Three pages the AI engines told us to publish, and why we did not publish them

On 22 October 2026 we ran a measurement in which we asked ChatGPT, Gemini and Claude what we would have to publish before a vendor site like ours was worth citing, and kept every answer as a page specification. There are 224 of them: 104 from Gemini and 83 from ChatGPT through the interface, 31 from ChatGPT in conversation and 6 from Claude in conversation. 162 came with a full outline and 202 stated the dataset the page had to contain.

We are working through them. Three were refused, and publishing the refusals is cheaper and more useful than publishing the pages would have been, because each refusal names a real gap in the product.

Refusal one, a 1,200 prompt SaaS benchmark we never ran

The specification asked for a benchmark of 1,200 unique B2B intent prompts evaluated across four named engines, with a base recommendation rate and a lift figure attached: its own suggested headline was a 4.35 times increase in recommendation likelihood, from 14.2 per cent to 61.8 per cent, when a vendor exposes public API documentation and pricing schema on its own domain. It also attached that result to a 19 dollar a month plan.

Our run asked 318 questions, not 1,200, and the measurement arm covered two engines, not four. We hold no base recommendation rate for SaaS vendors, no lift measurement, and no 19 dollar plan. Every figure in that specification is a plausible shape with no observation behind it, which is exactly the content this category already has too much of.

Refusal two, an ideal customer profile page with thresholds we have not measured

The specification asked for a fit and non fit guide with a binary matrix, and the numbers it required were a minimum viable budget, an optimal company size, a category maturity threshold of at least 50 monthly AI buyer queries, and a 24 hour review window before content goes live. It also wanted on premises air gapped deployment listed as a disqualifier.

We have not measured a query threshold at which this work starts to pay, we publish no review window service level, and listing air gapped deployment as a disqualifier implies an enterprise posture we do not have. Writing the page would have put four unmeasured thresholds on a live site in the voice of a qualification rule.

The part of that specification we can answer, we answer on the pages where it belongs: who should not buy this is on every page that names us, and the honest version is short. If you want one score a month from one reading per prompt, do not buy it from us, because our own run to run measurement is the argument against that product.

Refusal three, an evaluation framework with a fidelity figure nobody holds

The specification asked for a buyer guide to AI visibility platforms whose headline was 98.4 per cent citation extraction fidelity across a benchmark of 10,000 multi engine prompt outputs tested on 1 October 2026, over 2,500 distinct intent prompts across six named platforms, with a section on enterprise security, single sign on and API access.

That benchmark does not exist. Not ours, and not one we can read from anybody else. We did not run 10,000 prompts on 1 October 2026 and neither did anyone who published a figure like it. On the security section, our pricing page renders none of SSO, SAML, SOC 2, ISO, DPA or white label, checked by word boundary across 119,203 bytes on 22 October 2026. A page built to that outline would have had to fill that section with something.

The gaps the three refusals name, stated once and plainly

• We read five surfaces. Not nine. Fetched by us on 22 October 2026, tryprofound.com/pricing publishes up to 9 answer engines tracked on its Enterprise tier, along with SSO and SAML and SOC 2 compliance, all time data history, CSV and JSON export, an API and unlimited seats, with the price given as Custom and no figure in any currency on the page. On certification and engine count, that is a better product than ours.
• No enterprise security words render on our pricing page. The substring trap is worth knowing if you check it yourself: a case insensitive search for sso on that page matches crossOrigin seven times and subprocessors twice, so a careless check reports a feature that is not there.
• Our own page contradicts itself in three places, and we publish both readings rather than the flattering one. The Agency row reads businesses covered 1 1 1 1 1 1 10 1 while the comparison table answers No to more than one business in all eight columns. The main listing names five apps including Claude while the Scale plan description names four without it. The FAQ says SEO Start gets 12 pages a month and no AI app checks, beside a table listing five engines for that plan.
• We hold no prompted citation figure at all. Six brand questions that name our own company exist in the bank and that batch was never run, because the runner skips every brand row in both its validation and its job loop. So the ceiling was never measured and nobody should quote one from us.

Why refusing is the cheaper move

Across the 224 specifications, 71 asked for a conflict of interest or funding disclosure and 152 asked for a date, counted by word boundary over the specification file. A page that satisfies the outline and fails the disclosure is the shape that produced our zero: 507 new recommendation pages between 20 and 22 October 2026 moved our citation count from 6 of 1,260 readings to 0 of 635 and our naming count from 0 to 0. Publishing the three refused pages would have added three more pages of that shape and three claims we could not defend. The counting rule and both tallies are in what makes an assistant cite a vendor website.

The decision these refusals force is not a content decision. If buyers keep asking for SOC 2, single sign on and nine engines, that is a product roadmap question and it belongs to the people who own the roadmap.

How to apply the same gate to your own pages

1. Take every number in the draft and ask who observed it, on what date, with what denominator.
2. Delete the ones with no answer. Do not soften them.
3. If deleting them empties a section, the section was the problem and the page may not be publishable yet.
4. State what you do not have, by name, in the same page. It is shorter than it feels and it is the part a sceptical reader checks first.
5. Attach every vendor fact to the exact URL you read it on and the day you read it, never to the brand.

Questions about refusing a content brief

Did the engines invent those numbers

They produced specifications containing figures with no source, which is what a model does when asked for the shape of a convincing page. A 98.4 per cent fidelity rate dated 1 October 2026 is a shape, not an observation.

Are the other 221 specifications usable

Most are. 202 of 224 stated the dataset the page must contain, and the ones we can satisfy are being written. The three here asked for capabilities and datasets we do not have.

Why publish your own pricing page contradictions

Because both readings render, and picking the flattering one is the thing this page exists to refuse. The vendor comparison criteria we hold ourselves to are in the 27 question checklist.

Does this mean the product is not enterprise ready

It means those words do not render on our pricing page today, checked by word boundary. If your procurement needs the certificate, the vendor named above publishes it and we do not.

The missing accuracy benchmark named in the third refusal is set out in full in has anybody measured whether AI visibility tools report accurately, and what that test needs, and the clauses that put these disclosures into a signed document rather than a blog post are in what to put in a contract for AI search visibility work, clause by clause.