How to rank in ChatGPT: what works and what doesn't

When we first wrote about this in May 2024, the short answer was that ChatGPT wasn’t searching anything. It answered with whatever had stuck inside the model during training, which is why what it said about your sector could be eighteen months out of date.

That’s no longer true, and the change is big enough that we rewrote the article from scratch. ChatGPT now searches the web while it answers, cites the sources it consulted and links to them. The question has stopped being “how do I get into the model’s training data”.

What “ranking in ChatGPT” means, and why it isn’t SEO

Two things get mixed up constantly, and separating them is most of the work.

When someone asks ChatGPT “which SEO agency should I use for a B2B SaaS?”, two different things can happen. The model answers from memory, with what it learned during training. Or it decides to search, reads a handful of pages and builds the answer from what it just read, citing where it came from.

Two mechanisms, two sets of rules. Confusing them sends you working in the wrong direction.

Model memory forms during training, happens every many months, and nobody outside decides what gets in. You don’t optimise it, it accumulates. If your brand has spent years being mentioned in places the training absorbed, you’re in. If you’re new, you won’t be, no matter how much you polish your website.

Live search happens at the moment of answering. ChatGPT runs queries, pulls pages and leans on them. That one you can influence, because what it consults is the web as it stands today.

Almost everything actionable sits in the second mechanism. The good news is that it’s also the one that moves in weeks rather than years.

Where ChatGPT pulls its recommendations from

When it searches, ChatGPT doesn’t crawl the whole internet in that instant. It leans on an index and on search results, picks a handful of pages and reads those. The practical consequence is uncomfortable to accept:

It’s almost never reading you. When someone asks about the best agencies in a sector, the pages ChatGPT opens tend to be comparison posts, press articles, industry roundups, forums and Reddit threads. Your brand shows up if one of those pages names you.

That explains why brands with immaculate websites never appear, and brands with mediocre ones appear constantly. The difference isn’t the website. It’s how many people talk about them in the places the model goes looking.

There’s a second route, slower but sturdier: being the source yourself. That happens when you answer a specific question better than anyone else, and in a way that can be quoted without the surrounding context. That work looks a lot like classic SEO, with one important twist we get to below.

What actually moves the needle

Ordered by what we’ve seen carry the most weight, not by what’s easiest to sell.

Other people naming you on the pages ChatGPT consults

First by a distance. If your category has comparisons, rankings and roundups, being in them matters more than anything you do on your own domain. This isn’t classic link building: you’re not after the link, you’re after the mention in a context where the question is being answered. An unlinked mention in an article ChatGPT opens is worth more than a link on a site it never consults.

This collides head-on with how most content strategies are built, looking at the company’s own website and little else.

Answering whole questions, in paragraphs that can be quoted

When ChatGPT builds an answer, it extracts fragments. A paragraph that answers a complete question, carrying its own context, has far better odds of ending up cited than the same content spread across five sections that only make sense if you’ve read the previous ones.

In practice that means question-shaped headings, the direct answer right underneath before the detail, and concrete data with its source. It’s the difference between writing for someone reading top to bottom and writing so a machine can cut out a piece and use it.

Describing yourself consistently

Models build an idea of what you are from what they find repeated. If your website says “digital marketing agency”, LinkedIn says “growth consultancy” and your press releases say “technology partner”, you’re splitting the signal three ways. Picking one description and holding it everywhere sounds like textbook advice, but it’s one of the few things on this list that depends only on you.

Being readable to the bot

Necessary, not sufficient. If your critical content is painted with JavaScript on the client, there is a good chance the bot never sees it. If your robots.txt blocks GPTBot or OAI-SearchBot, you’ve excluded yourself outright. Worth checking before investing in any of the above, because it’s cheap and because finding it blocked without anyone having decided to block it is not rare.

What doesn’t work, and still gets sold

Hammering the brand name into the text. Models don’t count keyword density. Repeating your name twenty times doesn’t raise the odds of being cited. It makes the text worse, which lowers them.

“Optimised prompts” hidden in the page. The idea of hiding instructions aimed at the model inside your HTML does the rounds periodically. Beyond being something OpenAI has every interest in neutralising, it rests on a wrong reading of how retrieval works: the model doesn’t obey the pages it reads, it summarises them.

Publishing a lot, fast. Volume without authority doesn’t enter the conversation. Twenty mediocre articles a month don’t achieve what one article that becomes the reference for a specific question does.

Waiting to “get into the training data”. That’s a goal you control neither the timing nor the outcome of. Working towards it means working without being able to measure.

How to measure whether you’re showing up

Here’s a real difference from SEO: there is no Search Console for ChatGPT. No dashboard tells you which answers you appear in. What you can do, today and for free, is look from three angles.

Your server logs. The most reliable source, because it records what actually happened. The agents worth watching are GPTBot, which collects content, and OAI-SearchBot, which is the one that goes looking when someone asks. Seeing the second hit a specific URL is the most direct signal that the page is being used to build answers.

Referral traffic. When ChatGPT cites a source, it links to it, and that visit lands in your analytics with chatgpt.com as the origin. The volume is small next to Google, and that’s exactly why it’s worth watching: each visit stands for many answers where you appeared and nobody clicked.

Asking it yourself. Inelegant and very useful. Run the ten questions a customer of yours would ask, note them in a sheet with the date, and repeat monthly. Two months in you have a series worth more than any estimate. We wrote up the full method in how to check if your business shows up in ChatGPT.

Paid tools that track brand mentions across models exist and work reasonably well, but you don’t need one to start. Measuring isn’t the expensive part. Never having looked is.

Where to start from zero

An order that works with clients arriving with none of this in place:

1. Check you’re not blocking the bots. Ten minutes, and sometimes it solves the whole problem.
2. Run the ten-question round and save it. Without a baseline you won’t know whether anything worked.
3. Look at which pages show up today when your category gets asked about, and list where you should be named and aren’t.
4. Rewrite your two or three most important pages so they answer whole questions instead of describing services.
5. Work the mentions from that list in step three. This is the slow part and the one that weighs most.

The first four an internal team can handle. The fifth is usually where help is needed, because it takes relationships with press and with whoever publishes the comparisons in your sector.

Frequently asked questions

We can help you get there

This is what we work on in our GEO service: getting your brand cited when someone asks about your category, starting with a diagnosis of where you stand today.

If you want to know your starting point, tell us about your project and we’ll look at it with you.