Is your site ready for AI agents? A test on 8 Turkish sites

On 5 November 2026 Visa published a test in which 22 AI models were given 15 shopping tasks. 42.8% of the runs failed, and about half of the failed runs still ended in a purchase. By Visa’s criterion, when nothing matches what the customer asked for, the right move is not to buy.

We asked the same question on Turkish businesses’ own websites. On 9 November 2026 we gave Claude the kind of tasks a real customer would give it on eight sites: book a table, get an appointment, find a gift within a budget. The hardest moments for the agent came where what the user wanted and what the site offered drifted apart. The requested time did not exist, nothing fitted the budget, or the site did not say what the price covered.

In agentic shopping, the right decision is sometimes not to buy

In Visa’s test, 22 models from 12 labs each made three attempts at 15 shopping tasks, such as buying a TV or a laptop. To pass, a model had to find every one of the customer’s requirements and make the correct buy or no-buy decision. The overall pass rate was 57.2%. 54% of the failed runs still ended in a purchase.

According to news coverage (a secondary source), the always-on agents OpenAI described as Dots at its developer event on 29 September 2026 each run on their own cloud computer and browser. We did not measure from which country such a browser reaches your site. The coffee example below shows why that question can matter to a business.

What did we test on the 8 sites?

We gave each task in one sentence, the way a customer would write it to an agent. The agent reached the final screen on seven of the eight sites. By final screen we mean the screen just before the last button, where the site asks for contact, address or card details. The only site it could not get through was the coffee shop that does not sell to browsers outside Turkey.

The hair salon was the only flow where the agent never had to guess. It let the agent choose the service, the stylist, the day and the time, and showed booked slots as booked. The hotel price appearing in euros and the coffee site’s price in pounds most likely follow from our browser being outside Turkey.

Where the agent decided on its own

When the requested option is missing, the agent picks without asking

The restaurant’s Saturday evening slots of 18:00-20:00, 18:30-20:30 and 19:00-21:00 were full; the free one was 21:00. There was no 20:00 option. The agent chose 21:00 without asking the user. In real use, that is a decision the user should have been asked about.

The florist had nothing under 3,000 TL in its birthday category; the cheapest item was 3,995 TL. By Visa’s criterion, the right move here was to stop and ask the user. In our test the agent continued with the 3,995 TL item up to the delivery address screen, to see the rest of the flow. That was our deliberate choice; in real use, carrying on with an item 995 TL over budget would have been wrong.

Preselected time slots and ready-made bundles

At the same florist, the delivery day and the time window “10.00-16.00” came preselected. The task said “evening”; the agent changed the window to 14.00-20.00. The site had no exact match for “evening”, so this was a guess too. The 14:00 cut-off for same-day delivery appeared only in the banner at the top of the page.

The “Tamamlayıcı Ürünler” (complementary products) box on the product page offered a 6,080 TL bundle behind a single “Birlikte Sepete Ekle” (add to cart together) button: the 3,995 TL bouquet, a 595 TL balloon, a 1,295 TL scented candle and a 195 TL card reading “Geçmiş Olsun” (get well soon). An agent looking for birthday flowers that pressed this button would exceed the budget by another 2,085 TL and add a card meant for a different occasion.

What does the price cover?

At the boutique hotel, the first price shown was “From 217,80 €”. Only the cart revealed that this was the total for two nights with breakfast. The date box also expected plain text in month/day/year format; the agent filled it in by guessing the format.

The candle task asked for delivery by Thursday evening; we ran it on a Monday evening. The site only said “2 iş gününde kargoda” (ships within 2 business days). That is the time to dispatch, and no delivery date appeared anywhere before the checkout page. For shipping it said “Kargo sonraki adımda hesaplanacaktır.” (shipping will be calculated in the next step), so the final price of the 95 TL item was unknown until an address was entered. There was no gift-wrap option on the product, cart or checkout page.

Filters, search and forms should do what they say

The florist’s “1995-3995” price filter also listed items at 4,295, 4,995, 5,295 and 7,295 TL. An agent that trusts the filter may not notice it has gone over budget.

At the candle shop, a search for “lavanta” (lavender) returned 20 results. Only one of them was a lavender candle, in 15th place; the top results were room fragrance, cologne and a cinnamon candle. An agent that picked the first result would have bought something that was not lavender.

The florist’s delivery district field only accepted a choice from its suggestion list. Typing “Caferağa” returned an empty list; typing “kadikoy” without Turkish characters returned four suggestions.

An “appointment” is not the same thing on every site

The hair salon showed real free slots. The dental clinic’s form had no choice of time, treatment or dentist; it only asked for the branch, the day and a contact preference. Even if submitted, it would have been a call-back request, not an appointment. The restaurant’s page also described itself as a request and stated that it is not a reservation.

Finding the clinic’s form was the longest flow of the whole test. The “Hızlı Randevu” (quick appointment) box could not be clicked, the booking button in the menu was hidden at this screen width, and the button on the dentist’s card only appeared on mouse hover. The agent read the page code to work out that the button opens a pop-up; the flow took 11 actions.

What the page shows versus what the product data says

A product page can also carry a summary that visitors never see: product data. Written for search engines and AI tools, this summary uses a shared vocabulary called schema.org to state the product’s name, price, currency and stock status in a form machines can read. Under Google’s structured data guidelines, this data must be a true representation of the page content.

The coffee roaster’s site is in Turkish and belongs to a Turkish business. To our browser connecting from outside Turkey, the product page showed a price of £17.16. Where the add-to-cart button should have been, it read “Currently, we are not selling in your location”, and we found no way to change the country. The product data on the same page said “1150.00 TRY, InStock”. Search engines and AI were being told “in stock, 1,150 TL”, while an agent arriving at the same address could not buy.

We do not know whether this is a deliberate sales decision or an overlooked setting. An agent that sees both pieces of information has to decide for itself which one to trust.

JSON is the format the product data is written in. The florist’s data had a syntax error in that format; the page looked fine to visitors, but machines could not parse the summary.

Agents make mistakes too, so the final approval should stay with a person

The agent made its own mistakes as well. At the hotel the screen updated late; the agent assumed the room had not been added, pressed the button again and put a second room in the cart. The total rose to €435.60. The agent noticed and removed the extra room; in an unsupervised flow this error could have gone all the way to payment.

At the candle shop, the first “Sepete Ekle” (add to cart) click did not add the item; the second one did. The page had 20 buttons with the same label, and it is unclear which one the first click hit, so we do not count it as the site’s fault. Going over budget at the florist was, as described above, our own decision for the sake of observation.

On all five service sites, the final screen asked for personal details:

In our test, entering these details and pressing the final button always stayed with a person. Since 54% of the failed runs in Visa’s test still ended in a purchase, a screen that shows an order summary and asks for approval before payment is the last place where a wrong order can be stopped.

How can you see agent traffic on your site?

Microsoft Clarity analysed more than 5.2 million agent sessions between 13 August and 9 September 2026. A checkout event was recorded in 25.21% of agent sessions and 14.65% of human sessions. Clarity states plainly that these figures are not completed sales and do not prove that an agent finished a task on its own; some sessions may have been handed over to a person.

Clarity’s “Agent Activity” dashboard, announced on 4 November 2026, shows which agents visit your site, how that traffic changes over time and what agents do there, such as searches, form submissions, appointment bookings and checkout. If Clarity is installed on your site, you can see in your own data where agents move forward and where they stop.

Six things you can check on your site today

1. 01

   Show real free slots

   If you take requests rather than bookings, say so on the first screen; do not present a call-back form as an appointment.

2. 02

   Keep buttons visible on every screen

   Booking and buy buttons should be clickable at every screen width; avoid buttons that only appear on mouse hover.

3. 03

   Review default selections

   Preselected time slots and bundles should not replace the customer’s decision; a bundle should fit the purpose of the purchase.

4. 04

   State price and delivery in text

   Say what the price covers, the delivery date and the shipping cost at the earliest possible step.

5. 05

   Test filters, search and forms

   Try your price filter, site search and address fields with Turkish characters; they should do what they say.

6. 06

   Compare product data with the page

   Price and stock in your product data should match the page; also check what you show visitors from abroad.

What this test does not show

Eight sites, one agent, one attempt per site and one day do not make a sample. They are not enough to say whether Turkish websites in general are ready for agents, and we do not draw that conclusion.

We did not measure from which country ChatGPT’s or any other company’s agent connects to websites. Businesses that work only by phone or WhatsApp were not part of this round. The second-round tasks were prepared by the same agent, which knew which edge cases it was looking for. The findings should therefore be read as a record of the points where these sites pushed the agent into making a decision.

Readiness for agents starts where a good website starts for human customers: real free slots, a price that says what it covers, a clear delivery date and product data that says the same thing as the page.