How does ChatGPT decide which local businesses to recommend?

ChatGPT tends to recommend businesses that show up clearly on its map, have plenty of reviews or belong to a well-known chain. When we asked it for furniture stores in Cologne, it named home24 Outlet, IKEA, TRÖSSER and Revive most often and showed a star rating for three out of four recommendations. Claude built longer lists, dominated by large chains. In October 2026 we put typical customer questions to ChatGPT and Claude in three languages, ran a short test with Gemini and analysed around 970 recommendations. For a smaller business, the levers are complete listings, a steady flow of reviews and a website that describes each product range on its own page.

How did we test it?

We asked the AI assistants the kind of questions real customers ask and saved every answer. The aim was to record what these systems actually recommend today, using one city and one sector so that the results are comparable.

• Questions: in German, Turkish and English, from general ("Can you recommend a good furniture store in Cologne?") to specific: sofas, beds, kitchens, affordable furniture, designer furniture, delivery and service, good reviews and Turkish furniture. One question was typed the way people type into a search engine: "sofa kaufen köln empfehlung".
• ChatGPT: on chatgpt.com without logging in. This is what anyone using the free version sees, and no chat history can influence the result. Each question was asked twice, each time in a new chat.
• Claude: with web search switched on and the location set to Cologne, using the Opus 5.5 and Sonnet 5.5 models.
• Gemini: we received only a few answers, so we treat them as a first observation.

From every answer we recorded the businesses recommended, the ratings shown and the sources cited. Different spellings of the same store, such as "TRÖSSER Köln-Marsdorf" and "TRÖSSER", were merged.

What does ChatGPT rely on when it recommends a local business?

ChatGPT shows a map for local questions and puts ratings next to the businesses it names. In our data, a map appeared in almost every answer. Three out of four recommendations came with a star rating and one in three with a review count. ChatGPT does not say where those ratings come from.

The answers are short and selective. ChatGPT named four or five businesses per answer on average, and around 55 different stores across all answers. Being on that short list is what counts, because a customer who sees four names may not look any further.

The sources below each answer show what else ChatGPT reads:

• the businesses' own websites,
• city guides such as mrkoeln.de, koeln.de and Wogibtswas,
• the DISQ customer satisfaction survey in several answers and Trusted Shops in one.

In other words, ChatGPT does not rely on the map alone. What other websites publish about a business also feeds into the recommendation.

Which furniture stores in Cologne did ChatGPT recommend?

The stores named most often were home24 Outlet, IKEA, TRÖSSER and Revive. The table shows the share of answers in which each store appeared and the number of reviews ChatGPT displayed for it:

Store	Share of answers	Reviews shown
home24 Outlet	50%	7,449
IKEA	42%	20,324
TRÖSSER	32%	2,171
Revive	29%	1,496
BoConcept	24%	161
smow	24%	67
Möbel Leyendecker	16%	not shown
porta	16%	1,443

The four stores at the top had between 1,496 and 20,324 reviews. Further down, BoConcept and smow appear with far fewer reviews, but both are established chains.

Do ChatGPT and Claude recommend the same businesses?

Partly. The stores both assistants named most often largely overlap: IKEA, home24 Outlet, TRÖSSER, Revive, BoConcept, smow, Höffner and porta. They have one thing in common: they are either well-known chains or have more than a thousand reviews.

How they get there differs. Claude writes much longer lists, around ten stores per answer and about 200 different stores across all answers:

Store	Share of Claude's answers
IKEA	57%
Höffner	51%
BoConcept	46%
porta	39%
Revive	32%
XXXLutz	26%

Claude ran four or five web searches per answer on average but never showed a map, and ratings appeared in only about a third of its answers. Most of its answers advised checking the address and opening hours before visiting, and a few said openly that the list was partly based on general knowledge.

Gemini, in the few answers we received, worked differently again: no map, no ratings, and in the first answer business directories such as WerKenntDenBesten, ThreeBestRated and Kölner Branchen as sources. With so few answers, we draw no conclusions about Gemini.

Does the wording of the question change the recommendation?

Yes, considerably. A single word can replace most of the list. For the same city on the same day, these were the results for questions asked in German:

• Cheap: when we asked for affordable furniture of good quality, ChatGPT named second-hand and charity furniture stores such as Sozialistische Selbsthilfe Köln, fairstore and Emmaus. Of the big names, only home24 Outlet and IKEA remained. Asked a similar question in Turkish, ChatGPT added the discount chains POCO and ROLLER; in English, POCO appeared once.
• Design: when we asked for high-end designer furniture, the list changed almost entirely: pesch, smow, Qvest, Rheinstore and ligne roset.
• Delivery and service: both answers named TRÖSSER, Möbel Leyendecker and Höffner, only in a different order.
• Good reviews: when we asked which shop has really good reviews, every store named had between 4.6 and 4.8 stars: Revive, TRÖSSER, Möbel Leyendecker, BoConcept, home24 Outlet and home design by akaslan.
• Fast delivery: porta, Möbel Boss, home24 Outlet and ROLLER led. One answer also included a removal company, a sign that the AI sometimes reads a question differently and recommends businesses from another sector.
• Kitchens: when we asked about a new kitchen, ChatGPT named only kitchen studios and none of the furniture stores from the table.

Repeating a question does not give the same result either. When we asked exactly the same question a second time, about a third of the stores from the first answer were missing. Visibility in AI answers therefore cannot be judged from a single test, only across several questions and repetitions.

Does the language of the question matter?

For general questions, very little. For specific ones, a great deal. Asked in Turkish for a good or reliable furniture store, ChatGPT named almost the same stores as in German: home24 Outlet, TRÖSSER, IKEA and BoConcept. The only difference was the Turkish brand Bellona in one answer.

The picture changed completely when the question was about Turkish furniture. Asked in Turkish where to buy Turkish furniture in Cologne, ChatGPT returned an entirely different list: Istikbal, Can Möbel, home design by akaslan, Bellona, Bessi Home and Nevizade. None of the large stores from the table appeared.

In English, IKEA appeared in all eight answers, whatever we asked.

For a business with a clearly defined customer base, the lesson is that the content of the question matters more than its language. A store that uses its customers' own words, such as "Turkish furniture" or "storage bed", can appear in lists where the large chains are absent.

Can a relaunched website be recommended within weeks? The Plaza Crown example

Yes, at least for its own speciality. Plaza Crown is a furniture store with a showroom in Cologne-Nippes and a client of MAKK Digital. We rebuilt its online shop from scratch and manage it.

The new website went live on 23 August 2026, with pages released in stages. According to Search Console, 149 pages were in Google's index on 14 October, most of them added in September. On the day of our test, most pages had been on Google for only two to six weeks.

Even so, when we asked in German "Where can I find a good box-spring or upholstered bed in Cologne?", ChatGPT recommended Plaza Crown in both answers. In the first answer it was fifth, with plazacrown.de cited directly as the source. In the second it moved to first place, appeared on the map with 4.9 stars and was recommended "for a beautiful upholstered bed". That second answer came after ChatGPT announced that, without a login, it had switched to a simpler model. We therefore read the result with some caution.

In our view, the most likely reason is that the website describes each product range separately and clearly. Upholstered beds, storage beds, fabrics and sizes each have their own pages. When an AI reads them, it can identify Plaza Crown as a good match for exactly this question.

It is just as clear where there is room to grow. Plaza Crown does not yet appear for general questions, sofas or kitchens, nor in Claude's answers. There, stores with many reviews and well-known chains lead. The next steps for Plaza Crown are therefore more reviews and stronger listings on maps and directories.

How do you get your business recommended by ChatGPT?

From our data we derive five practical levers. None is enough on its own; together they improve your chances of being named.

1. Keep your listings complete. Your Google Business Profile and every map and directory listing need the right category, opening hours, photos and products. ChatGPT shows a map for almost every local question. Our article Why local SEO matters: getting found where your customers are explains the basics.
2. Ask for reviews regularly. Most of the frequently recommended stores have plenty of reviews. A short message after delivery or a card in the shop is enough. Do not offer discounts or gifts in return; Google prohibits incentivised reviews. Replying to reviews also shows that the profile is actively looked after.
3. Give each product range its own page. In the Plaza Crown example, a small shop that gives each product range its own page was first for exactly that question in one of two answers. "Three-piece suites" and "storage beds" belong on separate pages.
4. Be present on other websites. ChatGPT's sources included city guides, customer surveys and trust seals such as Trusted Shops. Your business name, address and phone number should be identical everywhere.
5. Answer your own customers' questions. If you serve a specific group, such as Turkish-speaking customers, offer a page in their words. There is far less competition for those questions.

We cover the background in more detail in How to get your business to show up in AI search.

Short on time? How can MAKK Digital help?

None of these is a one-off job. Keeping listings current, collecting reviews and shaping pages around your customers' questions takes steady effort, and in a busy shop it is exactly what tends to slip. We take on the whole process for you:

• Baseline: we measure which questions your business is named for on Google, Google Maps and ChatGPT, and where your competitors lead.
• Listings: we complete your Google Business Profile and the relevant directories and make sure your name, address and phone number match everywhere.
• Reviews: we set up a routine for asking your customers for reviews within the rules and prepare your replies.
• Website and content: we describe each product range on its own page in your customers' words, in English, German and Turkish if needed.
• Tracking: we measure regularly what changes on Google and in AI answers and report it in plain language.

This is the path we are following with Plaza Crown: website and product pages first, then reviews and listings.

To see where to start, try our free AI visibility check. We put your customers' questions to ChatGPT and analyse 25 answers, review your Google Business Profile and Google Maps listing, and email you the report within two working days. You can find out more about the service on our SEO and performance marketing page or contact us directly.

What are the limits of this study?

This study is not a fixed ranking but a snapshot. Please keep the following in mind:

• Scope: measured on 14 October 2026, in Cologne and for furniture only. We asked 19 questions twice in ChatGPT and twice with each of two Claude models, giving 114 answers from these two systems. The two answers from Gemini count as a first observation. Results may differ on another day or in another sector.
• Variation: AI answers change even for the same question, which is why we report shares rather than individual answers.
• ChatGPT without a login: logged-in users may see different results because of their history and other models. Without a login, ChatGPT switched to a simpler model at one point; we have recorded this.
• Claude via the API: we measured Claude through the API, not the app. The app may give different answers.
• The API and the app do not give the same results: in a measurement via the OpenAI API the day before, Plaza Crown was named much more often for similar questions. What users see in the app does not always match what the API returns.

A store's appearance in these lists is not a verdict on that store, and its absence says nothing about its quality. We are only reporting what the AI recommended on that day.