GEO SEO for webshops: 70 Dutch stores tested for AI access

GEO SEO for webshops starts with access: in Oase Creative's test on 24 October, 66 of 68 Dutch webshops allowed AI search bots in robots.txt, yet no AI user agent got through to the product pages of Coolblue and Jumbo.

What does GEO SEO mean for a webshop, and where does it start?

GEO SEO for a webshop does not start with copy. It starts at the door. An AI search engine can only recommend a product if its bot is allowed to fetch the product page and finds something useful there. We measured exactly that on 24 October 2026 across 70 Dutch webshops: the ten largest from the Twinkle100 ranking and 60 smaller stores from the winners and finalists of the Dutch Shopping Awards.

The result in three lines:

• The front door is open. 66 of the 68 measurable webshops allow all five search and user bots from ChatGPT, Claude and Perplexity in their robots.txt.
• The inner door is sometimes shut. Coolblue and Jumbo do not mention a single AI bot in robots.txt. Yet in our test no AI user agent got through to their product pages. A normal browser did.
• We have only checked half the shelf. Whoever gets in almost always finds a name and a price. Our test says nothing yet about returns and delivery.

So GEO SEO is an access question first and a content question second. If you want to know how AI search works first, read our article on GEO for webshops. This article covers the step before that: can the bot get in at all?

How did we test the 70 webshops?

We measured what a webshop promises in robots.txt, what the server does next, and which extra files are ready for AI agents.

Sample. The top 10 of the Twinkle100 2026 plus 60 webshops from the winners and finalists of the Shopping Awards 2025. That covers both the big players and professional small and mid-sized stores. 68 of the 70 were measurable: one returned a challenge page on robots.txt itself, one returned a server error.

Layer 1: robots.txt. For each domain we fetched robots.txt and determined, for twelve AI user agents, whether they may fetch the homepage. The parser follows the official robots.txt rules: groups per user agent, fallback to the general group, longest rule wins. We checked five files by hand. The classification was right in all five.

Layer 2: the product page. For twenty stores, ten from the top 10 and ten smaller ones, we requested one product page as Chrome and as five AI bots: GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot and PerplexityBot. OpenAI and Perplexity publish their user agents. Anthropic only names the ClaudeBot token; the full string comes from udger.com. We did this two ways, with a plain HTTP client and with a browser without JavaScript: 240 requests. The figures below come from the browser run. In the plain run Shopify rate-limited Chrome too, so that run tells us nothing there.

Layer 3: the agent layer. For all 70 domains we checked for an llms.txt, an agents.md or a UCP profile (/.well-known/ucp).

One limitation up front: we request pages from our own IP address. Real bots come from their own addresses. A CDN may still let verified bots through by IP, or be stricter. Treat layer 2 as an indication: it shows which stores filter on user agent. It does not tell us what the real GPTBot receives.

Do Dutch webshops block AI bots in robots.txt?

Barely. 58 of the 68 measurable webshops block none of the twelve AI bots we tested. Those that do block mostly hit the bots that collect training data, not the bots that look up answers.

The two exceptions on the search side are Amazon.nl, which refuses every AI bot, and Deloox, which blocks OAI-SearchBot. At every other store, robots.txt allows the search and user bots of ChatGPT, Claude and Perplexity.

Four stores separate training from search: they block GPTBot and leave OAI-SearchBot open. Those are Zalando, Breiwebshop, Gewoonstijl and Farfallina. All four block exactly the seven bots from our list that also appear in Cloudflare's managed robots.txt. That looks like a default setting, not a deliberate decision.

Nine stores explicitly grant AI bots access in a rule of their own, such as MediaMarkt. Not a single store blocks AI through the general * group.

How does that compare with other countries and sites?

For search bots, Dutch webshops sit on the open side, level with Canadian retailers. They block training bots more often: 6 of 68 keep GPTBot out, against 1 of 40 in Canada.

For comparison, two measurements by others:

• A study of Canadian websites found that 5% of 40 retail sites keep out at least one AI crawler, against a much higher share among news media.
• Cloudflare writes that less than 1% of its sites block search bots, while 17% take some measure against training bots. Since 15 September 2026, search, training and agents can be set separately there.

In our own sample the top 10 blocks more often than smaller stores: three of the ten big players block at least one AI bot, against 7 of the 58 smaller stores. With ten stores, that is a hint, not a trend.

The Canadian study's bot list may differ from our twelve, so read the comparison as a direction. The conclusion for GEO SEO holds: robots.txt is almost never the reason ChatGPT or Perplexity fails to mention a Dutch webshop's product. When it goes wrong, it goes wrong deeper.

Why does an AI bot sometimes still not get in?

Because robots.txt is a promise, not a lock. Real access is decided by your server, your CDN or your firewall. And those can refuse a bot that robots.txt politely allows.

We could measure 17 of the 20 product pages. Albert Heijn and Zalando also refused a normal browser from our address, and Picnic has no public product pages. Of those 17, four refused at least one AI bot that Chrome got past:

Coolblue and Jumbo say yes in robots.txt and no on the server. None of the five user agents got through to the product page, including ChatGPT-User and PerplexityBot. Those are the bots that fetch a page live when a customer asks in a conversation what a coffee machine costs at Coolblue. We repeated the Coolblue test by hand: the OAI-SearchBot user agent got a 403 again, the browser a 200. If their CDN also blocks the real ChatGPT-User, ChatGPT cannot fetch a live price there. TechCrunch reported in October that such blocks increasingly run through the firewall or the CDN, not through robots.txt.

Sokkendirect shows what a deliberate choice looks like: training bots get a challenge, search and user bots get through.

What does an AI bot find once it is in?

Usually enough to mention a product. Of the 17 measurable product pages, every page had the price in the raw HTML, and 16 had Product schema (JSON-LD) with a price. The exception was Amazon.nl, with no Product schema.

In the top 10, all five AI user agents could read the price at 5 of the 7 measurable stores. Among smaller stores it was 8 of 10. The five Shopify stores let every bot through in the browser run. So smaller stores do at least as well as the big players here.

Price alone is not enough to choose. An agent buying on someone's behalf also wants to know what delivery costs, how fast it arrives and whether it can be returned. We have not yet measured whether Dutch stores offer that in readable form. What an agent looks for on a product page is covered in our article on what AI agents read on your product page.

Picnic is a separate case: its range only exists in the app. For an AI search engine it simply does not exist. That is not a robots.txt mistake but the result of a channel choice.

Does an llms.txt or UCP profile improve GEO SEO?

Not by itself. About one in four stores has such a file, and usually the platform put it there, not the store.

We found a self-written llms.txt at 9 of 70 stores, including MediaMarkt and Wijnvoordeel. The top 10 barely does anything here: one in ten has an llms.txt, against 18 of the 60 smaller stores. Bol, Coolblue, AH, Jumbo, Picnic, Zalando and Belsimpel have none of the three files.

A measurement of American Shopify stores shows that present is not the same as useful. Citecycle examined 200 large Shopify stores. 183 had a UCP profile. Even so, the policy tool for agents returned the return window for only 33 of 176 stores. Shipping cost plus delivery time came back for 4 of 176. At 64 of 197, robots.txt itself blocked the policy pages.

At those stores the technology is switched on and the content is missing. We wrote about the Shopify template before in a separate piece on agents.md left at factory settings. Webshops on Magento, Lightspeed or WooCommerce usually do not have this layer at all.

How do you check in ten minutes whether AI search engines can read your webshop?

Four steps, in this order. Each step catches a mistake the previous one misses.

1. Read your robots.txt per bot. Deliberately open search bots (OAI-SearchBot, Claude-SearchBot, PerplexityBot) and user bots (ChatGPT-User, Claude-User). Training (GPTBot, ClaudeBot, CCBot) is a separate choice. Write that choice down, so nobody reverses it by accident later.
2. Request one product page as an AI bot. Use curl -A "<user-agent>" or a similar tool, for each of the five bots. A 403, a challenge page or a dropped connection while your browser gets in means your CDN or firewall filters on user agent.
3. Line up robots.txt and your CDN. If you use Cloudflare, check its AI bot settings and the managed robots.txt. Since September, search, training and agents can be set separately there. Make sure both say the same thing.
4. Check what is on the shelf. Price in the HTML, Product schema with price and stock, and readable delivery and returns pages. If you block /policies/ in robots.txt, remove that rule.

Only then does work on content, category pages and quotability pay off. For that we listed seven elements of a GEO-ready category page. Whether your AI traffic actually grows afterwards, you measure with AI referrals in GA4.

Where is the work once your robots.txt is already open?

At most Dutch webshops the door is already open. The work sits behind it, in three places.

• In the server. Coolblue and Jumbo refused every AI user agent on a product page in our test, while their robots.txt names no AI bot. Whether that is deliberate, we do not know. Anyone running the same CDN template may have the same issue without knowing it.
• In the default setting. The four stores that separate training and search probably do so through a Cloudflare template. An agent layer that exists usually comes from Shopify. If you do not choose, you get your supplier's choice.
• In the content. Price is almost always there. Delivery, returns and variants are the data an agent chooses on, and they are the least measured.

Our own position: search bots and user bots open, training a deliberate choice per brand, and robots.txt and CDN in line. If you want to be mentioned in AI answers, you first have to be readable. That ranking well in Google is different from being named by AI, we showed earlier in our article on ranking first in Google and still missing from the AI answer.

How does Oase Creative help with GEO SEO for your webshop?

We run the same test on your webshop: robots.txt per bot, your product pages requested as five AI bots, and the agent layer behind them. It takes us a few minutes and shows where a bot gets stuck. Then we line up robots.txt, CDN and your delivery and returns pages. We fill the agent layer with your brand and your terms.

That work falls under our SEO service. If your store runs on Shopify, we include it in our Shopify work. Want to know when a CDN update shuts the door again? Then we run the check automatically, through AI automation.

This test is from 24 October 2026. We will repeat it with the same 70 webshops and update this article when the numbers shift.

Send Louie the link to one product page. We request it as five AI bots and send you back what each bot got.