What AI says about your company: a half-hour check

What changed in search

A search result used to be a list: ten links, and the person chose for themselves. Increasingly the result is a finished answer. Google shows one above the links (AI Overviews), while ChatGPT and other assistants answer instead of the search engine entirely.

The difference for a business is fundamental. On a list of links you were eighth — someone might still have seen you. In an answer naming three firms there is no eighth place. Either you are in the answer or you are not.

Why this matters even if you are not at the top

An assistant is not obliged to take the first result. It assembles an answer from whatever it managed to understand and cross-check: company sites, directories, reviews, articles, discussions. A small firm with a clear website turns up in those answers regularly, while a large competitor with a brochure site full of adjectives does not.

This is one of the rare moments when a small business competes on equal terms with a big one. The window will close, though: once everyone starts working on it, the advantage moves to whoever invested earlier.

A half-hour check: five questions

Open ChatGPT and Google in an ordinary browser. Ask five questions in turn — phrased the way your customer would ask them, not the way a marketing specialist writes them.

• The choice question. "Who does your service in your city?" For example: "who builds websites for dental clinics in Kraków".
• The price question. "How much does your service cost in Poland in 2026?"
• The comparison question. "Your firm or a competitor — which should I choose?"
• The direct question about you. "What is known about the company name?" Here it matters not only whether you are named, but whether the details are right: address, phone, what you actually do.
• The sceptic's question. "Is it worth ordering your service from a small company?" This surfaces the objections the assistant considers typical for your market.

Ask each question twice — in separate chats or a new window. Answers vary, and a single run proves nothing.

What to write down

Set up a table with five columns and fill it in as you go: the question, whether you were named, who was named instead, what sources the assistant showed, and whether there are factual errors about your company.

That last column is usually the most valuable. An out-of-date phone number, an old address, a service you no longer offer, someone else's review attributed to you — an assistant repeats all of it confidently and without caveats, and the customer takes it as fact.

Where the assistant gets its answer about you

The order runs roughly like this, and it explains where the effort belongs:

• Your website — if there is anything on it worth quoting. Not slogans, but facts: what you do, for whom, what it costs, where you are, how long you have been working.
• Directories and maps — your Google profile, trade directories, company registers. This is where out-of-date details usually live.
• Reviews and discussions — review sites, forums, social media. The assistant summarises them, old complaints included.
• Publications — articles that mention you. One mention in a trade publication carries more weight than ten links from throwaway directories.

Three reasons you are not being named

• There is nothing on the site to quote. "Individual approach, a team of professionals, high quality" is text no answer can be assembled from. An assistant needs claims that can be checked and repeated.
• The facts are scattered and contradict each other. One phone number on the site, another in a directory, a third on the maps profile. A model will favour the firm whose details agree everywhere.
• No answers to the questions people ask. The site explains how good you are but not what it costs or how long it takes. Those are exactly the questions people put to an assistant.

What can be fixed on the site

The same things work as in ordinary search, but weighted heavily towards facts.

• Specifics instead of adjectives. "7–30 days", "from EUR 300", "working since 2010", "over 200 projects" — this is what ends up quoted word for word.
• One source of truth for contact details. Name, address, phone and opening hours must match on the site, in maps and in directories down to the comma.
• Answers to direct questions. A questions-and-answers section stating what it costs, what is included, what is not, and what happens after launch.
• Structured data. Markup for the organisation, its services and prices gives a machine structure instead of guesswork. Not magic and not a guarantee, but with everything else equal the site that is easier to parse wins.
• Writing for people, not for the crawler. Models were trained on human text and quote what reads like an explanation rather than a keyword list.

What not to do

Offers of "ChatGPT ranking" with a guaranteed place in the answer have already appeared. No one can guarantee that: the answer is built afresh for every query and depends on the wording.

Actively harmful: buying reviews, dozens of cloned pages for different cities, hidden text "for robots". Models reuse the quality signals of ordinary search, so the penalty lands there — and takes the rest of your traffic with it.

How to tell it is working

There is no ranking metric here, but the check is repeatable. Keep the five-question table and come back to it a month or six weeks after making changes: same questions, same wording, two runs again. Watch three things — whether you are named, whether the factual errors are gone, and whether the assistant cites your site as a source.

It is also worth reading your server logs: assistant crawlers visit the site, and the logs show which pages they read most.

What to do this week

• Ask the five questions, twice each, and fill in the table. Half an hour for the full picture.
• List every factual error about your company and correct it at the source — the site, maps, directories.
• Find the page on your site that says what it costs and how long it takes. If there is no such page, that is the first job.
• Replace three phrases about your "individual approach" on the home page with actual numbers.
• If you would rather approach this systematically than by feel, the Visibility in AI answers page sets out what the work involves.

In short

The check costs half an hour and needs nothing but a browser. Five questions, two runs each, a table with five columns. What follows is almost always the same finding: the assistant is not skipping your firm because of algorithms, but because there are no quotable facts on the site and because your details do not match across sources. Both are fixable without an agency — at least halfway.

Victor Parhimchik, founder of the IT Deweloper web studio