We would rather be audited
than believed

The command that started this

In August 2026 I was planning a group of new web properties and wanted to check something obvious before writing the brief: what does our own site actually serve to an AI crawler. So I fetched every route type on the agency site with the documented GPTBot user agent string and looked at the raw response bodies.

The blog came back fully rendered, because a prerender step had been written for those routes and it worked. The homepage came back as an empty root div and a set of head tags. The service page came back at roughly the same size, which against that baseline is a heading and about one sentence of real content.

Our most important page, the one a model would fetch if somebody asked an assistant what our company does, was blank. It had been blank for months. Nothing in Search Console said so, because Googlebot renders JavaScript and the AI crawlers do not, and no report anywhere in an ordinary marketing stack shows you that difference.

That is not an unusual situation. It is close to the default for any company that built its marketing site on a modern JavaScript framework in the last five years, which is most of them. The unusual part was noticing, and the only reason I noticed is that I had a specific reason to run a fetch as a bot on a Tuesday afternoon.

The second reason, which is about the field rather than the site

When I went looking for what to do about the problem, most of what I found was citation multipliers with no methodology attached, tools promising a single AI visibility score, and confident advice that the platforms' own published documentation directly contradicts. Meanwhile the actual primary sources sat there being ignored, because they say boring things and boring things do not get shared.

The gap is not a knowledge gap. Everything worth knowing in this field is public: the crawler documentation from each provider, Google's own optimisation guidance and its section on what you do not need to do, Microsoft's statements about schema, the server log study everybody cites, the measurement papers nobody reads. The gap is a discipline gap. Somebody needed to write the version that names its sources, labels its confidence and admits the parts nobody outside the platforms knows. That is this site.

Why this is its own property rather than a section of the agency site

Three reasons, and the third is the one that decided it.

They are different disciplines with different buyers. Outbound is a volume and infrastructure problem: whether the right people know you exist, solved with lists, deliverability, copy and consistency. AI search visibility is a retrieval problem: whether a machine can read and quote you when a buyer who already has the question goes looking, solved with rendering, structure, entity clarity and corroboration. A company that conflates them spends outbound money on a retrieval problem, or the reverse, and both are expensive mistakes.

Depth beats breadth in every retrieval system worth optimising for. A single site trying to be the authority on cold email deliverability and on AI crawler behaviour competes against specialists on both and beats neither. Splitting the topics means each property can go deep enough to be worth citing, which is the only version of this that works.

And this site is supposed to be a demonstration of its own subject. It ships real HTML on every route, an explicit crawler allowlist, one consistent entity graph, original first party data and a standing rule about which numbers it will not print. If it were a subfolder on an agency site, none of that would be independently checkable. As a standalone property it can be fetched with a bot user agent by anyone who doubts a word of it, which is the point.

How we decide what goes on a page

Four rules, and the fourth is the one nobody else in this category has adopted: every claim we publish carries a date on which it stops being trusted.

One. Every claim carries a confidence label

Confirmed means the platform stated it in its own documentation, or a study measured it with disclosed methodology and a sample worth trusting. Strong means several independent sources agree, including at least one primary, with no credible contradiction. Contested means the sources disagree, or the only sources are vendors selling the answer. Unknown means nobody outside the platform knows. A number reaches a live page only when it is confirmed or strong and the source can be named in the same sentence.

Two. Unknowns get published as unknowns

How OpenAI ranks candidates for citation. Whether ChatGPT runs its own index alongside its search partner. Which provider powers Claude's search. Whether any answer engine consumes dataset markup at all. We do not know, and neither does anybody selling you a service that depends on knowing. Saying so plainly is the single most useful thing a practice in this field can do right now, and it happens to be the kind of statement a model cannot synthesise from the rest of the web.

Three. Some numbers never go on a page, even to debunk them

The GEO category runs on citation multipliers with no method behind them: one for comparison tables, one for question and answer markup, one for content recency. Single vendor, no published methodology, no replication, no denominator. We name those claims as unsourced and we never print the figures, including on the pages that exist to refute them, because a table of banned numbers with the refutation in the next column is still a table of banned numbers, and a retrieval system lifting the left column carries them off unqualified.

Four. Every claim carries a recheck date, and the register is public

On denominators, this site applies a narrow version of a rule that runs across the whole group: no figure goes on a page without its instrument, its sample and its date in the same sentence. The wider argument about what campaign rates are divided by, and why two ratios with different denominators must never be set beside each other as if they were comparable, is go to market math and lives with the group property that owns that subject. We follow it. We do not re-explain it here.

Who runs this

It is worth explaining how somebody who spends his days on sending infrastructure ended up running a practice about being read by language models. The answer is that both jobs are the same job seen from opposite ends. Outbound asks whether the right people have heard of you. Retrieval asks whether the machine they ask about you can find anything to quote. The discipline that transfers is the one about denominators: knowing what a number was divided by, and refusing to publish it when you do not.

Who owns this site, and why that is a credential

InboundPros is owned and operated by Jānis Plūme and is part of the Outbound Pros group, alongside a managed B2B outbound agency and sibling properties covering LinkedIn outreach mechanics, multichannel cadence and go to market math. That appears in the footer of every page here, and in this site's schema as a parent organisation rather than only in prose.

The second reason it is a credential is mechanical. Language models resolve companies by corroboration across sources. Several properties that each independently define the same organisation, with matching sameAs and parent organisation markup and one shared author identifier, make that entity more resolvable rather than less. Hiding the relationship would have meant either omitting the schema that makes the family machine readable, or shipping schema that contradicts the copy. Both are worse than the truth, and one of them is the exact pattern Google's spam policy started naming in 2026.

The third reason is the plainest. This is the normal structure of a company, stated out loud. A disclosed brand family where each property owns a distinct topic, publishes original data and ships a working tool is a corporate structure, and the links between the properties are corporate links. The alternative, an undisclosed network of sites recommending each other, is the thing the spam policies are actually aimed at. We would rather be the boring version.

What would change our minds

Two findings, both written down in advance on the evidence register beside the claims they would break. A rigorous public re-test showing that the major AI crawlers now execute JavaScript would collapse the first gate of our whole model, and we would rewrite the site rather than defend it. Any major provider committing to llms.txt in a production answer surface would move our position on it from ship it and expect nothing to something warmer. Naming both in advance is the only version of that promise worth anything.

Contact