How to rank in ChatGPT

First, understand what you are optimising

There is no result list to climb. When someone asks ChatGPT for the best invoicing tool, a retrieval step gathers sources and the model writes a paragraph naming a handful of products. You are trying to be one of the names, and ideally one of the links.

Three states matter, and most tools collapse them into one:

• Cited — the answer links a page on your domain. This sends traffic.
• Named, not linked — the model knows you exist and typed your name, but did not have a page it wanted to link. This is the most fixable state there is.
• Absent — the buyer never sees your name.

Each engine has a different front door

They are not one audience. Getting into one is not getting into all, and knowing which pipe feeds which engine tells you where to spend a week.

• ChatGPT retrieves through a mix of Bing, its own crawler’s index, and a Google fallback. Being findable in Bing is not optional.
• Claude retrieves through Brave’s index. Brave’s crawler honours Googlebot permissions — so a robots.txt that blocks Googlebot makes you invisible to Claude as a side effect nobody intends.
• Gemini and Google AI Overviews retrieve through Google. Blocking Google-Extended removes you from Gemini’s grounding entirely.
• Perplexity is the most winnable for a young domain — it leans hardest on live retrieval and least on established authority.

Practical order for a new product: Perplexity first, then ChatGPT, then Claude and Google’s AI Mode, with AI Overviews last — Overviews stay the most rank-coupled surface, so they follow ordinary SEO rather than leading it.

What actually moves the number

1. Get mentioned somewhere that is not your website

This is the binding constraint, and it is not close. Branded mentions across the web correlate with AI visibility at roughly 0.66; backlinks manage about 0.22. Vendors typically account for only a small share of their own citations — the engines mostly cite pages about you, written by somebody else.

2. Get onto the listicles that already win your question

On commercial “best X” questions, ChatGPT cites a listicle essentially every time. Around a third of those cited lists sit on low-authority domains rather than major publishers, which means inclusion is usually a matter of asking rather than of authority. The question is which lists specifically — and that has a precise answer: whichever ones the engine cited when it answered your buyer’s question. That list is the work queue.

3. Claim the free directory profiles the engines treat as inclusion gates

G2 and Capterra profiles sit behind essentially every tool recommendation — 100% and 99% of ChatGPT-recommended tools have them. Review volume matters much less than existence; the free tier is enough to stop being structurally excluded.

4. Publish your own comparison page

A page that fairly compares you against your named competitors is the format engines reach for on commercial questions. It also gives them something linkable when they already know your name — the fix for the “named, not linked” state.

What does not work

• llms.txt. Around 97% of published files get no AI crawler requests at all, and no major engine has committed to reading one. Publish it if you like — it costs nothing — but it is not a lever. Our generator is free and says the same thing.
• Schema markup for citations. Matched-control analysis found a causal null, and Google removed FAQ rich results in May 2025. Schema is fine for what it is for; it is not an AEO strategy.
• Press releases and paid syndication. Cheap to buy, almost never cited.
• Trying to get a Wikipedia page. A company page will fail notability review unless you are genuinely notable. Wikidata is the achievable version.
• Gating your best content. A page behind a form cannot be retrieved, so it cannot be cited.

How to know whether any of it worked

Measure before you change anything, then keep measuring on a fixed cadence. Two traps make most people fool themselves here:

• A single run is noise. Ask the same question twice and the cited set moves; only about 2.2% of domains recur across three runs. One check is a snapshot, not a baseline.
• “Not cited” and “could not ask” are different facts. If a tool cannot reach an engine and records a zero, you will read a rate limit as a lost citation and go and fix something that was never broken.

That is the whole reason CiteRank asks on a weekly cycle, on every engine, and states its coverage in every report.