Who does AI recommend when someone asks for an online shop?

Anyone who types “Which online shop is good for running shoes?” into a chat window today does not get ten blue links but an answer with three to six names. For a retailer the question is therefore no longer only whether they are on page one of Google, but whether they appear in that answer. For the German market there have so far been mainly assertions about this. Here are figures.

Four findings

• The answers rely on review portals, not on the shops. The most frequently linked source across all 120 answers is trustedshops.de (34 answers), followed by dtgv.de (22), getestet.de (20) and chip.de (16). Whoever appears in a comparison gets into the answer. Whoever merely has a good website does not.
• ChatGPT and Perplexity recommend different worlds. Amazon appears in 30 of 60 answers from ChatGPT and in 6 of 60 from Perplexity. Perplexity puts OTTO at the top instead (19 of 60) and cites editorial sources.
• For niches ChatGPT links the Google Maps entry, not the website. For office supplies and tea, ChatGPT pointed to a shop domain in none of its answers but to 16 Google Business Profiles. A well-kept profile is the way in there, not the home page.
• Each sector has two or three fixed names. Fashion: Zalando and About You in all six answers. Books: Thalia and Hugendubel in all six. Pet supplies: Fressnapf and zooplus in all six. Behind them it thins out quickly.

Overall: who appears in the most answers

Share of the 120 answers in which the shop is named or linked as a source, across all 20 categories.

By category: the three shops named most often

In brackets: in how many of the six answers per category the shop appears.

Two models, two worlds

Amazon appears in 30 of 60 answers from ChatGPT and in 6 of 60 from Perplexity. The ten most frequent shops per model:

Who gets linked as a source

Domains the answers refer to, counted per answer. Review portals and media are marked.

What this means for a business

The models do not invent their recommendations. They read what is written about shops on the web and give it back in condensed form. Four things follow from that, and each can be verified.

1. Get into directories and comparisons. The sources table shows where the answers come from. A retailer profile at Trusted Shops, an entry in a sector comparison or a test by a specialist editorial team weighs more in the answer than any rework of your own home page.
2. Look after the Google Business Profile. Name, category, address, products, photos, reviews. For small retailers it is the entry ChatGPT links when it finds no comparison.
3. Look at the two models separately. Anyone who wants to appear in ChatGPT needs a presence on marketplaces and in directories. Anyone who wants to appear in Perplexity needs editorial teams and review portals that name them. There is no single measure for both.
4. Publish figures of your own. Answers that need a figure link the source of the figure. A retailer who discloses its own measurements, price series or analyses gets cited for them. This page is itself an attempt in that direction, and we measure every month whether it works.

Method

Questions. Three questions for each of the 20 sectors, phrased the way a buyer asks them, always with reference to Germany, never with a brand name, and asked in German. All 60 questions are at the foot of this article.

Models. ChatGPT (gpt-4o-mini, web search switched on) and Perplexity (sonar), queried on 15 September 2026 through the DataForSEO interface. Each question was put to each model once: 120 answers.

Counting. A shop counts in an answer if its name is in the text or its domain is linked as a source, at most once per answer. Linked sources were counted separately and divided into shops, price comparison sites, and review portals, media and forums. Google Maps references were not counted as a source but recorded separately.

Limits. Six answers per sector are a sample, not a full survey. The models do not answer the same way on every call; differences of one mention are therefore not robust, differences of four or more are. A mention is not a recommendation in the literal sense but an appearance in the answer. Two models were measured; Google AI Overviews, Gemini and Claude are missing. The measurement is repeated monthly, and a trend will then appear here.