SEO for ChatGPT: what actually changes whether it recommends you

Short answer: There is no ranking in ChatGPT, so "SEO for ChatGPT" is a misnomer that has stuck. There is no position to hold — an answer either names you or it does not, and it may name four competitors alongside you. The levers that appear to move it, in descending order of confidence: being crawlable by the indexing bot, being present on the third-party pages the retrieval layer already trusts for your category, having comparison and alternatives coverage of your own, and structuring pages so the answer arrives in the first hundred words. Everything past that is inference, and this post is explicit about which is which.

Why the SEO metaphor breaks

Classic SEO optimizes for a slot in an ordered list. Ten results, numbered, stable enough to track daily. The whole discipline — rank tracking, position averages, CTR curves — is built on that ordering existing.

Generated answers have none of it. Ask ChatGPT which CRM a five-person startup should use and you get a paragraph naming two or three products, assembled fresh, with no ordering that persists to the next query. Ask again an hour later and the set can change. There is no rank to track because there is no rank.

What replaces it is a rate: across the questions your buyers actually ask, how often does an answer name you? That is the metric, and because the underlying answers vary, it is an estimate with real uncertainty in it rather than a number. We wrote up the statistics of that in why your AI visibility score is probably noise; the short version is that three appearances in five answers is not 60%, it is 60% with a 95% interval running from roughly 23% to 88%, and only one of those two sentences survives being asked how sure you are.

Tier 1: things we can state as fact

The indexing crawler has to be able to reach you. ChatGPT's cited answers come from a retrieval layer built by OAI-SearchBot, which is a different token from the training crawler GPTBot. If your robots.txt disallows it, you are not in the index, and nothing else on this list can help. This is verifiable in ten seconds and it is the single most common own-goal in the category — usually committed by a team that meant to opt out of training. Our free readiness checker will tell you.

Your content has to survive without JavaScript. Crawlers that build retrieval indexes are generally not executing your bundle. A page whose body renders client-side is, to that crawler, an empty page with a nav on it. Check by fetching your own URL with curl and reading what comes back.

Blocking is forward-looking, not retroactive. A Disallow added today does not withdraw content already crawled and already in an index or a corpus.

That is the honest extent of the certainties. Notice that all three are about being readable, not about being preferred — they are necessary and nowhere near sufficient.

Tier 2: things with real evidence behind them

Third-party coverage appears to matter more than your own site. In most B2B categories, when we look at what a generated answer actually cited, the sources skew heavily toward listicles, review sites, comparison posts and community threads rather than vendor homepages. This makes intuitive sense — a retrieval layer synthesizing "which tool is best" has structural reason to prefer a page that compares several to a page that sells one — and it matches what we observe across accounts. It is a strong pattern, not a published ranking factor.

The practical consequence is uncomfortable for a marketing team: a meaningful share of the work is not on your own site. Getting into the "best X for Y" roundups that already rank is closer to digital PR than to SEO.

Comparison and alternatives pages are disproportionately effective. "X vs Y" and "alternatives to X" are among the most common shapes of buyer question put to an AI, and the engine quotes whoever wrote the page. If you have not written yours, a competitor has, and theirs is the framing being read back to your prospect. This is the highest-leverage content you can write on your own domain, and most companies have not written it.

Structured data helps machines resolve who you are. JSON-LD gives an engine a machine-readable statement of what you are and what you sell rather than something to infer from your layout. We went through which types actually earn their keep in schema markup for AI search — the honest answer is that three types done properly beats twelve done partially.

Answer-first structure is quotable structure. A model extracting an answer takes the passage that answers. A page that opens with context, then narrative, then finally the answer in paragraph nine has handed the citation to whoever led with it. This is the same discipline featured snippets always rewarded, which is why the two have converged in practice.

Tier 3: what is being sold to you as fact and is not

Nobody outside these companies knows the ranking factors. No provider publishes how its retrieval layer selects and orders sources. Every confident list of "the 12 ChatGPT ranking factors" is inference from a black box, presented with more certainty than the method supports. Some of the inferences are probably right. You should still know which kind of claim you are reading.

"llms.txt will get you cited" is unestablished. The file is a proposal from 2024, not a standard, and no major provider has publicly confirmed fetching it at crawl time. Publish one — it is cheap, low-risk and forces you to decide which twenty pages matter — but do not let anyone sell it to you as a lever. We laid out the full state of adoption in what is llms.txt and does it actually work.

Publishing volume does not appear to work. More pages is the oldest reflex in SEO and the least applicable here. There is no index-size race to win, and a domain full of thin generated articles is precisely the material a retrieval layer has least reason to trust. Content mills are selling the metaphor, not the mechanism.

Your Google rank does not carry over. Ranking first for a query and being absent from the generated answer to the same question is a common and expensive combination. They are different systems with different source-selection logic. Treat your Search Console data as evidence of what people ask — which is genuinely valuable for choosing prompts — rather than as a proxy for how you are doing.

A working order of operations

1. Confirm OAI-SearchBot and PerplexityBot can reach you, and that your content is in the HTML. Ten minutes.
2. Write the comparison pages you are missing — you versus each real alternative, and the "alternatives to [category leader]" page. Concede honestly where the other product is better; a page that concedes nothing reads as an advertisement to a human and, apparently, to a model.
3. Find the pages already cited for your category's questions, and work on being present in them. This is outreach, not publishing.
4. Rewrite the openings of your highest-intent pages so the first two sentences answer the title's question outright.
5. Measure the rate, not the anecdote. One screenshot of one ChatGPT conversation is a single draw from a distribution and tells you nothing you can run a roadmap on.

That last one is the part we built EchoRanked around: run your buyers' questions across ChatGPT, Perplexity, Gemini, Copilot and Google AI Overviews on a schedule, count how often each names you, and report it with a confidence band and the verbatim answers behind every number. We publish our own results with the losses left in, because a vendor in this category asking you to trust a number should be willing to show you theirs.

What "good" looks like

There is no universal benchmark and anyone quoting one is selling. A 20% appearance rate across a crowded category with eight credible competitors may be strong; 40% on prompts that name your product is meaningless. The comparison that means something is your own trend over time, on a stable prompt set, with a band wide enough to be honest about what you do not yet know.

Frequently asked questions

Can you actually do SEO for ChatGPT?

Not in the ranking sense — there are no positions to hold. You can improve the odds that a generated answer names you, by being crawlable to the indexing bot, present in the third-party sources the retrieval layer already trusts, and covered by your own comparison pages. The metric is an appearance rate across your buyers' questions, not a rank.

How long does it take to show up in ChatGPT?

Technical fixes such as unblocking the indexing crawler can take effect within a crawl cycle. Becoming a source the retrieval layer reaches for is slower and depends largely on third-party coverage you do not directly control. Any fixed timeline you are quoted is a guess.

Does ChatGPT use my Google rankings?

No, not directly. They are separate systems with different source selection, and ranking first while being absent from the generated answer for the same query is common. Search Console is still useful here — as evidence of the questions people ask, which is how you choose what to measure.

Is GEO the same as SEO for ChatGPT?

GEO — generative engine optimization — is the broader term, covering being cited or recommended across every generative engine rather than one product. "SEO for ChatGPT" is the same idea narrowed to a single engine, and the levers are largely shared.

Where to go next

Start by confirming the crawlers can reach you at all: how to check if ChatGPT can read your website. If you are about to make blocking decisions, the complete AI crawler list covers which token does which job. And if you want to understand what the engine is actually drawing on when it answers, where ChatGPT gets its information separates the three pipelines.