How ChatGPT decides whom to name
When a buyer asks ChatGPT for a recommendation, two things can happen. It may answer from what it learned during training, in which case the names it produces are the ones that were well represented and consistently described across the web when the model was built. Or it may retrieve: run a search, fetch a handful of pages, and write an answer grounded in them, showing those pages as sources. Questions about specific companies, prices, locations and recent events usually trigger retrieval.
You have a little influence over the first path and a lot over the second. Everything below is about being present, accurate and citable where retrieval happens.
1. Let the right crawlers in
This is the gate, and it is checked first because nothing else matters if it is closed. OpenAI documents separate user agents: one for training, one that indexes for ChatGPT search, and one that fetches a page when a user asks about it. Each can be allowed or blocked independently in robots.txt. A rule written to keep your content out of training, if written too broadly, also keeps your pages out of answers.
Open your robots.txt and look for the agents by name. If you want to be cited, the search and user-triggered agents need to be allowed. The same applies to the other engines: Perplexity documents its crawler, Anthropic documents its, and Google explains that AI features in Search use ordinary Googlebot crawling. Also check that a firewall or bot-protection product is not blocking these agents silently; robots.txt can say one thing while the server does another.
2. Make your key pages readable as text
Crawlers that fetch raw HTML see what is in the HTML. If your pricing, your services and your location only appear after JavaScript runs, a fetch sees a shell. Open your most important pages with JavaScript disabled in your browser; if the content is missing, it is missing for the engine too. Server-side rendering or static generation fixes this, and it is usually a task for whoever built the site rather than a content job.
3. Have one page that answers each question
Take the commercial and comparison questions from your buyers (“how much does X cost in Estonia”, “X vs Y”, “who does X for English speakers in Tallinn”) and make sure one page on your site answers each one plainly. The structure that works is unglamorous: a heading that matches the question, a first paragraph that gives the answer, then the detail. State facts rather than implying them. A page that says “transparent pricing” is not citable for a price question; a page that says what things cost is.
FAQ sections on those pages help in the same way, because they put a question and its answer next to each other in the text. Mark them up with FAQ structured data if you like; Google limits which sites get FAQ rich results, but the markup costs nothing and describes something real.
4. Fix what is wrong, at the source
Run your brand questions (“tell me about X”, “where is X located”, “what does X charge”) and note every wrong claim. Then find out where it comes from: often a directory listing, a review profile, a press mention or an old page of your own. Correct it there. Engines repeat what the most-cited sources agree on; if three listings say you are in Riga, the engine will too, however clearly your homepage says Tallinn.
Then make your own site the canonical statement of the facts: a single page with name, legal entity, locations, services, prices or price ranges, and contact details, in plain text, with Organization structured data that matches. Consistency across your site and third-party profiles is the lever here, not any one page.
5. Be present where the engine already looks
When you run your buyer questions and you are absent, look at the sources the engine cited for the competitors it named. Those pages are the engine’s idea of what defines your category: a comparison article, a directory, a professional register, a review platform, an industry list. Getting an accurate entry on the same kind of page is often faster than any change to your own site, because it puts you inside material the engine already retrieves for that question.
6. The small things, in proportion
A sitemap that lists your real pages. Structured data where it describes something that exists (your organization, a product with a price, an FAQ). An llms.txt file, which is a published proposal that no major engine has documented using; cheap to add, low expected effect. Visible dates on pages that change. None of these is a strategy. All of them remove small reasons for a page to be skipped.
What to expect
Changes to pages the engine retrieves live can show up once those pages are fetched again; changes that depend on training data wait for a model update, and no one outside the engine company can tell you when. The sensible stance is to make the changes, then measure the same questions the same way each month and judge from your own data. If your mention rate on commercial questions moves, you will see it; if it does not, you will know which fixes to revisit.
Where to start depends on where you are, which is what a check tells you. Our free check runs three buyer questions on ChatGPT and Perplexity and emails you the answers with your brand marked; the full audit runs fifteen questions twice on four engines, checks the crawl gate and page citability by code, compares every claim with your facts, and gives you the fixes above in the order that applies to you.