How to get your business recommended by ChatGPT

ChatGPT recommends a business when, for the question that was asked, your name is a plausible and supported answer. There is no switch in the product labelled “add my company”. There is also no honest agency that can sell you a guaranteed mention. What you can do is make the true answer easy to find, easy to quote, and consistent everywhere a crawler or a person might look.

This is a working sequence for a service business, written from the point of view of the work we do at iShack Innovation Consultancy. It does not include a success rate. We do not have a public set of before-and-after citation counts that we are willing to present as a benchmark. If we publish one later, it will show the prompts, the dates and the assistants.

Decide which ChatGPT you mean

People say “ChatGPT” and mean several behaviours.

Sometimes the product answers from the model alone. The reply has no footnotes. Changing your website today will not reliably change that reply today, because nothing fetched your website.

Sometimes the product searches or browses and shows sources. Those replies can cite a page that exists now. This is the behaviour you can work on. When you test, turn on the mode that shows sources, or use a product that always cites, such as Perplexity, as a second check. Google AI Overviews and Gemini are worth the same questions, because your buyers do not all use one app.

Write down which mode you used. A screenshot of an uncited answer and a screenshot of a cited answer are not the same measurement. Mixing them is how reports become nonsense.

Start from the question, not the homepage

List the questions a buyer asks before they are ready to be sold to. For a service business they tend to look like this:

• Who are reputable providers of this service in this country or city?
• What should I look for when hiring this kind of firm?
• How does provider A differ from provider B for a company of our size?
• What does this service cost, in ranges, for a business like ours?

Keep the list short. Ten questions you will actually re-ask each month are enough to see whether anything changed.

Include at least one question where you are not the flattering answer. If a competitor is the name everyone already says, that question belongs on the list. The point of the exercise is to see the recommendation a buyer would get, not the recommendation you wish they would get.

Ask the questions. Save the wording, the date, the product, whether sources were shown, the names mentioned, and the URLs cited. That file is the asset. A slide with a single green arrow is not.

Make the pages those answers would have to quote

Look at the URLs ChatGPT and Perplexity already cite for your questions. Read them. They are usually specific. They name a service, a place, and a criterion. They are rarely a homepage that says the company is passionate about excellence.

You need a page for each question you care about, or one strong service page that answers several of them without becoming a junk drawer. A useful page does five things in the HTML, not only after a script runs:

• The title and the H1 say what the page is about, in the words a buyer would use.
• The first paragraph answers the question. It names your business only where that is true, and it says who the service is for.
• The next sections add the conditions: where you work, the size of client you suit, what you do not do, and how an engagement starts.
• The facts match the rest of the site. One name. One set of locations. No phone number you do not answer. No street address that is a placeholder.
• The page links to the next most relevant page on your site, and those pages link back.

Fetch the URL without JavaScript before you celebrate. If the H1 is missing from that response, do not assume a crawler will reconstruct it. Shipping the words in the HTML is the fix.

Give the assistant a consistent entity

Models and the search systems behind them are better at repeating an entity that looks the same everywhere. You do not need a complicated knowledge-graph project to start. You need agreement.

Use one public name. Put it in the title of the homepage, in the footer, and in the first mention on each service page. If a shorter brand and a longer legal name both exist, use the brand in prose and the legal name where a contract would use it. Do not invent a third variant for “SEO reasons”.

Publish only the contact details you want repeated. An assistant will cheerfully quote a fake address if you put one in your structured data. We have seen that mistake, including on sites that should have known better. If you do not yet have a confirmed public address or phone number, leave them out. An email address you actually read is enough.

Use Organisation structured data for the name, the URL, and the profiles you really have. Add a phone or a postal address only when they are real. FAQ structured data is appropriate when the page shows those questions and answers to a human as well. Article structured data is appropriate on articles. Breadcrumb structured data is appropriate when the breadcrumbs are visible. Structured data that describes a page you do not have is a liability, not a tactic.

Keep the same name and the same description on the profiles you link to. One accurate profile is better than five neglected ones.

Write the way a citation is extracted

A cited sentence is usually a sentence that can stand alone. “We are the leading provider of innovative solutions” cannot stand alone. It has no subject a buyer can check.

Prefer sentences like: “We install and maintain commercial refrigeration for food producers in the Western Cape and Gauteng. We do not do domestic kitchens.” A model can quote that. A buyer can see whether it matches the rest of the site.

Say what you will not do. Boundaries make a recommendation safer. An assistant that is trying not to mislead a user will often prefer a source that limits itself.

Put prices in ranges if you publish prices, and say what the range includes. A page that says “contact us for a quote” on every question gives the assistant nothing to cite when the user asked what it costs. If you are not willing to publish a range, do not invent one. Answer a different part of the question instead, such as what changes the fee.

Do not add a block of synonyms under the real copy. Do not hide text. Do not write a paragraph whose only job is to repeat “ChatGPT” and your city. That style is easy to detect and embarrassing when a prospect reads it, which they will, because you sent them the link.

Earn the mentions you cannot write yourself

ChatGPT, when it cites, often cites a page that is not yours: a directory, a trade association, a journalist, a comparison written by someone else. You cannot edit those pages. You can be a source they would mention without being asked to game anything.

Practical versions of that, with no count attached:

• Be listed, accurately, in the associations and directories your buyers already use. Claim the listing. Use the same name and URL.
• Publish a small piece of original explanation your peers would forward: a method, a checklist, a definition, a record of how you scope work. One strong page is a better pitch to a journalist than a monthly drip of thin posts.
• When you pitch a writer or a partner, pitch a fact they can check. Do not pitch “we guarantee coverage”.

If coverage happens, it is because the story was useful. Treat it as a note in the monthly record, not as a numbered deliverable.

Measure without a made-up score

Once a month, re-ask the same questions, in the same products, and append the result to the same file. Note:

• the date
• the product and whether sources were on
• the exact prompt
• whether you were named
• who else was named
• which of your URLs, if any, were cited

Look for a change that survives more than one check. A single lucky answer is not a trend. A question that never cites you, across a month, is a real gap.

If your analytics show referrals from chat or answer products, record those too, with the caveat that many assistants do not pass a clean referrer. Absence of referral traffic does not prove absence of mentions. Presence of a citation in your prompt log does not prove a lead. Keep both, and do not collapse them into one percentage.

Share the log with the person who approves the work. If an agency will not show you the prompts, you cannot manage them.

A 30-day sequence that does not depend on us

Week one: write the question list, run it, and save the log. Fetch your key URLs without JavaScript. Fix the worst mismatch in your name, your locations, or an empty HTML shell.

Week two: choose three questions from the log where you should have been a reasonable answer and were not. Outline a page for each, or a revision of a page you already have. Put the answer in the first paragraph.

Week three: publish those pages, with normal titles, a canonical URL, and internal links. Add structured data only for what the page visibly says. Ask someone who did not write the page to read the first paragraph and tell you what the company does.

Week four: re-run the original questions and update the log. The sequence will not get you recommended on a deadline. It shows whether the basics are in place.

When a managed service is the right tool

Hire help when the questions are clear and the editing, the technical fixes, and the monthly log are more than your team will actually do. The work is ordinary: pages, facts, crawling, and a record. It is ordinary on purpose.

The version we sell is described on our AEO service page, and the fees are on the pricing page. A free starting point is the AI visibility review. If you only want a technical pass of a URL, the free site audit is the smaller tool.