A page that ranks can hand GPTBot an empty HTML shell

Two gates sit in front of any assistant answer that could cite your site. The first is whether a named user-agent is allowed to fetch the domain at all ^[2]. The second is whether the bytes that come back contain the explanation ^[5]. Two gates with two outcomes each gives four states, and only one of them puts product copy where a retrieval system can read it ^[1]. The state that costs you is the one that passes the checks you already run: fetch permitted, response nearly empty, deployment green ^[7].

The access gate fails quietly because the rule was usually not written as a decision. The dev.to post lists where such rules come from: a template, a hosting configuration, or a change someone made for a reason nobody documented ^[3]. Ordinary search traffic keeps arriving the whole time, so nothing surfaces ^[4].

The content gate is the more interesting problem, because the page is genuinely there. Many sites return an HTML shell and paint the copy after JavaScript runs, so a fetcher that does not execute scripts receives a mostly empty document ^[5]. A login, a demo gate or a chat widget produces the same result, with a human clicking around seeing a complete product while the retrievable page holds almost none of the category language a buyer would use ^[6]. That is the argument for reading the HTML instead of a screenshot ^[7].

The rest of the post's case needs a further assumption before it transfers. Its model is that an assistant builds a category shortlist from what it can retrieve, meaning pages it managed to fetch plus public text where people discuss the category by name ^[10]. On that model a competitor with a worse product gets named because more retrievable evidence exists that it exists ^[9], and sites present in both places become the default answer ^[11]. Nothing in the post measures any of it: no crawl sample, no share of ranking domains that disallow GPTBot, no logs from an assistant ^[17]. The first gate is still cheap to test on your own domain, at one request per named agent and a look at the response body ^[8].

On instrumentation the post is blunt. Watching an AI visibility score does not change the record the answers are assembled from, and the author wrote of such a score, "It is a thermometer, not treatment." ^[12] Paid placement gets the same reading. Ad impressions are not text an engine retrieves when it assembles a category answer, which the post gives as the reason ads are a poor first fix for never being named ^[13].

The author calls the audit table "deliberately unglamorous" ^[14]. That is the most credible sentence in the piece.