Fourteen reviews wrote down that they could not find an Enterprise price. Weeks later, not one of the fourteen publishes it

Every review we publish ends with a list of what the vendor would not tell us. We went back and asked all of them again, on the vendor’s own pages, on 28 August 2026. Of 98 questions whose vendor we could read, 6 are answered today. The single most common question in the whole list is what an Enterprise tier costs, 14 reviews wrote it down, and not one of the 14 vendors publishes that number now.

Disclosure: we sell AI visibility measurement, so we compete with most of the vendors in this study. Here is the whole method: the population is the 224 bullets in the 48 “what we could not verify” sections of our English corpus, the rule for which of them a page could answer was written down before the first request, the prediction and its retirement threshold were committed before anything was fetched, and every vendor was read on the URL our catalogue already records plus its apex and its /pricing on the same day. Then we widened the read, on a discovery rule fixed before that arm’s first request, and both arms are in our measurement register. Anyone can repeat it against us.

The short version

1. 6 of 98 answerable questions are answered today, 6.1 per cent. 92 are still open.
2. 14 reviews asked what an Enterprise tier costs. 0 are answered. That is not a low rate, it is a complete absence in the most repeated question of the set.
3. We predicted 20 per cent or more and wrote the retirement threshold at 8 per cent before fetching anything. The result crossed it, so this audit does not become a routine, and that was decided in advance.
4. 5 of the 6 answers were on the page our own catalogue already pointed at, which points at the vendor having started to publish rather than at us having missed a page. That is far too small a sample to lean on, and we say so where the figure is.
5. The largest class is not price at all. 30 of the 103 bullets ask about some other published limit, ahead of 25 asking about another price and 14 asking about Enterprise.
6. 5 bullets could not be adjudicated because the vendor’s page was behind a bot challenge or served its figures only after running JavaScript. They are their own bucket and never enter the rate.
7. Widening the read did not find them either. Following the links each vendor’s own page carries closed 1 of the 92 open questions and 0 of the 14 Enterprise ones. It is not that we did not look.
8. The case that started this study is the exception. One vendor had quietly begun publishing, we noticed, and the instinct to generalise from it is exactly what a single dramatic cell should not buy.

What we asked, and what counts as an answer

A review here ends the same way every time. After the price table and the arithmetic there is a section listing the things the vendor’s own pages did not say: what the top tier costs, whether all the engines named in the marketing are in the plan you are looking at, how many prompts the free tier really runs, whether the discount on the annual toggle applies to the tier you want.

Those sections are honest and they are also a standing commission. If a bullet says “the vendor does not publish X”, that sentence has a price: one request. So the question is what happens when you pay it, weeks or months after the review was written.

The population had to be defined before anything was fetched, because the tempting move is to pick the bullets that turned out interesting. The rule is ours and it is narrow: a bullet is page readable if its heading names a price, a plan, a limit, a trial, an engine, a model, a prompt quota, a seat, a currency, a billing term or a discount, and it does not say that nobody in the category publishes it, that it would need a controlled test, or that you have to ask in the demo. On the 224 bullets of the corpus that gives 103, across 40 reviews.

The rule being ours has a direction, and the direction matters. A bullet we misfiled reads here as unanswerable, so it sits outside the denominator and cannot be answered. The 6.1 per cent is a floor.

Each of the 103 was then put to the vendor again on three pages: the URL our catalogue records for that vendor, with the date it was read, its apex, and its /pricing. Four verdicts, all four written down in advance: answered on the URL the catalogue already pointed at, answered somewhere else on the same vendor’s site, still open, or could not read.

The class that never answers

Sorting the 103 by what the review was actually asking for produces six classes, and the interesting one is not the biggest.

The biggest is a catch-all: 30 bullets asking about some other published limit, the kind of thing that is only knowable from a plan comparison table that the vendor did not build. Another 25 ask about a price that is not the Enterprise one, a currency, a billing cycle, an add-on. 13 ask which engines or models a tier actually includes. 12 ask about prompt quotas or refresh cadence. 9 ask about the terms of a free trial or a free tier.

And then 14 ask, in fourteen separate reviews written on fourteen separate days, what the Enterprise tier costs.

All 14 are still unanswered. Not one of those vendors has put a number, a range, a floor or a starting point on the page. Every one of them still routes that question into a form.

This is worth separating from the general result, because the general result is a rate and this one is not. A 6.1 per cent answer rate across the whole set is a slow category. Zero of fourteen, in the question a buyer asks first, is a category norm. Nobody is defecting from it.

The six that did answer, and why six is not a finding

Six bullets flipped. One vendor’s monthly price appeared where before there had been only an annual figure. One vendor’s engine list turned out to be enumerated per tier after all. One published its agency partner pricing. One confirmed a saving percentage that had previously only been a badge. Two more resolved engine and tier questions of the same shape.

Five of those six were on the page our catalogue already recorded, and one was somewhere else on the same site.

The design predicted the opposite, and predicted it with a number. The argument was that a catalogue entry points at one URL, so if an answer exists today it is more likely to be on a page we never fetched than on the page we did. We wrote 55 per cent or more on another page as the prediction and 30 per cent or less as the point where the reading inverts.

It came out at one in six, so the prediction is refuted in the direction it declared. That would normally be the headline of the section. It is not, and the reason is that six is a sample you cannot say anything with. A single row moves it by a sixth. So the honest reading is that the split points weakly at vendors having changed rather than at our reach, and the number is published with that sentence attached rather than dressed up as a second finding.

What we could not read, and why it has its own column

Five bullets never got a verdict. Two belong to a vendor whose site answers automated reads with a bot challenge, and that vendor has refused several separate automated reads in this repository, which makes the refusal a property of its edge and not of any one of our scripts. Three belong to pages whose pricing exists only after JavaScript runs, so a plain read comes back with the layout and none of the figures.

A fact about our fetch is not a fact about the vendor. Folding those five into “still open” would have moved the rate by half a point and would have quietly asserted something we did not observe, which is that those vendors do not publish. Two of them very well might. So the bucket is separate, it is printed in the register, and it never touches the denominator.

The prediction we lost, and what losing it decided

This one was frozen and committed before the first request, and the interesting part is what the loss buys.

The prediction was that 20 per cent or more of the answerable bullets would be answered today. The reasoning was not idle: this category reprices fast. Our own price-freshness pass has caught one vendor publishing four different prices in twenty days and another moving from publishing nothing to publishing a full ladder. If prices move that quickly, the reasonable expectation is that the questions around them get answered too.

They do not. Repricing and publishing are different behaviours, and this category does the first far more than the second. A vendor will change what it charges three times in a month and still not tell you what the top tier costs.

The threshold below which the audit is not worth repeating was written at 8 per cent. The result is 6.1. So this study does not become a recurring pass, and that decision was made before the number existed rather than after it disappointed. That is the whole point of writing the retirement threshold down: an audit whose yield you assess after seeing the yield is an audit that never gets retired.

Two different fourteens, which is easy to misread

Our verified catalogue currently holds 14 active tools that publish no entry price at all. This study found 14 bullets asking what an Enterprise tier costs. Those are not the same fourteen and they are not the same question.

The first is a count of vendors with no public starting price anywhere. The second is a count of review bullets, written on different days about different vendors, most of which do publish a starting price and simply stop before the top of the ladder. A vendor can be perfectly transparent about its cheapest tier and completely silent about what happens above it, and most of these are.

We are naming the collision because a reader moving between our catalogue and this study will hit both numbers within a minute, and two identical figures next to each other are exactly the kind of thing that gets summarised into one.

We widened the read, and it changed one answer

The obvious complaint about the result above is that we did not look hard enough. Three pages is three pages, and a vendor could be publishing any of this one click further in.

So we went and looked, with the discovery rule written down before the first request of that arm so the result could not pick it: from the page our catalogue records, follow the links on the same domain whose path or text carries one of nine words, pricing, plans, docs, documentation, faq, help, support, enterprise or contact, ordered by that list, at most four not already fetched.

96 requests across 37 vendors, 86 readable pages, and 1 of the 92 open questions closed. That is 1.1 per cent. The prediction, frozen with the rest, was 20 per cent or more, with 7 per cent or less written down as the point at which widening the read finds nothing worth the trouble. It came out at 1.1.

None of the 14 Enterprise prices turned up. The prediction there was 0 or 1, with 4 or more forcing the headline of this piece to be narrowed. Zero.

The one that flipped is worth naming because it shows what a real answer looks like. A review recorded that a vendor lists “Custom Scheduling” as a feature and publishes no default frequency, which matters because frequency is what turns a prompt allowance into an answer count. That vendor’s FAQ, which is not its pricing page, says the report updates monthly by default and offers Update Now, Daily, Weekly or No Update. The answer existed, on a page one click away, and had for some time.

Two things about the second arm’s own limits, and the first one is a defect of our design rather than of the category. The instrument control asked how many vendors the rule found and read at least two new pages for, and the design wrote 70 per cent or more as the healthy case and below 50 per cent as the point where discovery, not publication, is the limit. It came out at 62.2 per cent, between the two, and the design did not name that band. So we publish the number and say the middle was ours to name and we did not. Four vendors had no qualifying link at all.

And four extra paths is still a ceiling. A zero here is a floor on what a vendor publishes and not proof they publish nowhere, because a page the recorded page does not link with one of those nine words is invisible to this arm by construction.

What the arm does buy, and it is the reason to run it, is that the headline stops depending on our reach. “Fourteen vendors do not publish an Enterprise price” and “we only read three pages” are compatible sentences. After 96 more requests they are not the same sentence any more.

What this does not say

It does not say these vendors are hiding anything improper. Not publishing enterprise pricing is a normal commercial choice, it is common far outside this category, and it has a defensible logic when deals are scoped rather than listed.

It does not say the questions are unanswerable. It says they are unanswerable from the page, on the day we read it. A vendor may answer any of them in a demo, over email or in a document behind a form. What this counts is what a buyer can find out without identifying themselves first, which is a narrower thing and the thing a buyer actually does at the start. “On a page we did not fetch” used to belong in that list and no longer does, because the section below went and fetched them.

It is one reading, on one day, of up to three pages per vendor in the first arm and up to four more in the second. It is not a time series. And the population is one we selected: these are the questions that struck us as unanswered while writing reviews, not the questions a buyer would necessarily ask.

What a buyer should take from this

The practical version is short. If a tier’s price is not on the page today, waiting will not put it there. 6.1 per cent of the questions our own reviews recorded got answered by the passage of time, and the one you most want answered got answered zero times out of fourteen. So the cost of that question is a sales conversation, and it is worth knowing that before you start comparing, not after you have shortlisted three vendors on published prices and discovered that the tier you actually need is priced in none of them.

The corollary for anyone selling in this category is the cheaper half of the same fact. Publishing a number where fourteen of your competitors publish a form is a differentiator that costs one page edit. Our engine-naming study found 11 of 77 vendors not naming a single AI engine on the page where they take money, and the same instinct is at work: the page is written for the deal, not for the person deciding whether to have one. If you want to know how your own pages read to a buyer and to an answer engine at the same time, that is what our product measures.

Common questions about what vendors leave unpublished

How many AI visibility vendors publish their Enterprise pricing?

In this study, none of the fourteen that were asked. Fourteen of our reviews recorded that they could not find what the vendor’s Enterprise or top tier costs. Re-reading all fourteen vendors on their own pages on 28 August 2026, on the URL our catalogue records plus the apex and /pricing, 0 of the 14 publish that figure. Every one of them routes the question to a contact form. The full row list is in our register.

Do software vendors eventually publish what they leave out?

Very rarely, on this evidence. Of 98 questions our reviews recorded as unanswered by the vendor’s pages, and which a page could in principle answer, 6 are answered today. That is 6.1 per cent, and it is a floor because the rule for which questions count is ours: a question we misclassified reads here as unanswerable. The other 92 are still open weeks after they were written down.

Is a vendor that reprices often also a vendor that publishes more?

Not in this category. Repricing is common here, our price-freshness pass has caught a single vendor publishing four different prices in twenty days, and we predicted before measuring that a category that moves that fast would also answer more of its open questions. It does not. Changing a number and publishing a number that was never there are different behaviours, and this category does the first far more often than the second.

What counts as an answer in this study?

A figure or a term a buyer can read on the vendor’s own page without identifying themselves. Three pages per vendor were fetched on the same day, the URL our catalogue records for that vendor, its apex and its /pricing, and the second arm followed up to four more links from the recorded page. A demo, an email reply or a document behind a form does not count, not because those are illegitimate but because the question is what a buyer can learn before making contact.

Why are five of the questions excluded from the rate?

Because we could not read those vendors’ pages, which is a fact about the fetch and not about the vendor. Two sit behind a bot challenge and three publish their figures only after JavaScript runs. Counting them as “still open” would have asserted that those vendors do not publish, which we did not observe. They are their own bucket in the register and never enter the denominator.