visible "updated" date and genuinely current content signal that you're a safe source for a question where staleness is a risk. Specifics here move fast, so I treat freshness as ongoing maintenance, not a one-time publish.

The structural layer: markup and crawlability

Good content is necessary but the plumbing has to be right too. Clean, semantic HTML with real headings beats a page where everything is a styled div — the model reads structure to find the answer. Server-rendered text the fetcher can read in the initial response is far safer than content that only appears after client-side rendering. JSON-LD structured data (FAQPage, Article, Organization) makes your claims and your entity unambiguous to machines; it won't rescue thin content, but on a genuinely useful page it removes friction between your answer and the engine extracting it.

And the boring one that sinks people: check your robots.txt and your crawler access. If you've blocked the user agent ChatGPT uses to fetch sources at answer time, you've quietly opted out of every citation. The full structural checklist — markup, crawlability, answer-first patterns, measurement — is what I laid out in the Answer Engine Optimization Playbook, and ChatGPT is one of the engines it targets directly.

Track whether it's actually working

You can't manage what you don't measure, and the right measurement for ChatGPT visibility is refreshingly concrete: test the real prompts.

Build a fixed list of the actual questions your customers ask — twenty to fifty is plenty to start. Run each one in ChatGPT with search enabled. Log whether your domain appears in the cited sources, and if so, what claim it's cited for. That gives you a prompt-coverage rate: out of the questions that matter, how many cite you. Re-run the same list on a schedule, monthly for most businesses, and watch the rate move as you ship pages.

Two things make this discipline pay off. First, keep the prompt list fixed so month-over-month comparisons mean something — changing the prompts every time turns your scoreboard into noise. Second, log the competitor who got cited when you didn't, because that tells you exactly which page you're trying to beat and what claim you need to own. I treat this as the core AEO scoreboard across every engine, and walked through the full approach in Track Your AI Visibility.

The honest caveat: ChatGPT's behavior is not deterministic. The same prompt can route to search one day and memory the next, and the cited sources can shift run to run. Don't over-fit to a single answer. Track the trend across a fixed prompt list over time, and let the direction — not any one screenshot — tell you whether the work is landing.

Where to start tomorrow

If you do nothing else: pick your ten highest-value customer questions, ask each one in ChatGPT with search on, and note which trigger citations and who gets cited. For the ones where you're missing, ship a tight, answer-first page that states the answer in the first sentence, matches the question's real phrasing, and is fully crawlable. Then re-run the list next month. That loop — find the queries that search, own the answer, measure the pickup — is the whole game, and it compounds.

The shift from links to answers isn't something to fear or fight. It's a new front door, and the businesses that show up at it deliberately are going to own a lopsided share of the attention behind it.

FAQ

Does ChatGPT actually cite websites now?

Yes. When ChatGPT runs a web search or browses to answer a question, it pulls a handful of live sources and attaches citations to the synthesized answer — a source name and a link you can click. Not every answer triggers search; timeless or general questions get answered from the model's own knowledge with no sources. But anything time-sensitive, local, product-related, or news-shaped tends to fire a retrieval step, and that's where citations come from.

How do I get ChatGPT to cite my site?

Be retrievable, be relevant, and be quotable. Retrievable means OpenAI's crawler can read your page and it's indexable. Relevant means the page directly answers the exact question being asked, not a loosely related topic. Quotable means the answer sits in a clean, specific, self-contained passage near the top — a sentence the model can lift without hedging. Miss any one of the three and you're invisible in that answer.

What's the difference between ChatGPT's training knowledge and its search?

Two separate paths. Training knowledge is what the model absorbed up to its cutoff — it has no live link and moves on a slow retrain cycle. Search is a live retrieval step that pulls current pages and cites them. For AEO you mostly target the search path, because it's the one that reads your page today and names you as a source. Earning your way into training memory is a slower, authority-driven game.

Should I block OpenAI's crawler to protect my content?

If you want ChatGPT to cite you, you can't block the crawler that lets it read and surface you. OpenAI uses separate user agents for training versus live search retrieval. Blocking the search/user-facing fetcher in robots.txt removes you from the answers entirely. Blocking the training crawler is a real choice some publishers make, but understand the tradeoff: visibility in the answer requires being fetchable when the answer is written.

How fast does ChatGPT pick up a new page?

On the live-search path, fast — once the page is crawlable and the query routes to search, a strong new page can show up as a cited source within days. The slow part is authority: ChatGPT leans toward sources it can corroborate, so a brand-new domain with one page competes harder than an established site with a track record. Publish, make it retrievable, then test the real prompts.

How do I track whether ChatGPT is citing me?

Test it directly. Build a fixed list of the real questions your customers ask, run them in ChatGPT with search on, and log which answers cite your domain. Re-run the same list on a schedule — monthly is plenty for most businesses. That prompt-coverage rate, tracked over time, is your real scoreboard. Screenshots and dashboards are nice; the discipline of the same prompt list every month is what actually tells you if you're winning.