Measurement · Export
What do exporter websites tell machines? 2026 measurement.
Most corporate websites of Turkey’s largest exporters in Bursa, Izmir, Gaziantep and Denizli do not declare their multiple languages to search engines, and some turn AI crawlers away at the server while staying open to an ordinary browser.
Istanbul — An exporter’s website addresses two readers: the foreign buyer who opens the site, and the machine that brings that buyer to the site — a search engine, an answer engine, an AI assistant. What the second reader sees usually goes unmeasured. This article looks through that reader’s eyes by measuring the websites of Turkey’s largest exporters in four industrial cities with the same method.
How was the sample built?
The starting point is the Turkish Exporters Assembly’s (TİM) Top 1000 Exporters list for 2025. Companies whose province column reads Bursa, Izmir, Gaziantep or Denizli, that disclose their name and that are not foreign trade companies were ordered by overall rank, and the first 30 companies from each province were taken; in Denizli, 16 companies met these conditions. Of the 106 companies in total, 103 were matched to a corporate website by finding the company’s legal name or brand on its home page.
The 18 matches that led to an international group’s global website were not measured: those sites represent the group’s headquarters, not the Turkish company’s own surface. Of the remaining 85 sites, 81 responded at the time of measurement. That number is the denominator of every ratio below.
What was measured, and what was not?
The method is the same as that of this site’s Measurement Observatory: for each site the home page, robots.txt and llms.txt were read; one probe was sent under the identity of each AI crawler, plus one control request. The probe requests did not hide their identity — the request header carried the address of the measurement. Page content was not stored; only countable facts were recorded. Speed and accessibility were not measured in this round: both require a real browser, and a single run does not represent their distribution.
Multilingual, but does the machine know it?
Of the 81 sites that responded, 31 declare more than one language on their home page with hreflang tags. This is a lower bound: a site that offers a language switcher without hreflang looks monolingual by this measure. But the same holds for the search engine — without hreflang, whether the German page is shown to someone searching in German is left to the search engine’s guess. The language of exporting is often the language of the sale; a machine that does not know that language means the right page does not appear in the right market. How this gap is closed is in the multilingual site architecture guide.
The cities differ: in Bursa, 10 of the 22 responding sites use hreflang, while in Denizli the figure is 3 of 12 sites. The samples are small; the difference is an indicator of a trend, not a definitive ranking.
A rules file is a request; the server’s response is a decision.
Are AI crawlers left at the door?
Only 2 sites explicitly forbid AI crawlers in their robots.txt file. But a rules file is a request; the server’s response is a decision: 10 sites opened the page to an ordinary browser yet refused access to a probe requesting the same address under an AI crawler’s identity. In Gaziantep, that is 5 of the 24 responding sites.
This is not a verdict. The block may be a deliberate security decision, or the default setting of a protection layer in front of the server. But the outcome is the same: that site cannot be a source in the answer an AI assistant gives to the question “who makes this product?”. The site owner needs to choose this knowingly — whether what robots.txt says matches what the server does can be seen with the machine reading tool.
Security header and structured data
Almost all the sites open over a secure connection; but only 21 send the HSTS header, which makes the secure connection permanently mandatory for the browser. The number of home pages carrying structured data (JSON-LD) that describes the product, the company and the contact details in a form the machine understands is 21. 7 sites have an llms.txt file — the effect of this file is disputed and our own measurement does not count it as a lever; the number only shows how widespread it is.
Limits
The measurement was taken on 23 September 2026, at a single moment and from a single location; the decisions of protection layers in front of servers can vary by country and time of day. Only the home page was read. A company’s base is the province of its registered office in the TİM data, which may not be where production takes place. Company names and domains are not published; every number is read from this article’s measurement log and is updated with the same method when the measurement is repeated.
What is expected from the websites of these four cities’ industries is covered separately on the city pages; the general structure is on the B2B website page.