A quarter of small Swedish web shops never show up when ChatGPT is asked

What the study did

As EIN Presswire reported, KLOTA AB, a digital marketing agency in Borås, Sweden, ran 189 small and mid-sized Swedish online stores through its own AI visibility analysis on November 4, 2026. Each store had annual revenue between SEK 3 million and SEK 30 million.

For every store, the agency wrote three shopping questions in Swedish based on the homepage, leaving out the store's name. It put them to ChatGPT and to Claude, both with web search switched on, then checked whether the store showed up in the replies. In all, that produced 1,123 answers.

How many stores went missing

The headline figure: 28 percent of stores were not named in any of ChatGPT's three answers. Claude left out 29 percent. About 19 percent were missing from both services.

The two assistants did not agree with each other. Claude named stores in 47 percent of its answers, and ChatGPT in 39 percent. One store in five was mentioned by only one of the two, so a store can look fine in one tool and absent in another.

When a store was missing, the assistant often pointed elsewhere. In four cases out of ten, it recommended one of Sweden's 100 largest online retailers instead.

Text, not access, set stores apart

Getting in the door was rarely the issue. 98 percent of stores allowed the crawlers (automated visitors) that ChatGPT, Perplexity and Claude use to read pages, and no store blocked them in server tests. About one in three had an llms.txt file, a plain-text note meant to guide AI systems. It made no visible difference: 16 percent of stores with the file were never mentioned, and 15 percent without it.

What did matter was how much a page said. Stores with fewer than 220 homepage words were named in 2.4 of 6 answers on average. Those with more than 970 words reached 3.3. Stores with at least one heading phrased as a question averaged 3.4, against 2.6 for those with none.

Combining the two traits helped most. Of the 50 stores with at least 500 words and a question heading, just 3 were never mentioned. Among stores with little text and no such headings, one in four was never mentioned. Revenue did not clearly matter: bigger stores in the sample were not reliably named more than smaller ones.

Gaps an AI cannot fill in

The study also looked at what homepages leave out. 40 percent showed no concrete terms for shipping, delivery or returns. Only 11 percent had fully completed structured data, the hidden labels that describe a business to machines.

KLOTA's founder, Anders Karlsson, put the logic simply: "If there is nothing to quote, there is nothing to recommend." His advice is to test visibility with the questions real customers ask, in more than one AI service, and to repeat the test a week later, since answers vary between runs.

How far to trust it

Read this with care. KLOTA ran and paid for the study, and it sells services in this field. The release says the stores were picked in random order and that KLOTA's current and former clients were excluded. It also says the gaps for text length and question headings are statistically significant in a Mann-Whitney test, a method for comparing two groups.

But the authors themselves say the results show correlation, not cause. AI answers change between runs, and answers fetched through a developer interface may differ from those in the consumer apps. The release says the findings were first covered by the Swedish trade site ehandel.se, and that the full method is published in Swedish on klota.se.

For context, the release cites the Swedish Internet Foundation's report Svenskarna och internet 2026, which says 52 percent of Swedes use ChatGPT.