How to rank in ChatGPT

AI search·13 minutes·26 October 2026

How ChatGPT decides which products to name

Ranking in ChatGPT does not mean what it means in Google. There is no list of ten results and no position to hold. When a buyer asks which tool to pick, ChatGPT names three or four products in a paragraph, and the question is whether yours is one of them.

For questions about current products and vendors, ChatGPT usually searches the web before it answers. It writes its own search queries, often a rephrased version of the question with “best” or the current year attached, reads a handful of the pages that come back, and summarises them. The names in the answer are, for the most part, the names that appear on those pages.

That gives you two ways in. Your own pages can be among the ones it reads, or your product can be named on the third-party pages it reads. In the categories we have measured, the second matters far more than the first: in our study of 152 cited sources in ITSM software and proxy servers, not one of them sat on the named vendor’s own website.

Two different outcomes are worth separating. A mention is your product named in the answer. A citation is one of your pages linked as a source. Buyers act on mentions; citations are how you find out which pages the model trusted.

Step 1. Measure where you stand

Before changing anything, write down the questions your buyers actually ask and run them. Not “what is ITSM”, but “which ITSM tool should a 300-person company use”, “ServiceNow alternatives for mid-market”, “cheapest residential proxies for scraping”. We use a set of two hundred, because a smaller sample swings too much from week to week to tell a real change from noise.

For each answer, record four things: whether you are named, in which position, which competitors are named instead, and which sources are cited. Position matters more than it looks. Being named first and being listed fifth after the main recommendation are different outcomes, and almost nobody reads past the third name.

Then keep the set fixed. Re-run the same questions every quarter. If you change the sample between runs, the chart shows a change in method rather than a change in result. When we ran this for the two clients in our proxy and ITAM/ITSM case studies, both started from the same place: named in none of the answers.

Step 2. Collect the sources it cites for your competitors

The baseline tells you who gets named. The sources tell you why. Take the answers where competitors are named and you are not, and list every page cited in them. This is the single most useful document in the whole exercise, because it turns “be more visible in AI” into a list of specific pages with specific owners.

Expect the list to be long and unglamorous. In our study, the 152 citations pointed to 141 different pages on 131 domains, and only a handful were cited for more than one vendor. Most of them were “best X” rankings: 90 percent in ITSM, about half in proxies. And most were not written by media. In ITSM, 84 percent sat on the blogs of other software companies writing comparison lists for their own category.

You cannot guess this list from a keyword tool. Two pages with similar rankings in Google can play completely different roles in an AI answer. The only reliable way is to read the citations in real answers, competitor by competitor.

Step 3. Get onto the pages ChatGPT already reads

This is where most of the result comes from, and it is closer to link building than to content writing. For every page on the list, work out what it takes to appear on it. In our data, five routes covered almost everything:

Ask the publisher. For most rankings, the first step is a message asking what inclusion requires. Some vendors add credible products for free, some want a review account, some charge. In ITSM this was the route for 82 of 90 pages.

A mention inside an existing article. Where a page discusses the topic without a list, a short, accurate mention with a link can be added to a paragraph that already exists. This covered eleven of the proxy sources.

An affiliate programme. In niches where comparison sites earn commission, as proxies do, those sites rank the providers who pay them. Without a programme of your own, you are absent from a third of the sources.

A profile on review and directory sites. Aggregators list providers that have a profile page. Getting one is usually a form and some patience.

A sponsor block. A few homepages in the proxy set carried paid placements that the assistants read like any other text.

Two rules apply to all five. Be accurate: a publisher who adds you on the strength of claims that turn out to be inflated will remove you, and the assistant will repeat whatever is wrong. And prioritise pages cited for several competitors first, because a single placement there shows up in several answers at once.

Step 4. Make your own pages easy to quote

Your own site still matters, for two reasons. It is the page the assistant reads when someone asks about you by name, and it is what the publishers in step 3 check before they write a line about you. A model lifts a paragraph, not a page, so the structure of the text decides whether there is anything to quote.

What works: a direct answer in the first paragraph, with the claim before the explanation. Concrete figures rather than adjectives: prices, limits, setup time, supported systems. Comparison tables with named alternatives. References to sources a reader can check. What does not work: a positioning paragraph above the answer, feature lists without numbers and pages where the price says “contact sales”.

One page type deserves special attention. Nearly every source in our study was a list with named alternatives. If you sell in a category, publish a fair comparison of that category on your own site, including the rows where a competitor wins. It is the page type assistants are built to read, and it is also the one that converts best in classic search, as we covered in the SaaS SEO strategy guide.

Step 5. Make sure ChatGPT can read you at all

The technical part is short but it fails more often than you would expect. Check three things.

robots.txt allows OpenAI’s crawlers. OpenAI uses separate user agents for search (OAI-SearchBot), for fetching pages a user asks about (ChatGPT-User) and for training (GPTBot). Blocking OAI-SearchBot keeps your pages out of ChatGPT’s search answers. Some security plugins and CDN bot rules block all three without anyone noticing.

The content is in the HTML. If your prices, comparison tables or plan limits load only through JavaScript, assume a crawler may not see them. View the page source, not the rendered page, and check that the facts you want quoted are there.

Your facts agree with each other. If the pricing page, the docs and a three-year-old press release give different numbers, the model has no way to know which one is current. Pick one source of truth and update the rest.

And llms.txt: add it, because it costs nothing, but do not expect it to change the answer. There is no settled standard, and most platforms do not use it. Content and third-party mentions decide this.

How long it takes

Faster than classic SEO, and less predictable. In our proxy case study, AI work started in month sixteen of the engagement. The first mentions appeared six weeks later, and eight months in the client had 486 mentions in AI answers, from zero. In the ITAM/ITSM case, four months after starting the client was named in 31 percent of answers across the fixed query set.

ChatGPT is usually the slowest of the major assistants to move, because it leans heavily on external roundups and directories, and getting onto those takes months rather than weeks. Perplexity tends to pick up changes first. Plan for a first signal at six to eight weeks and a measurable share at four to six months.

One honest caveat: the volume is still smaller than search. Being named in AI answers is worth a lot per visit, because the person asking is choosing a vendor, but it supplements organic search rather than replacing it.

What does not work

Writing for the prompt. Stuffing a page with “best tool for X” phrasing does nothing. The model does not match your text against the question; it reads the pages its own search returned.

Fake reviews and planted forum threads. They are easy to spot, platforms remove them, and an assistant that picks one up will quote it alongside the replies calling it fake.

Mass placements on lists nobody reads. A page that exists only to sell slots rarely ranks for anything, so the assistant’s own search never finds it. The pages that matter are the ones already cited, which is why step 2 comes before step 3.

Waiting for training data. Being in a model’s training set helps with brand recognition, but for questions about current vendors the answer is built from live search. That is good news: it means the result can be changed this quarter.