How do you get recommended by ChatGPT?

1. Be readable — necessary, not sufficient

If the systems cannot read your pages, nothing else can help. Three failures are common.

Client-side rendering. In December 2024 Vercel reported that, in its observations, GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot and PerplexityBot did not execute JavaScript. If essential content is assembled in the browser, some crawlers may miss it. Test by disabling JavaScript and reloading — if essential content disappears, fix that before anything else on this list.

Crawler access. Check what your robots.txt actually serves, not what is in your repository. Hosting providers and CDNs increasingly inject managed rules that block AI crawlers by default, and the file you wrote may not be the file being served.

Entity clarity. One consistent company name, one canonical description, coherent Organization markup, and consistent details across the site and any external profiles. Systems resolve entities; ambiguity costs you.

This layer is a floor. Completing it does not make you recommendable — it makes you eligible.

2. Say something a competitor cannot say

This is where much of the outcome is decided, and it is not a technical task. If your description is "AI-powered platform for enterprise", a model has no basis to prefer you and every reason to produce the category sentence with your name attached.

A useful test: write the claim, then substitute your closest competitor’s name. If it reads as equally true and equally relevant, it is not yet doing positioning work. The fix is rarely a claim of total uniqueness; it is usually a more specific buyer, workflow or piece of evidence — or a combination of capabilities that matters to that buyer.

Specificity means nouns a competitor cannot borrow: a named method, a defined segment, a measurable threshold, a certification, a constraint you accept that others do not. Adjectives are free, which is exactly why they carry no information.

3. Get it corroborated — the part that decides it

Self-description is the weakest kind of evidence — for buyers, and plausibly for models, since everybody praises themselves. Agreement between independent sources is much stronger.

So the programme is: get the specific claim repeated, accurately, by publications, analysts, communities, comparison sites, partners, customers and the catalogues that list your category. Not with your marketing language — with the claim itself, stated in their words.

This is slow. It is also the reason the outcome is defensible: no software product performs it, and it cannot be bought convincingly. Purchased mentions and syndicated filler are detectable by anyone paying attention, and a record built from them is a liability with a delay on it.

4. Publish things worth citing

The most efficient way to get corroborated is to publish work other people need to reference: original research, real benchmarks, a documented method, data nobody else has. When someone cites it, your claim travels with the citation and you did not have to ask.

This is also the one asset in this market that cannot be counterfeited. In a category full of manufactured proof, a study someone can check is worth more than a page of logos.

5. Measure, and be honest about variance

Sample the answers deliberately, on a schedule, with the same question set — because absence produces no signal anywhere else. Run each question several times: these systems vary between sessions, and a single run is an anecdote.

Then accept the limits. You will not be able to attribute revenue to a change in an AI answer with any rigour, and a vendor who offers to is guessing.