How to Get Your Business Mentioned in ChatGPT and Perplexity

There is a large and growing body of advice on this question, and most of it shares one flaw: it cannot tell you whether any of it worked.

That is the thing to fix first. You cannot improve what you have not measured, and the measurement here is simpler than the tactics. It is your appearance rate: of the questions a buyer would actually ask, how many produce an answer that names you.

Everything below is ordered by how verifiable it is. The first two sections you can check yourself today. The third is mechanism you can reason about. The last is the part nobody can honestly promise, and it is worth knowing which is which.

Measure your appearance rate first

Write down between ten and twenty questions a real buyer would type. Not keywords — questions, in the words they would use. For a New Zealand service business that looks like "who are the best X in Auckland", "should I use X or Y for Z", "how much does X cost in New Zealand".

Then, for each one:

1. Ask it in ChatGPT, Perplexity and Google’s AI results.
2. Record whether your business is named in the answer.
3. Record every source the answer cited, not just whether you were one of them.
4. Record the date.

Your appearance rate is the count of answers that named you, over the count of questions you got a usable answer to. Keep those two numbers separate, because a question that failed to return an answer is not a question you lost — it is a question you did not measure, and averaging it in as a zero is how a measurement quietly becomes a fiction.

The citation list is the more useful half. It tells you who currently owns each answer, and whether the field is concentrated in a few sources or spread thin across many. A question where thirty different domains are each cited once is a question nobody owns. That is where a good page can win. A question dominated by the vendor being discussed is one you are unlikely to take.

Re-run the same questions monthly, with the same wording. Changing the wording between runs makes the comparison worthless.

Check that the crawlers can reach you

This is the most concrete thing on the list and the most commonly skipped. If the systems that build these answers cannot fetch your pages, nothing else you do matters.

Each of them identifies itself with a user agent, and your robots.txt decides whether it is welcome. The robots.txt specification is old and simple, and the trap in it is specific: a file that names a few search engines, allows them, and then ends with User-agent: * / Disallow: / blocks every AI crawler without ever naming one. Plenty of sites are in exactly that state without knowing it.

The agents worth naming explicitly:

OpenAI documents its crawlers, Perplexity documents its own, and Google documents Google-Extended separately from Googlebot — which is the part people miss, because allowing Googlebot does not allow Google-Extended. Common Crawl is not an answer engine at all; it is a public corpus, which makes it a slower and more permanent decision than the others.

This is a deliberate choice with a real trade-off. Allowing these crawlers means your content may be used in training. Blocking them means you are not in the answers. There is no configuration that gets one without the other, and anyone telling you otherwise is selling something.

Second, check that your pages are readable without JavaScript. Fetch your own page with curl and look at what comes back. If the body arrives empty and the content is assembled in the browser, some crawlers will see nothing.

Make the answer easy to extract

Now the mechanism, which is reasonable but less directly verifiable than the two sections above.

Answer in the opening. These systems quote from early in a document far more often than from late in it. A page that spends three paragraphs on context before answering the question in its own title has put the quotable part out of reach. Put the direct answer first and the reasoning after it.

Use structured data, correctly. Schema.org markup describes what a page is about in a form a machine does not have to infer. Google’s introduction to structured data is the clearest primer. An Organization block that states who you are, and an FAQPage block on pages that genuinely answer questions, are the two worth doing. Invalid markup is worse than none, so validate it.

Be specific about place and constraint. Across the questions we measure for New Zealand automation buyers, the citation lists are almost entirely offshore — a pattern visible in the measurement itself, not an assumption. If nobody has written the New Zealand answer to a question New Zealanders are asking, that gap is the opening, and a genuinely local answer has very little competition in it.

Write the page the question deserves. If the question is a choice between three tools, a page that lists the strengths of all three does not answer it. A page that resolves to one and says why the others are out does.

Read the citation list, not just your own name

Most people run the measurement, see their name is absent, and stop. The citation list is where the actionable part is, and it answers three questions your own absence cannot.

Is the field concentrated or fragmented? Divide total citations by distinct domains. Close to one means each source was cited once and nobody owns the answer — the field is wide open and a better page can take it. Two or three and above means a handful of sources are cited repeatedly, and you are trying to displace an incumbent rather than fill a vacuum.

Is the most-cited source the vendor being discussed? If the question is about a product and the product’s own site is the top citation, that is close to unwinnable, and the sensible response is to pick a different question rather than a better article.

How many of the cited domains serve your market? This is the one that matters most for a New Zealand business, and it is the easiest to check: count the .nz and .au domains. Where that count is zero across dozens of citations, the entire answer is being assembled from sources written for somewhere else — which is an opening, and a specific one.

Do this per question rather than in aggregate. A business can be locked out of one question and have a clear run at the one next to it, and an averaged figure hides exactly that.

Mentioned, cited, and linked are three different things

Worth separating, because they have different causes and different fixes.

Mentioned means the answer named you in prose with no source attached. Pleasant, hard to attribute, and it usually means the model knows of you from training rather than from a live fetch.

Cited means the answer names a source, and the source is you. This is the one worth optimising for: it survives the model being retrained, because it comes from a retrieval step rather than from memory.

Linked means the reader can click through. Whether a citation is clickable is the answer engine’s decision, not yours.

Record all three separately. Collapsing them into "did we appear" makes a training-data mention look identical to a live citation, and only one of those is something your content can influence this quarter.

On llms.txt

You will be told to add an /llms.txt file. Here is the honest position.

It is a proposal, published in 2024, for a file that helps agents navigate a site. Several documentation sites publish one — OpenAI’s and Perplexity’s own documentation both reference theirs. That is evidence that vendors publish the file for their docs, which is not the same as evidence that their crawlers read yours.

There is no published commitment from OpenAI, Perplexity or Google to consume an llms.txt on your site, and no measurement in the public domain showing it changes whether you get cited. It is cheap and harmless, so add one if you like. Do not add it instead of the two sections above, and be suspicious of anyone presenting it as the mechanism.

A sitemap.xml is a different matter, because the sitemap protocol is long-established and documented by Google as something its crawlers use. Keep the lastmod dates truthful. A sitemap claiming every page changed today is a sitemap that gets discounted.

What nobody can promise

Three honest limits.

There is no submission form. You cannot register with ChatGPT or Perplexity the way you can with a directory. What you can do is be crawlable, be clear, and be the best available answer — and then wait, because indexes lag by weeks. Judging a new page’s performance before it has had that time measures the crawl schedule, not the page.

Nobody outside these companies knows the ranking mechanism, and it changes. Treat any specific claim about how the systems weight sources as a hypothesis, including the ones above.

And appearing in an answer is not the goal, it is the proxy. The goal is a buyer who arrives already convinced. Measure the appearance rate because it is measurable, but do not mistake it for the outcome.

Where to start on Monday

Run the first section. Twenty questions, one afternoon, written down with dates. You will learn more from the citation lists than from any article on this subject, including this one — because those lists are about your market rather than about the general case.

Then check robots.txt. It takes minutes and it is the only item here that can be silently, completely wrong.

If the answers turn out to be dominated by pages that do not serve New Zealand buyers, that is the opening, and our guide to business process automation in New Zealand and our work on AI agents for New Zealand businesses are examples of the shape of answer we think those questions deserve. If you would rather have this run as a measured programme than a project, that is what our automation service exists for.

Common questions