Kiwop Labs · Study
AI in Spanish ecommerce 2026: what the stores show the AI
We measured 300 Spanish online stores on 1 and 2 September 2026 and checked every data point again on 5 September 2026. This is not a survey: it is what any AI crawler or agent finds when it arrives. Method, dataset and scripts, all open.
Five things the data says
- Spanish ecommerce is not shutting the door on AI. Only 10.7% block an AI crawler in their robots.txt and 5% block GPTBot. When a store does block, it mostly blocks CCBot, ClaudeBot, Bytespider and Meta's agent.
- But it is not talking to it either. Of the 216 product pages we could locate, only 27.8% expose the product with schema Product. The AI gets in, but on seven out of ten pages it finds no price, no stock and no reference it can read without guessing.
- llms.txt has gone from curiosity to norm in two months, and half the stores never decided it. In the July pilot, 1 in 20 had one. Today 26% of the 300 serve one, but 13% is the automatic template Shopify gives every one of its stores. 13% wrote it themselves.
- Size explains almost everything, in both directions. The big ones expose more product with schema (34.4% against 17.7% of the small ones) and the small ones have more llms.txt (36.6% against 17.3%), because they are the ones running on Shopify.
- Automated support is still a minority: 23.2% have a chatbot detectable on the home page (Oct8ne, Zendesk and Salesforce, in that order) and 30.6% a search engine with personalisation. Neither of the two is generative AI in itself.
The llms.txt figure, with the small print
We detected an llms.txt on 86 stores and read every one of them again. 78 are real text files. Of those, 39 start with "Agent Instructions" and are the template Shopify serves to all its stores without anyone asking for it (of the 44 stores in the sample we identified as Shopify, 39 have it). The other 39 were written by the store. The rest were an ordinary page returning a 200 or stopped responding, and are discounted. Publishing "29 % already talk to AI" without that distinction would be false: the honest figure is 13% by the store's own decision and 13% by platform default.
Among those that wrote it themselves: MediaMarkt, SHEIN, Samsung, Casa del Libro, PlayStation, Sprinter, Tradeinn, ManoMano, LG, General Optica.
By size
Three bands of about a hundred stores each, by estimated traffic in the sampling frame. Percentages over the band total; schema Product, chatbot and search over the stores where they could be measured.
| Band | n | Measurable | Block an AI crawler | Block GPTBot | Verified llms.txt | Own llms.txt | Schema Product (pages) | Chatbot | Personalised search |
|---|---|---|---|---|---|---|---|---|---|
| Large | 98 | 86 | 11.2 % | 8.2 % | 17.3 % | 13.3 % | 34.4 % | 22.1 % | 25.6 % |
| Medium | 101 | 89 | 8.9 % | 2 % | 23.8 % | 13.9 % | 32.9 % | 28.1 % | 39.3 % |
| Small | 101 | 96 | 11.9 % | 5 % | 36.6 % | 11.9 % | 17.7 % | 19.8 % | 27.1 % |
| Total | 300 | 271 | 10.7 % | 5 % | 26 % | 13 % | 27.8 % | 23.2 % | 30.6 % |
By sector
Ten sectors. Cells with fewer than 25 stores are indicative, and we say so: they show direction, not fine differences.
| Sector | n | Block an AI crawler | Verified llms.txt | Own llms.txt | Schema Product (pages) | Chatbot |
|---|---|---|---|---|---|---|
| Generalist and marketplaces | 62 | 12.9 % | 14.5 % | 14.5 % | 29.3 % | 16.7 % |
| Home | 35 | 20 % | 42.9 % | 22.9 % | 30.8 % | 38.2 % |
| Fashion | 33 | 6.1 % | 33.3 % | 9.1 % | 40.9 % | 25 % |
| Sport | 28 | 14.3 % | 17.9 % | 10.7 % | 45.5 % | 30.8 % |
| Electronics | 28 | 10.7 % | 28.6 % | 25 % | 21.1 % | 16.7 % |
| Jewellery and luxury | 28 | 7.1 % | 53.6 % | 14.3 % | 27.8 % | 33.3 % |
| Beauty * | 24 | 0 % | 20.8 % | 0 % | 17.6 % | 35 % |
| Health and parapharmacy * | 24 | 16.7 % | 12.5 % | 8.3 % | 31.6 % | 8.7 % |
| Wine and drinks * | 23 | 8.7 % | 26.1 % | 13 % | 10 % | 0 % |
| Food * | 15 | 0 % | 6.7 % | 0 % | 8.3 % | 33.3 % |
* n < 25
Who gets blocked
The AI crawlers blocked most often in the robots.txt files of the sample: CCBot (20), ClaudeBot (18), Bytespider (18), meta-externalagent (18), GPTBot (15), Google-Extended (14). GPTBot does not top the list: blocking concentrates on the training crawlers (CCBot, Bytespider) and on the Anthropic and Meta agents.
Who answers
Chatbot providers detected on the home page: Oct8ne (27), Zendesk (17), Salesforce Chat (12), Gorgias (6), Genesys (2), Tidio (2). This is detection by script signature, so a chatbot loaded after the first interaction does not show up.
What to do with this if you run a store
- Let in whoever you want recommending you. Check your robots.txt: blocking GPTBot, ClaudeBot or PerplexityBot takes you out of the shopping answers of those assistants, and it does not protect you from the training crawlers unless you name them.
- Explain the product in a language the machine reads. Schema Product with price, availability, brand and reference on every product page, and a feed consistent with what the user sees. That is what separates a mention from a sale.
- Check what you are already saying. If your store runs on Shopify, you have an llms.txt you did not write: read it. And if you have none, write a short, true one rather than a long, templated one.
Method
- Sample: 300 online stores trading in Spain, chosen by a public sector × size rule over the Semrush Trending Websites Spain frame (June 2026), with documented screening and rescue. 53 domains were left out because they were not measurable (global marketplaces with no locatable product page, apps and so on).
- Two passes: an HTTP request identified as KiwopResearchBot and a headless Chromium session with a normalised Chrome user agent and an honest navigator.webdriver, no fingerprinting evasion. Signals are merged: a store counts as accessible if it answered either of the two.
- What is measured: existence and rules of robots.txt for 14 AI crawlers, existence of llms.txt, JSON-LD types on the home page, location of a product page and whether it carries schema Product, chatbot signatures and personalised search signatures on the home page.
- Later verification: every llms.txt detected was read again and classified (own, platform template, false positive, no response). The platform of each store is detected by signatures on the home page with an identified fetch, and is indicative only.
- Nothing was bought, no account was created and no CAPTCHA was solved. An anti-bot block is a result, not an obstacle to get around.
Limits
- It measures what the store exposes, not what it gets: schema Product on a page does not guarantee that an assistant will cite it, and its absence does not prevent it.
- It is a two-day snapshot. A store with its CDN in challenge mode that day comes out as not measurable; that is why we publish the accessibility figure instead of hiding it inside the percentages.
- Chatbot and search detection goes by known script signature: it underestimates anything built in-house.
- Platform detection gets no answer on a third of the stores (they block the identified fetch) and is not guessed: it is published as "no response".
What comes next
The second part is the one nobody has published for Spain: can an AI agent with a browser reach the basket with a specific product? The protocol is written and preregistered before measuring (40 stores, identical task, closed result categories, no anti-bot evasion). It is published in the fourth quarter of 2026, and level 1 repeats every quarter to build a series.
Dataset and scripts
- Dataset for this edition (JSON, one row per store with every signal) ·
2026-09.json - Index of editions (JSON)
- Measurement, verification and agentic protocol scripts (GitHub)
How to cite
Kiwop Labs (2026). AI in Spanish ecommerce 2026: what the stores show the AI. https://www.kiwop.com/en/labs/ai-ecommerce-spain
Data and text under a CC BY 4.0 licence: you can use and quote them, with a link to this page.