How a company appears in ChatGPT answers — two routes and five decisive factors

A company appears in ChatGPT’s answers by one of two routes: either the model knows the company from its training data, or it finds the company through a web search at the moment of answering. Both can be influenced, but with different methods and on different timescales. In this article I go through both routes and what a Finnish company should do in practice.

How ChatGPT chooses who to mention

When a user asks ChatGPT for a recommendation, say ”what is a good accounting firm in Tampere”, the answer is produced in one of two ways.

From memory. The model has been trained on an enormous body of text, and if your company appears in it often enough and consistently enough, the model can talk about it without searching at all. The companies that get here are the ones written about a lot: in news, in listings, in discussions, on Wikipedia.

By searching. In newer ChatGPT versions the model performs a web search and builds the answer from the pages it finds, which it also cites. Here the game resembles search visibility: the answer includes whoever’s page turns up in the search and answers the question most clearly.

For a small or medium-sized Finnish company, search is the realistic route. Influencing training data is slow, but you can reach search-based answers within weeks, as long as the content is built to serve as raw material for an answer.

Five things that decide search-based visibility

1. Let the crawlers in. Check that robots.txt does not block OpenAI’s crawlers. If the crawler cannot reach the page, the page cannot end up in an answer. This takes a minute to check: open your robots.txt and look for GPTBot blocks.

2. Direct answers. A language model lifts answers from pages that answer the question immediately, not after a 500-word introduction. If your page is titled ”What does bookkeeping cost?”, the first paragraph has to answer it with numbers.

3. Verifiable facts. Prices, timescales, comparisons and definitions end up in answers more often than opinions and marketing lines. ”Bookkeeping typically costs an SME X to Y euros a month” is a quotable sentence; ”we offer quality bookkeeping at a competitive price” is not.

4. Consistent basic details. Name, industry, location and services stated the same way everywhere: on the site, in the Google Business Profile, in directories. Contradictory information weakens the model’s confidence, and an uncertain model leaves you out.

5. Mentions elsewhere. ChatGPT’s search favours sources that are trusted. Industry listings, press coverage and customer stories on third-party sites raise the odds of ending up in an answer, in the same way they raise Google visibility.

How to influence the training data

The slower route is getting the company into the body of text future models are trained on. In practice that means leaving traces on the public web: interviews, guest articles, listings, open discussions. A single mention changes nothing, but years of consistent presence show up as the model ”knowing” the company without searching.

Here it pays to be patient and honest: the training cycle for models runs in months or years, and nobody can promise that your company specifically will end up in the next model’s memory. That is why search-based visibility is where the effort should go first.

Test where you stand before doing anything

Ask ChatGPT ten questions your customers would ask while buying. Not ”what is Company Ltd”, because nobody asks that, but ”which firm should I choose for X”, ”best Y service in Finland”, ”who should I buy Z from”. Record:

• Is your company mentioned? In what tone?
• Which competitors are named?
• Which sources does the answer cite?

The cited sources are the most valuable finding. They tell you directly which sites and listings are worth getting into. The model has just shown you its own bibliography.

Two routes into an answer

ChatGPT can mention your company in two different ways, and they need different work.

The first is training data. If your company was written about widely enough before the training cut-off, the model may know of you without searching. This route is slow, cannot be influenced quickly, and mostly favours companies that were already well covered.

The second is search at answer time. The service runs a search, reads pages and cites them. This route responds to work you do this month, and it is where a small company has any chance at all.

Almost all practical advice concerns the second route. When someone says a new page appeared in ChatGPT within a week, they are describing search, not training.

What the search route requires

Three conditions have to hold, and failing any one of them makes the other two irrelevant.

Your page has to be fetchable. The search function uses its own crawler, and if a firewall or bot protection turns it away, the page cannot be read no matter how well it ranks. Check response codes in the log rather than assuming robots.txt tells the whole story.

Your page has to rank. Sources are picked from search results, so a page outside the first page of results is rarely read. This is why AI visibility work sits on top of search visibility rather than replacing it.

Your page has to answer the question directly. Among ten candidate sources, the one whose answer can be lifted without rewriting tends to be the one cited.

Writing a paragraph that survives summarisation

A cited passage usually shares the same shape, and it is a shape you can write on purpose.

It opens with the answer rather than context. It contains a number, a range or a named condition, because those are the parts a summary keeps. It stands alone, meaning a reader who sees only that paragraph still understands it. And it is between forty and eighty words, long enough to be complete and short enough to quote whole.

Compare two openings on pricing. One says that costs depend on many factors and every project is different. The other says the work typically runs between two and five thousand euros, with the range driven by the number of pages and whether translation is included. The second is quotable and the first is not, and the difference took one sentence.

What your own name returns

Before optimising for customer questions, find out what the service already says about you. Ask it plainly what your company does, and read the answer for three things.

Whether the facts are right. Wrong service lists, an old address or a merged description with a similarly named company are all common, and they usually trace back to stale information somewhere public.

What it does not know. Gaps show what is missing from your site rather than what is wrong with it.

Where it got the information. If the answer cites a directory listing rather than your own site, your site is not the strongest source about your own company, which is worth fixing before anything else.

Why the answer changes between attempts

The same question asked twice can produce different sources and different companies. This is not a fault and it is not something you can optimise away.

The search runs afresh each time, results shift, and the model chooses among candidates without a fixed ranking. Personalisation and conversation history add more variation. In practice this means no single answer proves anything, in either direction.

Measure across a set instead. Twenty questions, run monthly, recorded in a spreadsheet. The share of answers mentioning you is a stable number even though each individual answer is not, and it is the only figure worth reporting to anyone.

A realistic order of work

For a company starting from nothing, this sequence spends effort where it returns most.

First confirm the crawler can reach you and receives a normal response. Then ask what the service says about your company and correct anything factually wrong at its source. Then pick the five questions customers ask before buying, and write one page for each that answers it in the opening paragraph with a number in it. Then check the pages rank at least on the first page of results for their question, since that is the gate to being read.

Only after that is it worth measuring, because before that the measurement has nothing to detect.

Frequently asked questions

Why does ChatGPT not know my company?

Most likely because there is too little consistent text about your company on the public web. For a small company this is the normal starting point, not a failure. You can still reach search-based answers once your site answers buyers’ questions directly.

Can visibility in ChatGPT’s answers be bought?

No. There are no ad slots inside ChatGPT’s answers and mentions cannot be bought from OpenAI. Visibility is earned with content and mentions, the same logic as in organic search.

How quickly can ChatGPT visibility improve?

In search-based answers, within weeks, if the content is sorted and crawlers can reach the site. Influencing what the model knows permanently takes months to years.

Does the same work help in Perplexity and Google’s AI answers?

Largely yes. All three lift answers from well-structured content that answers directly. The differences are in how they weight sources: Perplexity cites most eagerly, Google leans on its own index.

How long does it take for a new page to appear in ChatGPT answers?

Through the search route it can happen within days, since the content is fetched at answer time. Through training data it takes until the next model generation, which is months away and not something a single site can influence.

Why does ChatGPT describe my company incorrectly?

Usually because the strongest public source about you is not your own site. Old directory entries, outdated listings and pages about a similarly named company all feed in. Correcting the details at their source works better than adding more content.

A half-minute quick test: the AI visibility test asks a language model what it knows about your company and shows the answer as it comes. What the full GEO programme contains is described on the service page.