Study · November 2026
Out of 26 Polish B2B software vendors, an AI agent sent to find a price comes back with an answer from four. The other twenty-two stop it at one of eight barriers. Almost none of those barriers is a decision not to sell to agents.
Every row below is a barrier met in practice, not a hypothesis. A vendor can appear in several rows. This report names no companies, because the point is not to judge individual vendors but to show a mechanism that repeats across the whole sample.
| Barrier | Vendors |
|---|---|
| The /pricing or /cennik address returns 404 | 6 |
| The page shows a number that is not a price | 3 |
| Price only behind a contact form | 2 |
| Pricing page rendered only by JavaScript | 2 |
| robots.txt forbids reading the pricing page | 1 |
| Amounts in an image file, machine-unreadable | 1 |
| Price only after logging in | 1 |
| Server refused automated access (403) | 1 |
Six vendors out of twenty-six publish a price in a machine-readable form. Two of those six forbid automated clients from reading the pricing page in robots.txt. An agent that respects those rules, which means every agent operating properly, comes back with a price from four vendors out of twenty-six.
Two rows work in opposite directions. Twenty-three percent of the sample returns 404 at the address an agent tries first, because that is what a pricing page is called across the rest of the internet. At the same time eleven percent of the sample shows a number that is not a price, most often share capital in the footer. The first group is invisible to an agent. The second is more dangerous, because the agent finds something and reports it to the user as a price.
Beyond the price itself, four technical facts decide whether an agent can work with a site at all. All of them are binary and machine-checkable.
Nearly half the sample did the visible thing. One vendor did the thing that lets an agent answer the question about cost.
| Infrastructure element | Vendors | Share |
|---|---|---|
| llms.txt file | 11 of 26 | 42% |
| Public API documentation | 11 of 26 | 42% |
| Any offer type in structured data | 5 of 26 | 19% |
| Named rules for AI crawlers in robots.txt | 2 of 26 | 8% |
| Price in structured data | 1 of 26 | 4% |
Ten vendors have llms.txt and no price in structured data. Five vendors describe their product with an offer type and four of them left the price field empty. In one case the offer object carries only availability and a link to a demo form, which is exactly what an agent does not need.
Twenty-four vendors out of twenty-six name no AI crawler in robots.txt, which in practice means they allow all of them. Two vendors have such rules and both are permissive. No vendor in the sample deliberately blocked an AI bot. The only AI-bot block in the entire sample is accidental: it comes from a rule covering all automated clients, with a comment pointing to a framework default that reached production. The same rule also blocks Google.
The order follows the measurement, not preference. At the top are the things that block agents most often and cost the least.
| # | What to do | Why it sits here and not lower |
|---|---|---|
| 1 | A pricing page at a predictable address, /pricing and /cennik | Six vendors out of 26 return 404 there. It is the first address an agent tries |
| 2 | Amounts as text on the page, not in an image file | An image is empty to a crawler, a model and a screen reader. The cost is rewriting one page |
| 3 | Check your own robots.txt against the pricing page | One vendor blocks its own pricing page there, another blocks everything with a rule that also covers Google |
| 4 | A pricing page visible without executing JavaScript | Two pricing pages in the sample are empty to any tool that does not render the page |
| 5 | Remove numbers that are not prices from around the pricing page | Three vendors put a share-capital figure within an agent's reach |
| 6 | Structured data of type Offer with the price field filled in | One vendor out of 26 does this. It is the only format an agent takes a price from without guessing |
| 7 | An llms.txt file | Eleven vendors out of 26 already have one. It is last because on its own it does not answer the question about cost |
Positions six and seven are in this order deliberately, and that is the practical conclusion of the whole study. The market did the thing in seventh place first and skipped the thing in sixth. llms.txt is easy to deploy and visible, so it reaches the plan first. A price in structured data requires a conversation about whether the company wants to publish a price at all, so it does not reach the plan.
The first five positions are a few hours of work for anyone with access to the site. None of them requires a decision about the pricing model. The sixth one does.
Missing prices increase monotonically with distance from self-service. The order is the one the contract-size hypothesis predicts: the longer and more intermediated the buying process, the less often a price is published.
| Sales model | N | Amount in text | Image only | No amount |
|---|---|---|---|---|
| A self-service path exists | 4 | 2 / 50% | 1 / 25% | 1 / 25% |
| Sold by the vendor's own sales team | 14 | 3 / 21% | 0 / 0% | 11 / 79% |
| Sold through implementation partners | 8 | 1 / 12% | 0 / 0% | 7 / 88% |
In this entire cohort not one vendor published a price range or a starting-from price. Either the number is specific or it is absent. The intermediate forms, which cover 11% of the wider SMB sample, do not occur here at all.
The self-service group has only four vendors, so its percentage should be read as a direction, not a measurement. It is consistent, however, with the wider SMB sample, where self-service dominates and 6% of vendors lack an amount.
A missing price is not one phenomenon. The categories below emerged from verification; they were not assumed before the data was collected.
| Species | What the buyer sees | Vendors |
|---|---|---|
| No pricing page | Nowhere to look. The only path is a form or a phone call | 9 |
| Open price list | Amounts in text, readable by a human and a machine | 6 |
| A pricing page that explains the absence of a price | The vendor devotes a page to why it will not give a price | 1 |
| A machine-unreadable price list | The amounts exist, but in an image file | 1 |
| Price behind a form disguised as a calculator | The tool looks like a calculator but computes nothing on the page | 1 |
| Price list behind a login | Prices are specific, but only after registration | 1 |
Two species look similar in the statistics and differ fundamentally. A price list behind a login means prices are specific and available without talking to a salesperson; you only need an account. A price list in an image file means the vendor publishes prices openly and in good faith, but to a crawler, a language model and a screen reader that page is empty. The first barrier is deliberate. The second is invisible to whoever put it there.
One accounting software vendor has a section on its home page about fair pricing: it declares transparency, no hidden fees, and invites you to compare its prices with the competition. There is no link to a price list anywhere in the site navigation, and no amount on any page checked. This is not silence about price. It is a contradiction between a marketing promise and the contents of the site.
Another vendor has a page explicitly called a price list, with category and product-line filters, and zero amounts in the text. Every price sits in an image file. To a human looking at the screen the price list is complete. To a search engine, an AI assistant and a screen reader it does not exist. At a moment when more and more questions about software cost go to a language model first, that is a difference with commercial consequences.
One vendor publishes a complete price list with specific amounts and at the same time forbids all automated clients from reading that page in robots.txt. The ban enumerates both currency versions separately, so it is not the result of an accidental pattern. The product has a price, a human will see it, and an agent sent by that same human is not allowed to read it.
One vendor devotes its pricing page to explaining why it will not give a single price, and lists the factors that drive implementation cost. It is the only case in the sample where a missing price is described and justified. The other eighteen vendors without a visible amount simply have no page anyone could look at.
Sampling frame. A list of Polish B2B software vendors built by hand, from industry knowledge and desk research, then verified vendor by vendor against two conditions: whether the company still operates and whether it sells its own software. Software houses and integrators that describe themselves as service providers rather than product vendors were dropped. So were entities that turned out to be dead, acquired or rebranded into a consumer product.
Conflict-of-interest exclusions. Companies connected to the author of the study were removed from the sample.
Segmentation. The dividing variable is the sales model in three categories: a self-service path exists, sold by the vendor's own sales team, sold through implementation partners. The variable is independent of the outcome, because it describes the purchase mechanism rather than whether the company publishes a price.
Two-stage coding. Stage one was an automated screen. Stage two was inspecting every case in a real browser, with JavaScript executed. Every negative result was verified, as was the context of every amount found.
Why the second stage is mandatory. Tools that fetch a page without executing JavaScript do not see client-rendered price lists. In this study that produced fourteen incorrect readings, all in the same direction: the company was showing prices and the measurement claimed there were none. Without the second stage the report would have overstated price hiding by more than ten percentage points.
Data source. The vendors' own websites only. No directories, comparison sites or aggregators.
What was not done. No robots.txt restriction was circumvented. One company in the frame blocks automated access and was marked as missing data, counted in no denominator. No contact forms or quote requests were submitted.
A correction made during analysis. The first version of this result said 86% had no amount. That figure was computed on a denominator skewed toward negatives, because browser verification targeted the cases without a price. After completing the cohort and computing the result by sales model, the numbers are those in the tables above.
Searching for a number-plus-currency pattern on a page whose address contains "pricing" produces false results in both directions in this sample. These are not edge cases: they affect one vendor in five.
| What the tool sees | What it actually is | Times |
|---|---|---|
| A seven-figure amount in zloty | Share capital in the site footer | 2 |
| An amount on a page at the pricing address | A blog article about price lists, after a redirect | 1 |
| Zero amounts on the pricing page | The price list exists, the amounts are in an image file | 1 |
| Zero amounts on the pricing page | A price calculator rendered by JavaScript | 1 |
| Zero amounts | Specific amounts visible once JavaScript runs | 1 |
The first two rows are false positives: the tool would record as a price a number that is not one, and report it to the user. The last three are false negatives. In this study the second kind of error occurred fourteen times before manual verification.
The conclusion for anyone who wants to repeat this study: the count of amounts found on a page is not data. Data is an amount whose context someone checked.
The sample is not random. The frame was built by hand, so the results describe this list of companies, not the whole market. The core cohort is 26 vendors, and the smallest group in a table only four, so one company shifts that percentage by 25 points.
The measurement is a snapshot from a single day. Companies change their sites.
The study does not claim that any of these companies is losing money today because of AI agents. That cannot currently be measured honestly from the outside. It measures readiness, not lost revenue.
Nor does it claim that Polish B2B SaaS hides prices. In the wider SMB sample the opposite is true: 83% publish a specific amount. The problem affects the part of the market that sells expensively and through intermediaries.
Finally, it does not claim that each of the eight barriers is a mistake to be fixed. A company that deliberately does not want its price quoted by an AI assistant is making a commercial decision. The sample suggests, however, that such deliberate decisions are rare: out of nineteen vendors without a visible amount, one stated plainly why it does not publish a price.
It claims two things. First: an agent that respects robots.txt and is sent to find a price at a Polish B2B software vendor comes back with an answer in one case out of seven. Second: the differentiating variable is the sales model and contract size, not the fact that the company is Polish.