Can artificial intelligence see Polish service businesses?
43% of the sites studied publish no structured data — a language model can read such a site, but gets not a single fact about the business to quote: no address, no opening hours, no type of business. Among guesthouses it is 69%, and the same on Hel.
Contrary to intuition, blocking is not the problem. Only 2% of sites shut the door on AI crawlers — and 5 of those 9 blocks are a Cloudflare default, not the owner's decision. The problem is subtler: the site lets the crawler in and gives it nothing concrete to remember.
Why we studied this
More and more people no longer type “hotel Karwia” into Google, but ask ChatGPT or Perplexity “where to stay in Karwia”. The answer is built from what those models' crawlers managed to read earlier on business websites. A business the crawler did not read will not appear in the answer — not because it is worse, but because it is not there.
This is measurable. You do not need to ask a model or guess. It is enough to check what a crawler gets when it visits the site.
How we measured — and how you can check it yourself
For each site we downloaded two files: the home page and /robots.txt. Then we checked seven things, each with a point weight. A site starts at 100 points of visibility and loses the weight of every thing the crawler is missing. The weights add up to exactly 100, so 100 = the crawler reads everything, 0 = it reads nothing. You can repeat each check in a browser in a minute — see the right-hand column.
Scale change. Until 1 September 2026 we reported the result the other way round — as invisibility, where 100 meant “nothing visible”. From 2 September we report visibility: 100 means “everything visible”. It is the same measure and the same measurements, just counted from the other side. If we sent you your site's result before that date, the current equivalent is 100 minus that number.
| What we check | Points deducted | How to check it yourself |
|---|---|---|
| robots.txt blocks AI crawlers (ChatGPT, Claude, Perplexity) | 30 | type yourdomain.com/robots.txt in the browser — look for a “Disallow: /” line under a crawler's name |
| content is added by a script — the AI crawler reads an almost empty page | 20 | Ctrl+U (view source), then Ctrl+F and the business name — if it isn't there, the crawler can't see it either |
| no schema.org structured data — AI has no facts about the business | 20 | Ctrl+U, then Ctrl+F and “application/ld+json” |
| no address and phone number in the page text | 10 | Ctrl+U, then Ctrl+F and your own phone number |
| no meta description | 8 | Ctrl+U, then Ctrl+F and “meta name="description"” |
| no H1 heading | 7 | Ctrl+U, then Ctrl+F and “<h1” |
| no sitemap declared in robots.txt | 5 | yourdomain.com/robots.txt — look for a “Sitemap:” line |
We checked blocks for 13 crawlers, by the names they identify themselves with in robots.txt: GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, anthropic-ai, Claude-Web, PerplexityBot and others. What we deliberately did not do: we did not ask any model whether it knows a given business. A model's answer changes with every query, so it can be neither repeated nor verified.
Overall result
Average visibility 82.8 / 100 (median 86). Sites below the 60-point threshold, meaning more than one thing is missing: 41 (10%).
| What is missing | Sites | Share |
|---|---|---|
| no schema.org structured data | 184 | 43% |
| no sitemap declared in robots.txt | 135 | 32% |
| no H1 heading | 102 | 24% |
| no meta description | 100 | 24% |
| no address and phone number in the page text | 93 | 22% |
| content added by a script (the crawler reads an empty page) | 12 | 3% |
| robots.txt blocks AI crawlers | 9 | 2% |
Who shuts the door on AI crawlers? Most often not the owner
9 of 424 sites have an entry in robots.txt that turns the ChatGPT, Claude or Perplexity crawlers away. We checked each of those files. 5 of 9 are an identical, automatically generated “Cloudflare Managed content” block — the CDN provider's managed robots.txt, with an ai-train=no policy and a list of crawlers with Disallow: /. The other four are manual entries, most often in WordPress.
Since 1 July 2025 Cloudflare blocks AI crawlers by default for new domains, and the managed robots.txt with the ai-train=no signal was switched on automatically on millions of domains. For a content publisher that is protection. For a guesthouse, clinic or law firm that wants to be recommended, it is a closed door the owner usually does not know about — because the hosting or the site's builder set it, with one click.
The fix is just as simple: one click in the Cloudflare panel, or your own robots.txt with an explicit Allow: / for the crawlers you want to appear to. This is what ours looks like.
By industry
| Group | Sites | Avg. visibility | Below threshold | Most common gap |
|---|---|---|---|---|
| Guesthouses | 116 | 75.7 | 18 (16%) | no structured data — 69% |
| Clinics and practices | 90 | 89.5 | 5 (6%) | no H1 heading — 30% |
| Hotels | 69 | 83.2 | 5 (7%) | no structured data — 46% |
| Estate agencies | 64 | 83.1 | 5 (8%) | no structured data — 47% |
| Law firms | 37 | 91.6 | 0 (0%) | no meta description — 27% |
| Beauty salons and spas | 28 | 85.5 | 1 (4%) | no address and phone number — 36% |
By city
| Group | Sites | Avg. visibility | Below threshold | Most common gap |
|---|---|---|---|---|
| Karwia | 38 | 77.5 | 7 (18%) | no schema.org structured data — 55% |
| Władysławowo | 35 | 78.9 | 3 (9%) | no schema.org structured data — 60% |
| Hel | 32 | 71.3 | 9 (28%) | no schema.org structured data — 69% |
| Sopot | 26 | 76.5 | 3 (12%) | no schema.org structured data — 69% |
| Warszawa | 122 | 88.1 | 4 (3%) | no schema.org structured data — 23% |
Groups with fewer than 20 sites were left out — 19 sites in the industry split and 88 in the city split. A percentage of six sites is not a study result, only an anecdote.
What this means for a business owner
- 1. Check your own robots.txt. Type your address followed by
/robots.txt. If you see the name of one of the crawlers there and under itDisallow: /— or a “Cloudflare Managed content” block — the models are not allowed onto your site. It is a fix of a few minutes, and most often nobody ever set it deliberately. - 2. Add structured data. This is the most common gap in the whole study. It is a dozen or so lines describing the business in a format a model understands without guessing: name, address, phone, opening hours, type of business.
- 3. Check that the content is in the page code. If Ctrl+U does not show your business name, a script adds the content in the browser. A person will see it, a crawler not always.
The same seven checks work on a competitor's site. If yours scores 40 points and the one next door 0, it is the one in the answer, not you. Checking both takes two minutes.
None of these things guarantees that a model will start recommending the business. They only guarantee that it will stop failing to see it. Anyone who promises first place in a ChatGPT answer promises something that can be neither set nor verified.
Methodology, limitations, data
- The sample is not random. These are service businesses from six industries, collected with Deep DevOps' commercial activity in mind. The result should not be generalised to the whole Polish economy.
- We oversampled the coast on purpose. 114 properties from Karwia, Władysławowo, Hel and Puck (Puck is below the threshold of 20, so it has no row of its own) were chosen so that the seaside municipalities could be described separately. Because accommodation does worse in this study than clinics or law firms, this raises the overall result — in the first measurement of 307 sites, without that batch, it was 37% instead of 43%. When comparing industries or cities with each other, use the tables below, not the overall average.
- We measured the home page. Subpages may look different.
- The measurement is from the day of the scan (18 August 2026 – 29 August 2026). The site may have changed since — and if the owner has fixed it, they are right and we are not.
- Business names appear nowhere — neither here nor in the downloadable data. Every business studied can get its individual result on request.
- Aggregate data to download: dane.json (CC BY 4.0 licence — cite with a link to this page). Breakdowns for five cities are public — links in the “By city” table above; I will prepare another breakdown at an editor's request.
External sources
Want your site's result?
If your business was in the study, the result already exists — one sheet, describing how to check each thing yourself. If it was not, enter the address in the scanner: the same method, the result straight on screen. No offer and no conditions.
Check your site →Author: Piotr Schumann, Deep DevOps — I build websites and run my own guesthouse in Pomerania, so the 69% among guesthouses is not a statistic to me, but neighbours.