How do I get my website cited by ChatGPT?

Last updated 11 September 2026. Written by Oyekola Obajuwon, founder of Growveloper, who builds the website, the marketing and the follow up for owner run businesses in Nigeria, the United States and the United Kingdom.

Get cited by ChatGPT by answering the question in your first paragraph, on a page whose heading is the question itself, and by being mentioned on the sites the model already reaches for. We ran 62 prompts against our own site and were cited zero times. The pages that beat us were not longer or older. They were clearer.

That is the answer. What follows is the measurement behind it, including the part that was unflattering, because the unflattering part is the useful one.

Does ChatGPT actually read my website, or just its training data?

Both, and the difference decides what you can change.

When ChatGPT answers a question about which web developer to hire in Lagos, it does not recall that from training. It runs a search, pulls a handful of pages into the answer, and links them. That set of pages is called the candidate set, and getting into it is a retrieval problem you can influence this month.

The other half is slower. The model already holds associations about who does what, built from everywhere your name appears, and that is why two businesses with identical pages get different treatment. Kevin Indig's framing for this is the cleanest one going: traditional search is crawl, index, rank, and answer engines are retrieved, cited, trusted. Three different filters, and the last one is not about your page at all.

So there are two jobs. Make each page the easiest possible thing to quote. Then be a name the model has met before. Most advice covers the first and goes quiet on the second.

What kinds of pages does ChatGPT cite?

We stopped guessing and measured it. We set up five topics in Microsoft Clarity, which watches what AI assistants say when they answer questions in your category, and ran 62 prompts through ChatGPT on 9 and 10 September 2026. Growveloper was cited zero times.

Sort that list by what the domains are rather than who owns them, and there are only three kinds of thing on it.

Directories

themanifest.com, seo.com, subcontractorhub.com. Pages whose entire job is to list businesses in a category. The model reaches for them because a directory answers "who does this" in the format the question was asked.

Tool vendors

semrush.com, zapier.com, hubspot.com, trainerize.com. Companies that sell software and publish the definitive explainer for the thing their software does. They have been writing the reference page for a decade.

Articles that answer the exact question in the first paragraph

Everything else, including domains nobody has heard of. ztabs.co. kivolaro.com. adrianrbdev.com. These are not authority plays. They are pages that happened to be shaped like an answer.

That third bucket is the whole opportunity, and it is the only one of the three you can enter this week.

Why does my service page never get cited?

Across all five topics and all 62 prompts, a Growveloper service page was cited zero times. Not cited rarely. Zero.

This is the finding that changes what you do on Monday, because the service page is the page most businesses put their effort into. Ours are not bad. The web development page states what gets built, what it costs and how long it takes. It is simply not an answer to a question. It is an offer, and an offer answers "why us", which is not what anybody typed.

A service page says: here is what we sell, here is the proof, here is the button. An answer engine is looking for a passage it can lift and attribute. A sales page has no liftable passage in it, because every sentence depends on the one before it to make its case.

So the service page is not the thing that gets you cited. It is the thing the citation points at, along with the proof pages behind it, like the VIP Creative Studio rebuild, which is the page that has to do the convincing once somebody actually arrives. You need something else to do the citing, and that something is an article that answers one question completely.

What do the pages that do get cited have in common?

In July we took the queries our buyers actually use, found the pages Google's AI Overview and ChatGPT were quoting, and opened nine of them in the page source to count what was there instead of assuming.

Read down the columns and the answer is embarrassingly simple. Schema varies from the full stack to nothing at all. FAQ blocks are present on five and absent on four. Length runs from 910 words to 4,860, a five fold spread, which means length bought nobody anything.

One column has no variance. All nine answer the question in the first sixty words. Not one of them opens with a story.

Two of the nine are worth staring at. Apollo's page is the most structurally disciplined thing in the set. Twelve of its twelve headings are literal questions. It carries no JSON-LD at all, and it still gets quoted. Meanwhile gtmepulse.com is 910 words with almost no schema, no FAQ block and no question headings, and it gets quoted anyway, because it published original data with a stated method: 228 practitioners surveyed across 32 countries. It is the shortest page in the set and it beats the three thousand word vendor guides.

The pattern across both: answer first, and have something of your own to say.

Why did our own best page still get cited zero times?

Here is the part I did not enjoy writing.

Look at the last row of that table. The three vendor tax is our own article, and on the structural checklist it is the best page in the set. It answers its question in the first forty words. Its first heading is the question. It has a named FAQ block. It emits the full schema stack, which on our site means BlogPosting, FAQPage, Question, Answer, Person, WebPage and Organization. It cites sixteen named outside domains.

Six of the seven things the cited pages share, and we had them. We were cited zero times out of 62 prompts.

That is the finding. Structure is necessary and it is not sufficient. A page can pass every item on the checklist and still be cited zero times, because the checklist only decides whether you can be quoted. Something else decides whether you get reached for.

Two things were missing, and only one of them is a page problem.

The page problem was small and dull: no visible last updated date, and no tables. Ryan Law's study of 17 million AI citations found that AI cited content is measured 26% fresher than content ranking in ordinary search results, so a page with a 2026 publish date and no modified date reads as older than it is. And every report we ran showed tables getting quoted more than prose, because a table is already a passage.

The other one is not fixable on the page, and it is the real answer to the question this article is named after. In July we searched our own invented phrase, "three-vendor tax", and found our page sitting at position seven on its own term. Not first. Seventh. The model had no reason to reach for a domain it had barely met, on a phrase we coined, so it reached for the directories and the tool vendors it meets every day instead.

Being quotable is a page property. Being reached for is a reputation property. Nobody sells you the second one in a checklist because it takes months and it happens mostly on sites you do not own.

Do I need llms.txt, schema markup or a longer article?

Short answers: helpful, helpful, and no.

On longer, the table above settles it. Nine cited pages ranged from 910 to 4,860 words. Write until the question is answered and then stop.

On schema, be honest about what the evidence supports. Apollo gets cited with none. The smallest domain in the set uses the fullest stack and wins its query. The fair reading is that schema helps a small unknown domain punch up and does not rescue a page that buries its answer. Ours carries the full stack and it was not enough on its own, which is exactly what you would expect if schema is a helper rather than a gate.

On llms.txt, the same. It is a plain context file at the root of a site, from the llmstxt.org convention, listing your important pages so an assistant can find them without crawling everything. We publish one, and we generate it from the content management system on request so it cannot go stale when a new article ships. It is worth twenty minutes. It is not worth a strategy.

And the thing worth reading before you buy anybody's AI visibility package: Google's own documentation on optimising for generative AI features states that its AI features run on core search ranking, so there is no special markup and no special file required, and it warns that writing separate content aimed at machines can trip its scaled content abuse policy. There is a research finding in the same direction. The 2024 GEO paper from Aggarwal and colleagues at Princeton, published at KDD, tested optimisation methods against a benchmark of real user queries and reported visibility gains of up to 40%, and also reported that which method works changes by domain. So the honest version is: the levers are real, and nobody can hand you the ranked list of them for your category.

Allowing the crawlers, though, is a genuine pass or fail. If your robots file blocks the agents that fetch pages at answer time, no amount of structure matters, because the page never enters the candidate set. Ours names sixteen answer engine and AI search agents explicitly and allows them, GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot and the rest, while crawlers that only harvest training data are left out. Check yours. It is a five minute job and it is the cheapest thing on this page.

How do I check whether ChatGPT is citing me?

You cannot fix a number you have never seen, and almost nobody has seen this one.

Set up topic tracking in Microsoft Clarity, which is free. You give it the topics your buyers ask about, it runs prompts against the assistants, and it reports which domains got cited and how often, including whether yours appeared at all. That is where our zero came from. It took an afternoon to set up and it replaced a year of guessing.

Then do the manual version too, because it tells you something the report does not. Open ChatGPT and type the exact sentence a buyer would type. Not your keyword. The sentence. "Who can build a website and connect it to WhatsApp for my cleaning business in Lagos." That is a real prompt shape, and it is why we keep a page for cleaning companies rather than one page for everybody. Read which businesses get named, then open the pages it links and ask what they have that you do not. Usually the answer is a heading shaped like the question.

Write down what you find and the date. That number is your baseline, and a baseline you wrote down beats a dashboard you did not understand.

Can a small site really win this?

This is the part where the evidence is on your side.

A December 2025 study by Zhang and colleagues compared what six AI search engines cite against what two traditional search engines return, across 55,936 queries, and found that 37% of the domains AI engines cite are unique to them. More than a third of what gets cited never appears in the ordinary search results for the same question. The ranked list is not the gate any more.

Our own July sweep says the same thing from the other end. On one query the AI Overview cited seventeen domains while only seven showed up in the organic results, and several of the cited ones were businesses nobody has heard of. On the questions small business owners actually ask, we watched an assistant cite a personal Medium post, a Facebook group thread and an unrelated Russian language site. That is not a competitive field. That is a vacuum with a few people standing in it.

Which is the honest reason to bother. Not because AI search is the future, but because right now the pages answering these questions are mostly bad, and being better than a Facebook thread is an achievable standard.

Two cautions so you size the prize correctly. Nathan Gotch's point is that even winning the "AI cites your site" slot gets you one of roughly thirty five citations feeding a typical answer, so the goal is citation count across many sites rather than one perfect page. And the same measurement that produced our zero will produce your zero, and then it produces a one, and the gap between those is measured in months.

What should a small business do first?

In this order, because each step makes the next one measurable.

First, measure, so you have a number to beat

Set up the topic tracking, run the prompts, write down the zero. This takes an afternoon and costs nothing. Every decision after this is otherwise a guess with a budget attached.

Second, unblock the crawlers

Check that your robots file allows the agents that fetch pages at answer time. Five minutes. If this is wrong, nothing else you do counts.

Third, write one article that answers one question completely

Not a service page. One question, in the words a buyer types, as the title. The answer in the first sixty words, before the story. Every heading a sub question somebody would ask on its own. One table if the question is a comparison or a price. A named source behind every number. Three to five questions and answers at the end, for the versions of the question you did not lead with. Then a visible last updated date, because freshness is measured. If you want the shape rather than the description, what growth engineering is is built to it, heading for heading.

Fourth, put your name where the model already looks

The directories get cited, so get listed in the ones for your category and keep the name, address and description identical across all of them. Answer the question in the places the assistants quote constantly, which in our measurements meant forums, LinkedIn and YouTube far more often than vendor blogs. This is the slow half and it is the half that decides whether anyone reaches for you.

Fifth, put a date in the calendar to do it again

Ninety days. Re run the same prompts, compare, and refresh the article with what changed. The measurement is the programme. One article is a coin toss.

We are running exactly this on ourselves, in public, starting from a documented zero. If you want the diagnosis done properly on your own site rather than the version you can run in an afternoon, that is what the Growth Audit is for, and what "one connected system" means when we say it is on the growth marketing page.

But you do not need us for the first four steps, and you should not wait for anybody to start them. Go and type your own buyer's question into ChatGPT tonight. Whatever it says back is the only competitive research that matters, and it is free, and most of your competitors have never once looked.