Measurement Practice
When does a well-known shop drop out of AI recommendations?
What we measured
Twelve German-language purchase queries, six access paths to AI systems, five days in August 2026, 465 answers. We did not give the models any competitors. Among 47 brands counted, the shop ranked 13th: known, but not leading.
Same brand, two pictures
| Questions | Examples (paraphrased, asked in German) | Mention of the shop |
|---|---|---|
| 4 questions in which the shop appeared regularly, mainly about the manufacturer, a product feature or delivery time | “Which Swiss manufacturer has the best online shop?”, “Which online shop delivers to Zurich within 3 days?” | 28 of 152 answers · 18% (95% interval 13–25%) |
| 8 questions in which it hardly appeared, mainly about the best shop for a product, a shop comparison or returns | “Which online shop offers a returns option?”, “Best value for money, compare at least three shops” | 8 of 313 answers · 3% (95% interval 1–5%) |
In 143 answers, a model committed to a single shop. None of them picked this shop.
One possible reading, which we treat as a hypothesis: the models know the provider as a manufacturer, but not as a shop where you order a specific product.
The returns cascade
| Question (paraphrased, asked in German) | Mention | 95% interval |
|---|---|---|
| Which online shop delivers to Zurich within 3 days? | 6 of 43 · 14% | 7–27% |
| … the same, plus a returns option | 2 of 44 · 5% | 1–15% |
| Which online shop offers a returns option? | 0 of 43 · 0% | 0–8% |
The more weight returns are given in the question, the less often the shop appears. Three questions, however, show a direction, not an established difference: the intervals of the first and the last row only just overlap.
What the shop's own pages said
The following is not a measurement result but a finding that anyone could check on the pages. On returns, there were three statements at the same time: the terms and conditions on the brand site ruled out taking goods back. The terms and conditions on the shop site ruled out exchanges. In the checkout process there was a withdrawal form with a 14-day period under German law. On top of that came differing shipping costs and free-shipping thresholds between the two domains, and differing information on payment methods.
A person usually reads one of these pages. An AI system can read all of them at the same time.
The measurement does not show whether these contradictions are the reason. Two best-shop questions without any reference to returns also produced zero mentions, and the delivery-time question reached 14% despite the differing shipping costs.
A second case: different provider, same pattern
In a second pilot measurement, for a Swiss manufacturer that sells to businesses and to private customers, the mention depended not on the type of question but on the product asked about:
| Products asked about | Mention of the manufacturer |
|---|---|
| Products for business customers (5 questions) | 20 of 177 answers · 11% (95% interval 7–17%) |
| One product for private customers (6 questions) | 0 of 205 answers · 0% (95% interval 0–2%) |
| Other products and general supplier search (14 questions) | 15 of 489 answers · 3% (95% interval 2–5%) |
The same manufacturer sells the product for private customers in its own shop. In more than 200 answers to questions about it, the manufacturer did not appear. If you only measure products for business customers, you see a provider that is in the field. If you ask about the product that private customers buy, you do not see it.
This fits the reading above, which we continue to treat as a hypothesis: the models know both providers as manufacturers, but not as a shop where you order a specific product. With this provider, the shop also delivered an empty page without JavaScript; more on this in What do AI systems search for before they answer, and what does that mean for SEO?
What follows from this
- Measure the questions where the choice is made. Asking about the best Swiss manufacturer's shop, we would have measured 23% mentions; asking about returns, zero. Both numbers are correct. What matters is which questions your customers ask.
- First, clean up what is in your own hands. Return period, shipping costs, delivery time and payment methods belong on your own website, unambiguous, free of contradictions and machine-readable. That pays off even without AI.
- Then measure again. With the same questions and the same systems. Only that shows whether the answers move.
More on how we derive the questions we measure: Which prompts actually count?