Producer websites are open to everyone. One in six offers a spec sheet.

We examined, from the outside, the 249 websites listed by producers and exporters in ICE's directory for foreign buyers. The sites are open to everyone, search engines and AI assistants included. The spec sheet a buyer wants to see is offered by just one food producer in six.

The study is by Itria AI (itria.io) and is independent of ICE: the directory is simply the public source the website addresses came from. The measurements were taken on 26 September 2026. The method, checks, corrections and the data for every site, without company names, are on this page and in the file you can download at the bottom, under a CC BY 4.0 licence. If you export, the practical pages start with the hands-on export guide for small food producers.

Where export deals really happen

In person, at trade fairs such as Anuga and SIAL, and through samples. The CBI, the Dutch government centre that helps exporters from middle- and low-income countries sell into Europe, says so in its guide on how to find European buyers of processed fruit and vegetables: most lasting partnerships still come from personal introductions, sample testing and factory visits.

“Processed food is a face-to-face business, but a website gives you global visibility” (CBI, guide to European buyers of processed fruit and vegetables, updated 19 September 2023)

Before the meeting comes the search. European buyers look for new suppliers on Google, typing the product name and words like “producer” or “supplier”, sometimes with the country of origin, and according to the CBI appearing on the first page for those searches is “very important”. When it's the producer who makes the first approach, the CBI recommends a phone call followed by an email.

On the website itself, the CBI says, product documentation should be available to buyers: full specifications, including size, weight, packaging, shelf life and nutritional values, which is the information a buyer wants to see and that helps them decide. As an example it points to a producer that publishes a PDF sheet for every product: that's the spec sheet this study went looking for.

The study measures what can be seen from the outside: what a buyer, a search engine and an AI assistant find on the website. How many export enquiries actually come in through the website is known only to the company receiving them. That figure is missing from the public sources we read, and we don't estimate it here.

What we measured: the buyer, the search engine, the AI assistant

A company website now has three readers: the foreign buyer who opens it, the search engine that indexes it, and the AI assistant that reads it on someone else's behalf. We measured five things the first one needs, and six that decide whether the other two can find the company and understand what it does.

Measure	What counts as a yes	Why it matters
English version	English hreflang, a link to /en/, an EN language switcher, or a home page already in English.	It's the foreign buyer's language.
Contact channel	A form with an email or message field, a mailto link, or a written address.	Without one, the buyer doesn't write.
Email readable by an AI assistant	The address in plain text in the served HTML, not obfuscated.	Assistants that read pages on the fly don't run JavaScript (Vercel).
Spec sheet	Downloadable: a product sheet as a PDF, image or web page, linked somewhere between the home page and a product page, and opened. On the page: a section headed technical sheet or technical data, with the fields below. A general catalogue doesn't count.	The CBI recommends keeping it on the website, available to buyers.
Terms for foreign buyers	An export or wholesale price list, a minimum order for resellers, an Incoterm, an ex-works price. Checked case by case.	Without them, every offer starts with an email asking a question.
robots.txt open to Google and Bing	Googlebot and Bingbot can read the home page.	Block them and you drop out of search results.
robots.txt open to AI search	OAI-SearchBot, PerplexityBot, Claude-SearchBot and Claude-User can read the home page.	OpenAI: sites that block OAI-SearchBot don't appear in ChatGPT's search answers.
Indexable home page	No noindex in the meta robots tag or the X-Robots-Tag header.	Google drops a page with noindex from its results altogether.
A title that says what the company does	The title contains something besides the company name and words like Home or Welcome.	Google asks for descriptive titles and is explicit about it: no “Home”.
Meta description	Present and not empty.	It's one of the sources for the snippet shown under the result.
Company structured data	The home page declares schema.org Organization or LocalBusiness.	It helps Google understand who the company is and tell it apart in results.

The reason behind each row comes from the documentation of the companies that set the rules: the pages from OpenAI, Anthropic and Perplexity on their crawlers, and Google's guides on noindex, titles and organisation data. We also measured sitemaps and product structured data, reported separately because none of these sources asks a small site for them.

Why we didn't measure llms.txt

Because Google says it doesn't use it. In its guide to generative search features, updated on 10 July 2026, the entry on llms.txt files says no special files are needed to appear in search, not even in AI features, and that creating them “will neither harm nor help” visibility, because Google Search ignores them.

“You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn’t use them.” (Google Search Central, guide to generative search features)

itria.io has an llms.txt file too, meant for other services that do read it. We mention it because the study's rule applies to us as well: we measure what search engines and assistants say they use. The same guide adds that structured data isn't a requirement for generative search, which is why we present it for what Google says it does: tell it who the company is.

The method: the ICE directory, the served HTML, the product page in a browser

The frame is ICE's public directory “Find your Italian partner”, designed for foreign companies and agents looking for Italian products or partners. The listings were collected between 19 and 31 August 2026 through searches for Puglia-related terms, such as Puglia, Apulia, Bari and Salento, and for products, such as preserves and cheese.

The directory doesn't record the region, and the search matches the text of the listings, so the companies found come from all over Italy. The listings gave 249 different websites, and we read them all, with no sampling. The sample represents the companies presenting themselves to foreign buyers in that directory; Italian food exports as a whole are beyond its reach.

For each site we read robots.txt, the home page, and the contact and product pages if linked from the home page, on 26 September 2026 between 13:34 and 14:39. We read the HTML the server sends, which is what an AI assistant that doesn't run JavaScript sees, using a tool that declares its identity, respects robots.txt and waits at least 2.5 seconds between requests.

Each measure has three possible outcomes: yes, no, and not measurable. A site that refuses the request or returns an error is counted separately and left out of the percentages. Pages, headers and robots.txt files were saved, so every measure can be rechecked against the code as it was that day.

For spec sheets we went one click further. On the evening of 26 September we opened in a browser, one site at a time, the first product page of every food website, plus up to two trade or download pages linked from the home page or the product page. Every candidate document was opened and read: general catalogues, brochures and price lists don't count.

The directory: 26 of 249 addresses lead nowhere useful

10.4% of the addresses the ICE directory shows to foreign buyers lead to something other than a company website. 14 domains no longer exist, 2 home pages return an error, and 10 pages are something else entirely: a default server page, two sites under maintenance, a WordPress install never set up, a parked domain, and a domain now taken over by an online casino.

Outcome for the 249 addresses	Addresses	Share
Read	195: 121 food websites across 122 addresses, 71 from other sectors, 1 unclassifiable, 1 shop under maintenance	78.3%
Home page almost empty in the HTML	14: 5 built in JavaScript, 7 that aren't really websites, 2 minimal but working	5.6%
Domain no longer exists	14	5.6%
Home page returns an error	2	0.8%
robots.txt unreachable or returning an error	14, counted as not measurable	5.6%
Access blocked or refused	6	2.4%
Other, not measurable	4: two sites that answer a browser but not the tool, a domain back online, an address mistyped in the listing	1.6%

The 10 pages that aren't websites, one by one: a server's default page, a server placeholder, a site under maintenance, an online shop under maintenance behind a password page, one being updated, a WordPress install showing the sample page with noindex switched on, a parked domain, a redirect to a domain that no longer exists, a domain now hosting an online casino, and a home page made up entirely of PHP error messages. All the addresses without a website were reopened in a browser on 26 September.

Anyone curating a directory of companies for foreign buyers could run this check in an hour: open the addresses. For the company, the website listed in a directory is a promise made to a buyer it hasn't met yet.

The table: 121 food producer websites

Of the 195 sites read, 121 belong to food and drink producers or distributors. We classified every one of them individually, reading the title, description and home page text: the directory also returns companies from other sectors, from cosmetics to machinery, because the search is text-based. Percentages are calculated on the 121 websites: two addresses in the directory lead to the same winery.

What the buyer finds	Websites with a yes	Out of 121
Contact channel in the code	120	99.2%
English version	93	76.9%
Email readable by an AI assistant	107 (7 obfuscated, 7 missing from the code)	88.4%
Downloadable spec sheet, between the home page and a product page	19	15.7%
Spec sheet on the page, with the fields	10 (7 of them without a downloadable sheet)	8.3%
Terms for foreign or wholesale buyers	1 (2 counting a borderline case)	0.8%

A foreign buyer can almost always find a way to get in touch. They can download a spec sheet on one site in six, and find the terms for ordering on one in 121. Every missing piece of information turns into an email asking for it.

And response time matters: in the 2007 InsideSales.com and MIT study, companies that called back within five minutes were up to a hundred times more likely to reach the lead, as we explain in the five minutes that decide an enquiry.

The cost of the questions that arrive by email adds up to hours nobody bills for: the bill for replying by hand runs to six lines. Two documents are all it takes to join the producers who offer them: the food product spec sheet template, with its field-by-field English version, and the export price list with EXW and DDP.

Where the spec sheets are: one in two wineries, one in sixteen other producers

Downloadable spec sheets are concentrated in wine: 13 of the 25 wineries offer one (52.0%), against 6 of the other 96 food producers (6.3%). The sheet almost always sits on the individual product page, one click from the home page: 17 cases out of 19. In the other 2, it's on the page listing the wines.

Spec sheet	Wineries, 25	Other food producers, 96
Downloadable	13 (52.0%)	6 (6.3%)
On the page, with the fields	5 (20.0%)	5 (5.2%)
At least one of the two	16 (64.0%)	10 (10.4%)

We split wineries from other food producers after the read, once we'd seen where the sheets were, using a rule written down before counting: a winery is a site whose main product is wine, and companies making both oil and wine stay with the others. The downloadable file includes the column, so the split can be rechecked.

Outside wine, the 6 downloadable sheets belong to three olive oil producers, a truffle producer, a confectioner and a dairy. Of the 19 documents, almost all are PDFs: in 3 cases the sheet is an image or a scan, and in 1 it's a wine's electronic label, a web page linked as “Technical sheet”.

Producers who have a sheet but don't show it

On 3 websites the spec sheet exists but can't be downloaded: a page of sheets behind a password, a download area that requires a login, and a “Technical sheet” link that opens a form to have it sent by email. We didn't count them as a yes, and no form was filled in.

On 4 websites the price list, prices or product list only appear after filling in a form or registering. On 11 the only product document is a general catalogue or brochure, and on 28 there is no individual product page at all: the products are all on one page or in a gallery, or not listed. Of the 27 trade or download pages read, only one offers downloadable sheets.

Search engines and AI assistants: robots.txt is open, the page says little

All 121 food websites let in Googlebot, Bingbot and the four crawlers ChatGPT, Claude and Perplexity use to build their search answers. None blocks even the training crawlers, and none uses Cloudflare's managed robots.txt, which according to its documentation only shuts out training. The door is open: what's missing is inside the page.

What search engines and assistants find	Websites	Out of 121
robots.txt open to Google and Bing	121	100%
robots.txt open to AI search crawlers	121	100%
Indexable home page, no noindex	121	100%
Page served over HTTPS	120	99.2%
Title made up only of the company name	22	18.2%
No meta description	43	35.5%
No structured data saying who the company is	53	43.8%
Product structured data on the pages read	9	7.4%
Sitemap declared or present	113	93.4%

A title like “Home - Company name” only tells the search engine and the buyer what the company is called, and Google explicitly asks sites to avoid “Home”. A home page without Organization data leaves Google to work out who the company is and to tell it apart from others with the same name. These are an afternoon's fixes, and they help all three of the site's readers.

With assistants, the game is measured in citations, because clicks rarely follow: the Pew Research Center found that people who see an AI summary on Google click a result in 8% of visits, against 15% for people who don't see one. How to work on both fronts is covered in getting found on Google and ChatGPT.

The email AI assistants can't see: 11.6%

14 of the 121 food websites have no email address an AI assistant can read: 7 obfuscate it and 7 don't include it in the code. With Cloudflare's obfuscation, for example, the address in the HTML becomes “[email protected]” and only turns back into plain text when a script runs in the browser, as Cloudflare's documentation explains. The itria.io case, and the check to run on your own site, are in our lab note on the email address AI assistants can't see.

AI assistant crawlers read the HTML exactly as it arrives: Vercel measured this across roughly 1.3 billion requests in a month, and none of the major AI crawlers runs the page's JavaScript. Ask an assistant for that company's contact details and it finds the website, but not the address.

What we corrected before publishing

A first read, on 25 September, found 2 spec sheets and 5 sets of terms for foreign buyers out of 128 sites. Checking the positive cases one by one, we found two rules that were too broad: one counted the word “Wholesale” in a menu and a retail basket's minimum order, and the other classified a vending machine manufacturer as a food business. We reread every website with the corrected rules.

Measure	First read, 25 September	Second read, 26 September
Food websites	128, classified by keyword	123, classified one by one
Downloadable spec sheet	2	1: the other belonged to a vending machine manufacturer
Terms for foreign buyers	5	1: the others were the word Wholesale in a menu and in a presentation, a spam text, and a retail basket
Home page almost empty	14, all put down to JavaScript	14: 5 JavaScript, 9 not really websites

Then came a third read, on the evening of 26 September. The second had looked at three pages per site and counted 1 spec sheet out of 123. Once we also opened a product page in the browser, the downloadable sheets came to 19, and rereading the websites turned up two errors in the count of food websites: all percentages are now calculated on 121.

Measure	Second read, 26 September afternoon	Third read, 26 September evening
Food websites	123	121: two addresses lead to the same winery, and an online shop under maintenance moved to the addresses without a website
Downloadable spec sheet	1, on three pages per site	19, including a product page. There were already 3 within those three pages: the rule didn't recognise “factsheet” or a sheet published as a web page
Addresses without a working website	25 out of 249	26 out of 249

The checks, and how far to trust them

Every “yes” for spec sheets and terms for foreign buyers was read individually in the saved code, and every measured site was classified individually. This case-by-case reading was done by an Itria AI agent, on the code saved on the day of the measurement, using rules written in advance. The positive cases and the addresses without a website were then reopened in a browser: this confirmed the positive cases and removed 7 addresses from the directory count, including a page that redirects to the brand's website and two sites that do respond to a browser. The borderline case is an FAQ for importers that explains what the minimum order depends on without stating it: we counted it as a no, and counting it would take the figure from 1 to 2 out of 121.

In the third read every sheet was opened and read: PDFs with a program that extracts their text, scans and images looked at one by one, and on-page sheets read in the code of the page opened in the browser, heading and fields. On 1 site the fields of the on-page sheet weren't in the code, so we counted it as a no.

robots.txt and noindex were checked on every file, and the tool correctly detects a block on a test robots.txt. Titles: every case reread individually. Descriptions, structured data and sitemaps: 20 sites drawn at random with a fixed seed and reread in the code, with the same outcome in 20 cases out of 20. With 20 cases, the true error rate for these three measures could be as high as about 15%, by the rule of three.

One limit of scope remains. For spec sheets we opened one product per site, the first listed on the product page: a site that only puts sheets on some products may come out without one, and a sheet behind restricted access isn't counted. The other measures are based on the three pages read on the afternoon of 26 September.

Downloading the data, and how to cite it

The file has one row for each of the 249 addresses, with date, outcome, sector and every measure, including those from the third read of the product page, with no company names and in shuffled order. The data is licensed under CC BY 4.0: you're free to reuse it, crediting Itria AI (itria.io) with a link to this page.

File	Contents	Link
Study dataCSV, 249 rows, CC BY 4.0	Date, outcome, sector, the measures for buyers and for search engines and AI assistants, the pages read, and for food websites the product page, the downloadable, on-page or hidden sheet, and whether it's a winery.	studio-siti-produttori-alimentari-2026.csv

Questions and answers