ChatGPT citations vs Google rankings: what carries over and what doesn't

Somewhere in your analytics is a page that ranks third for a question your customers ask. Ask ChatGPT the same question with search on, and the answer quotes two roundups, a Reddit thread, and a competitor's comparison page sitting around position fifteen. Your page was found. It was not used.

That gap confuses people who have spent a decade on SEO, because it looks like the rules changed. They did not change so much as split in two.

Key takeaways

An assistant does two things where a search engine does one: it searches, then it reads. Ranking gets your page into the candidate set; being quotable decides whether it survives the read. Everything that makes a page findable still matters. Everything that makes a page rank without stating a plain answer no longer pays. And the assistant rarely searches the keyword you optimised for, because it rewrites the question first.

One step became two

A search engine returns a list. The ranking is the output, and the click is yours to earn.

An assistant answering a question with web search runs a different sequence. It rewrites your customer's question into one or more search queries of its own. It fetches a handful of results. It reads them — usually the text it can extract quickly, not the rendered page a human sees. Then it writes an answer from the fragments it can actually quote, and cites what it used.

Two filters, not one. The first is retrieval and looks a lot like SEO. The second is extraction, and has almost no equivalent in ranking. A page can pass the first and fail the second, which is exactly what a #3 ranking with no citation is: found, opened, and found unquotable. Our pillar on how ChatGPT chooses sources walks through the whole sequence in more detail.

What carries over from SEO

Most of the unglamorous half.

• Indexability. If crawlers cannot fetch it, nothing downstream happens. That includes the AI crawlers specifically: GPTBot, OAI-SearchBot, PerplexityBot and friends are separate user agents and are commonly blocked by accident. Our robots.txt checker reads yours in a second.
• Server-rendered text. Retrieval fetches HTML. A page whose content arrives only after JavaScript runs is, to a fetcher, close to empty. This is the single most common technical cause we see, and it is invisible in a browser.
• Titles and headings that match the question. Both systems use them to decide what a page is about. A heading phrased as the customer's question is the cheapest fix on this list.
• Topical depth across the site. A site with a pricing page, a docs section, an about page and comparison pages is easier to place in a market than a single marketing page, for a ranking algorithm and a retrieval index alike.
• Being on the sources that rank. The roundups and directories that occupy the top ten for "best X" are the same pages the assistant fetches. Getting listed there was always good SEO; it is now the fastest route into an answer you cannot otherwise reach.
• Freshness signals. A visible date and recent updates help in both places, and a stale page is discounted in both.

If you have done this work, you have not wasted it. You have earned candidacy.

What does not carry over

• Winning on aggregate authority. Between two pages already in the candidate set, the one that states the answer in plain sentences tends to get quoted, not the one from the bigger domain. Authority helps you get retrieved. It does not make an unquotable page quotable.
• Ranking without answering. The classic ranking page — long, keyword-rich, the actual answer implied rather than stated — is the exact shape that fails the read step. There is nothing to lift.
• Facts in images. A price in a graphic, hours in a banner, a comparison table rendered as a PNG: all invisible. Every fact you want repeated has to exist as text.
• Interstitials and gates. A cookie wall, an age gate or a "subscribe to continue" overlay can be the whole page as far as a fetcher is concerned.
• Your keyword targeting. This is the one that surprises people most, so it gets its own section.

The assistant does not search your keyword

Your page targets "invoicing software for freelancers". Your customer types "I'm a freelance designer in Berlin, what should I use to invoice clients in other countries?" The assistant does not paste that into a search box. It composes its own queries — often several, often narrower, often including constraints your keyword research never contained: the country, the currency, the integration, the price ceiling.

So the retrieval set for a real customer question is frequently not the SERP you have been optimising against. It is the SERP for a question adjacent to yours. This is why a page that ranks well can be absent from answers entirely: it never entered the race that was actually run.

The practical consequence: stop optimising only for the head term and start covering the constraints. A page that says, in sentences, who it is for, what it costs, which countries and currencies it handles, and what it does not do, is retrievable for a much wider set of rewritten queries than a page optimised for three words.

Why a page on the second page of Google gets quoted

The pages that survive the read step tend to share a shape:

1. One page, one question. The title is the question. The first paragraph is the answer. Everything else is support.
2. Facts as sentences. "Ledgerly is £12 a month for one user, with unlimited invoices and Stripe payouts in 30 countries." That sentence can be lifted whole. A pricing table with the number in a cell often cannot.
3. Specific over comprehensive. The narrow page that answers the constraint beats the exhaustive guide that mentions it in passing, because the assistant is matching a rewritten, constrained query.
4. Third-party framing. Comparison and roundup pages get cited disproportionately because they answer a decision question directly, and because an assistant assembling a recommendation prefers a source that already weighed options.

That is also the uncomfortable answer to why competitors appear in answers where they do not outrank you: they are usually not winning on their own site at all. They are being quoted from someone else's page. We broke down the seven causes in why ChatGPT cites your competitor and not you.

Compare the two lists yourself

This takes an afternoon and settles the argument for your own site.

1. Pick ten questions your customers ask before they know your name.
2. For each, record the Google top ten — just the domains.
3. Ask the same question in a fresh ChatGPT chat with search on, and record the domains it cites, plus whether you were named.
4. Put the two lists side by side.

Three things fall out. The overlap between the lists is partial, and the non-overlap is where your assumptions are wrong. The cited set leans towards roundups, community threads and directories rather than vendors' own pages. And where you rank but are not cited, you now have a specific page to fix rather than a vague worry — usually one that never states the answer in a liftable sentence.

The method for running this repeatedly, and the four numbers worth writing down each week, are in how to track ChatGPT citations. If you would rather see it for your own domain without the spreadsheet, our free scan asks twelve of these questions and shows you the answers verbatim, with the cited domains next to each one, and the benchmarks show which sources decide answers in a given market.

What to do with the difference

Three moves, in order, once you can see both lists:

Make your ranking pages quotable. Start with pages that already rank in the top ten for a question you care about. Add the direct answer as the first paragraph. Convert facts trapped in images or tables into sentences. Give the page a heading in the customer's words. These pages have already passed retrieval; you are only fixing the read step, which is the cheapest work available to you.

Add pages for the constraints. One page per real buying question — the integration, the country, the price bracket, the alternative you are compared against. Short, specific, dated.

Get onto the sources that get cited. For the questions where nothing of yours appears at all, the route in is the roundup, the directory or the thread that does. That is slower and mostly not a writing task; the citation checklist covers what each of those sources needs from you.

None of this replaces SEO. Ranking is still how you enter the candidate set, and a page nobody can find will not be quoted by anything. But ranking is now the qualifier, not the prize. The prize goes to whoever wrote the sentence the assistant could lift.