Are Portuguese online stores ready for AI agents?

400 online stores of Portuguese brands and retailers, each on its own domain, gathered from 197 public pages on 41 sites: trade and producer association lists, award lists and press articles that name Portuguese brands by sector. Each store has its source on file.

We left out marketplaces, foreign retailers with a Portuguese site, stores whose only source was sponsored or more than ten years old, repeated brands (one store per brand), and the stores of current and past SO-MA clients, so we are not grading our own work.

The sample is not random: these are brands visible enough to appear on those lists. The results describe these 379 stores, not all Portuguese e-commerce.

On 19 October 2026, from Lisbon, using the code of SO-MA's public checker (the one on this page), run from a script rather than the site's server. The reader identifies itself as "SO-MA-research (+https://so-ma.studio/indice)".

One request at a time per store, with a one-second pause between requests; home page redirect hops and the two robots.txt files (with and without www) follow without a pause, as in a browser. Before anything else we read each server’s robots.txt and request no path it disallows. Public pages only: home page, robots.txt, sitemap.xml, llms.txt and one product page linked from the home page. No logins, no forms, no stored cookies.

We keep only the check results and counts (words, images, fields): no page content and no data about the stores’ customers. The measurements behind these figures made 2472 requests, robots.txt included.

We published the Index on 16 October 2026 with 204 ready stores, 17 below 60 and 79 scoring 100. We then found two errors in our reader. The errors were in our code, not in the stores.

The first read any response larger than 2 MB as empty: 12 home pages, 1 product page and 2 sitemaps. On 19 October 2026 we fixed it and measured those 15 stores again, and the figures became 212, 5 and 83. The reader now reads the first 2 MB of each response, like SO-MA's public checker.

The second counted the time each request spent waiting in our queue as the store's time: we make one request at a time per store, and the time limit started when the request joined the queue. On a slow store, a sitemap or product page that existed could run out of time before it was requested. Each request's time limit (8 seconds, 10 in total with redirects) now starts when it leaves the queue, as in the public checker, where requests run in parallel.

With both errors fixed, we measured all 400 stores in the sample again on 19 October 2026. Every figure in the Index comes from that measurement: 215 ready stores, 8 below 60 and 83 scoring 100.

Ten checks with fixed weights, the same as the checker. Ready: 85 or more. Partly ready: 60 to 84. Not ready: below 60.

We pick the 20 with the highest score. When the cut falls in a tie, ties are broken, in this order, by more basic accessibility checks passed (0 to 3), a higher share of images with alternative text, more words readable without JavaScript and the alphabetical order of the domain. Word count is not one of the ten checks, so the 20 appear in alphabetical order and unranked. No other store is named.

The product check looks on the home page for a link with /products/, /product/ or /produto/ in the address, and only counts the page if it is still a product page after redirects (/produtos/ is left out: in Portuguese stores it is almost always a listing). In 196 of 379 stores we found no product page that answered (more common outside Shopify, where product pages use other address formats); those stores get 5 of 15 points on this check, as in the checker. A product page without product data scores 0, so finding one can lower the score compared with finding none. That is why the "product without data" gap is counted only among stores where a product page was found.

We read the first 2 MB of each response, like the checker on this page: on a very large page, anything after that (for example product data at the end of the code) does not count. A store that takes more than 8 seconds to answer a request (10 with redirects) goes without that response, also as in the checker.

The 21 stores we could not measure, especially the 15 that refused even an identified reader, are probably among the most closed to agents. They are outside every figure, which makes the picture look better.

We only read robots.txt. Blocks at the firewall or CDN, such as the AI bot blocking Cloudflare offers, are not visible in this measurement.

Speed was measured from Lisbon on an office connection and varies through the day. It is a snapshot of the day each store was measured: stores change.

The study measures whether a store can be read and bought from by agents. It does not measure whether ChatGPT, Gemini or Perplexity actually recommend each brand.

21 of the 400 stores on the list were not measured and are not in any figure: 15 stores refused our reader; 4 stores had a robots.txt that returned an error or did not answer, and the rule in that case (RFC 9309) is not to read; 1 store did not answer within 8 seconds; 1 store redirected too many times. Refusing the reader means a 4xx response (403 or 429), almost always bot protection.

Stores that did not answer, or whose robots.txt did not answer, got a second attempt the same day; stores that refused the reader (403 or 429) did not.

Since 19 October 2026, a home page read with no text and no title, or a bot check page served instead of the store, never gets a score: the store counts as not measured. The rule also applies to earlier measurements. No store in this edition is in that case.

SO-MA sells the Agent-Ready Audit and builds Shopify stores. That is why we publish the whole method, use a checker anyone can run, and leave our clients’ stores out.