How ChatGPT, Perplexity and Google AI Overviews decide which businesses to recommend

Short answer: when someone asks ChatGPT, Perplexity or Google AI Overviews for a business recommendation, the system runs a live web search, reads a handful of pages and names the companies that those pages describe clearly, concretely and consistently. It is not a ranking you can buy. It is a selection from the sources the system can reach and trust. According to OpenAI, ChatGPT alone had 900 million weekly users in February 2026, so being named there is no longer a side issue.

We run these checks for Austrian companies every week, and the pattern repeats: ask for “the best web agency for SMEs in Linz” or “an estate agent in Vienna’s 19th district” and the answer ends in a short list of three to five names. Rarely the biggest firms. Mostly the ones whose website says plainly what they do, where and for whom, and who are mentioned on other sites as well. This article explains how the three big systems get there and what you can do about it. The practical step-by-step is in our guide how to get cited by ChatGPT; the service page is GEO.

Where does an AI answer get its knowledge from?

A language model can “know” your business in two ways.

• Training data: what the model learned before its cut-off date. You cannot change this quickly, and for small local businesses it is usually thin or outdated.
• Retrieval: the model searches the web while answering and quotes what it finds. This is where almost all local and commercial questions are decided today, and it is where you can act.

For a question like “which agent should I hire to sell my flat”, the system has to be current and local, so it searches. That makes a normal, well-built, crawlable website more important, not less.

How do ChatGPT, Perplexity and Google differ?

One detail matters for Google: AI Mode breaks a question into many related sub-questions (Google calls this “query fan-out”) and combines the results. A page can therefore be quoted for a question it never ranked for directly, as long as it answers one of the sub-questions well.

What do the three systems have in common?

None of the providers publishes a ranking formula, so what follows is what Google’s guidance says and what we see in our own checks. Four things come up every time.

1. The crawler can get in. The most common reason a site is never named is banal: a security plugin, a hosting default or a blanket “block all AI” rule locks out the search crawlers along with the training crawlers.
2. The page answers the question early. Passages that give the answer in one or two sentences are easy to lift out and quote. A page that opens with “we offer holistic solutions” gives the system nothing to quote.
3. The facts are concrete. Location, service area, price ranges, timeframes, number of projects, certifications. Models quote what is specific.
4. Other sources confirm it. Google reviews with text, entries in directories and the chamber of commerce, trade press, local news. Your website says what you claim; other pages confirm whether it is true.

Research points the same way. The GEO study by researchers at Princeton and other institutes (KDD 2024) tested text changes on a simulated AI search engine: adding statistics raised visibility in answers by about 41 percent and adding quotations by about 28 percent. That is a laboratory result, not a guarantee for your page, but it matches what we see: concrete pages get cited, vague ones do not.

What gets a local business named?

For “near me” and “in Vienna” questions the confirmation signals carry extra weight. In practice that means:

• One identical company profile everywhere: same name, address and phone number on your website, Google Business Profile, the WKO directory, Herold and industry directories. If a model cannot tell whether “Muster GmbH” and “Muster Immobilien” are one firm, it hesitates.
• Google reviews that name the service and the place, not only stars. “Great website for our practice in Graz” says more than five stars with no text.
• One page per service and location, each with real content, not the same text with a different town name.
• Structured data (schema.org) for company, services and articles. Google says it is not required for its AI features and there is no special schema for them, but it keeps your company facts consistent for every machine reader.

Why the technical foundation matters

An analysis by Vercel and MERJ found that the large AI crawlers, including those from OpenAI and Perplexity, do not execute JavaScript. A page whose text only appears after scripts have run can look empty to them. This is why we build sites on Next.js with server-rendered pages: the finished HTML, with headings, answers and structured data, is there on the first request. If you run a script-heavy theme today, see from WordPress to Next.js, and if you are planning a new site, our web design starts from this foundation.

What stays out of your hands?

• Training data. You cannot edit what a model learned last year.
• Wording of the answer. The same question can produce different lists on different days.
• Any guarantee. Nobody can sell you a fixed position in an AI answer. Be sceptical of anyone who does.
• llms.txt as a lever. The file is cheap to add, but Google says its search ignores it and the other providers have not confirmed using it. We generate it automatically, but we do not sell it as a visibility trick.

How do you test your own visibility in ten minutes?

1. Write five questions a customer would really ask, for example “Which web agency in Vienna builds websites for doctors?” or “Who builds a house in the Mostviertel?”
2. Ask each one in ChatGPT, Perplexity and Google (AI Mode or AI Overviews), logged out or in a private window.
3. Write down which companies are named and which sources are linked.
4. Check whether your company appears, whether the description is correct and which of your pages was used.
5. Repeat after four weeks. One run proves little because answers vary.

If you would rather not do this by hand, the free website check looks at the technical basics, and the GEO report shows real ChatGPT answers for your industry in about a minute.

Written byChristoph SpeiserProject lead & sales

Meet the team →