How to Show Up in ChatGPT: What 525 Citation Tests Taught Us

Showing up in ChatGPT means being the source the assistant quotes and links inside its answer, which happens only when the assistant runs a live search for that question and finds a page specific enough to quote. Since 1 July 2026 we have put the same 26 buyer questions to ChatGPT, Gemini and Google AI Overviews every six days, 525 graded checks in total, logging whether each one mentioned us, cited us, and whether it searched at all. This guide is what that data says, including the parts that contradict the usual advice.

What does it mean to show up in ChatGPT?

There are two different outcomes and most guides blur them. A mention is the assistant naming your business in prose. A citation is a linked source card the reader can click. Citations are the ones worth chasing: they carry a click, they persist in the answer, and they are the number that responds to work on the site.

Across 525 checks our citation rate went from 13 percent on the first audit to a band of 19 to 24 percent, where it has sat for two months. Our mention rate over the same period bounced between 4 and 25 percent with no trend at all. If you measure only mentions you will conclude your work is doing nothing, or everything, depending which week you look at.

Why is most advice about ChatGPT visibility untested?

Because almost nobody publishes a denominator. The standard checklist (write answer first passages, add schema, allow the AI crawlers) is mostly correct, but it arrives with no measurement of which parts matter or how often the result changes. We ran ours against a fixed panel so the same questions get asked the same way every time, which is the only way to tell a real move from a good week.

Finding 1: if the assistant does not search, you cannot win

This is the largest effect in the dataset and it is not close. We log whether each answer triggered a retrieval, meaning the assistant actually fetched pages, or whether it answered from training data.

• When retrieval fired: cited 33 percent of the time (n=256)
• When it answered from memory: cited 0 percent of the time (n=149)

Not "less often". Zero, across 149 checks. Nothing you publish this month sits in a model trained last year, so on a question the assistant feels confident answering unaided, your website is not a participant in the conversation. It is invisible by construction.

That split reorders the whole playbook. Before asking how to get cited, ask whether the question even triggers a search. Broad definitional questions ("what is business automation") get answered from memory. Specific, current, technical, named entity questions ("does Towbook have an API to sync invoices to QuickBooks") send the assistant looking. You want the second kind.

Finding 2: you get cited on narrow questions, never on broad ones

Sorting all 26 prompts by citation rate produces an unusually clean split. Every question we win names a specific piece of software. Every question we never win is generic.

Thirteen of our 26 prompts have never produced a single citation across 270 checks. Note the last row. The question this article is about is one we have never once been cited for, and we are publishing anyway, because the reason is the useful part and we come back to it below.

Naming software is necessary but not sufficient. Prompts naming Shopify, Fathom, GoHighLevel and Housecall Pro also sit at zero. What actually predicts a citation is a narrow platform with a thin published record. AppFolio and Towbook are closed ecosystem tools whose APIs almost nobody has documented properly. Shopify has ten thousand tutorials and an official docs site that will outrank you forever. Specificity is a moat only where the ground is not already occupied.

Finding 3: your own domain is not the only source that matters

Counting every cited domain across all 525 checks:

Our own site is the most cited domain on our own panel, which took a year of writing. But YouTube and Reddit together were cited 167 times, more than our site, and neither is something you optimise with schema. When an assistant wants a human account of whether a tool works, it goes where humans argue about tools.

The practical read: on questions you can win, on site work is the lever. On broad hiring questions it is not a lever at all, which is why our 13 zero prompts stay at zero no matter what we publish. Those need third party presence, and the data points at video and forums.

How to show up in ChatGPT

Step 1: pick questions that trigger a search

Write down the questions a buyer would type, then ask each one in ChatGPT yourself and watch what it does. If it answers instantly from memory, deprioritise it. If it pauses and cites sources, that question is winnable and belongs on your list. This takes an afternoon and will save you a quarter.

Step 2: go narrow enough to be the best available source

Pick the platform, the task, and the failure mode. "AppFolio investor reporting" beats "property management automation" because the first has three articles written about it and the second has three thousand.

Step 3: let the AI crawlers in, then check the raw HTML

GPTBot, ChatGPT-User, PerplexityBot, ClaudeBot, Google-Extended and CCBot all need to be allowed in robots.txt. Then verify your structured data is in the server rendered HTML rather than injected by JavaScript. We found our own JSON-LD was client side only, which meant every AI crawler that does not run JavaScript saw no structured data on any page of the site. One command tells you:

curl -s https://yoursite.com/some-page | grep -c 'application/ld+json'

If that returns zero, the schema you believe you have does not exist as far as those crawlers are concerned.

Step 4: answer in the first 60 words

Put a direct, self contained answer immediately under each heading, before any context. Extractors lift the first clean answer they find, so a paragraph that opens by setting the scene gets skipped in favour of one that opens with the answer.

Step 5: publish numbers you actually have

Concrete attributed figures get extracted far more often than claims do. This applies especially to your own operating data, the one category of fact a competitor cannot copy. If you have no numbers, the fastest route to some is measuring one thing for a month.

Step 6: measure with a fixed panel, on a schedule

Same questions, same wording, same cadence, written to a database. Ours runs every six days across three surfaces. Without a denominator you cannot tell a change from a Tuesday. Track citations, and track whether retrieval fired, because the second explains most of the first.

Where it gets complicated

The surfaces disagree with each other. On one audit ChatGPT, Gemini and AI Overviews each cited us on five of 26 prompts, but only one prompt was cited by all three. Treating AI search as a single channel will mislead you: a prompt you win on Gemini can be one you have never won on ChatGPT.

Citation is not traffic. Our most cited page earns single digit clicks a month from Google. Being in the answer is worth having, but justify the work on traffic alone and the numbers will disappoint you. The value is being named at the moment somebody is deciding.

The fix for a zero prompt is usually not on your website. This is the uncomfortable one, and it is why our own AI visibility prompts sit at zero. Generic hiring questions get answered from aggregators, forums and video. We could write ten more pages about AI visibility and the number would not move, because no amount of self published content persuades an assistant to treat you as the neutral answer to who somebody should hire.

What we would not bother with

Paid directory listings, on this evidence. Clutch, GoodFirms, DesignRush and hellodarwin appeared in zero of 525 checks. They may be worth buying for other reasons, but not this one. We would also not spend time on FAQ schema expecting rich results in Google, which were removed for most sites in 2026. Keep emitting it, because it helps machine parsing and costs nothing, but the visible question and answer text is what earns the citation now.

Frequently asked questions

How long does it take to show up in ChatGPT?

Our first citation appeared within a few weeks of publishing genuinely specific technical content, but the rate took about two months to settle into a stable band. Expect a quarter before you can tell a real change from noise, and only if you are measuring on a fixed panel rather than spot checking when you remember to.

Does ChatGPT use the same sources as Google AI Overviews?

Not reliably. On our panel each surface cited us at a similar rate but on largely different questions, and only one prompt in 26 was cited by all three on the same day. Optimise for the mechanism, live retrieval finding a specific page, rather than for one product.

Can I pay to appear in ChatGPT answers?

No. There is no placement to buy among the cited sources. The nearest paid lever is sponsoring content on the third party sites assistants already cite, which on our panel means video and forums far more than directories.

Does schema markup actually get you cited?

It helps machines parse you and it is cheap, so do it. It is not the binding constraint. Our largest measured effect by a wide margin was whether the assistant searched at all, which no amount of markup controls. Fix retrieval eligibility first, then markup.

What if my industry only has broad questions?

Then on site content is the wrong tool, and it is better to say so than to spend a year proving it. Thirteen of our own prompts are exactly this. The route there is third party corroboration, being the person in the video or the thread that gets quoted, not another page on your own domain.

Is this worth it for a small business?

It is worth it when your buyers ask specific questions about named software you genuinely know, because you can become the best available answer inside a month. It is a poor use of money when your category is broad and crowded, and honest measurement will tell you which one you are in within about eight weeks.

We have run this panel on our own site since July, and it changed what we publish more than how we publish. Our AI search optimization page covers how we run the measurement, and the AppFolio investor reporting build is the work behind the prompt we win most often. If you want a straight answer on whether your buyers ask the kind of question an assistant actually searches for, book a call and we will tell you in the first five minutes.