What we have actually tested about crawler access, extractability, schema and citation behaviour, with the instrument and the date attached to every number.
No, the major AI crawlers do not execute JavaScript. They read whatever HTML arrives in the first response and nothing else, which means a client rendered site hands them an empty container. We fetched our own routes with the documented GPTBot string and published the byte counts, including the ones that embarrassed us.
Training crawlers and search crawlers are different machines with different consequences, and one robots.txt line written by someone who did not know that can remove you from AI answers entirely. Every documented agent string, what it is for, and the specific cost of denying each one.
Retrieval scores passages rather than pages, which is why a narrow self contained section can enter an answer the whole page would never have ranked for. The traits that separate a quoted passage from an ignored one, the five tests we run before publishing, and the widely repeated citation statistics we refuse to print.
Mostly the same discipline, and Google says so in its own documentation. Two differences are genuinely new: the AI crawlers do not run JavaScript while Googlebot does, so a site can pass classic SEO and fail retrieval outright, and retrieval scores passages instead of pages. Everything else being sold as new is not.
Google states plainly that structured data is not required for its generative features. Microsoft has said on the record that schema helps its models understand content. Both are true at once, and the reason to ship it is entity disambiguation rather than a ranking effect. Which types still pay for themselves, and which are dead.
An assistant is not reading your about page and getting it wrong. It is reconciling sources that disagree with each other, and the loudest disagreement wins. The five step correction loop we run, starting with the part most teams skip: finding out which sources the model is actually reconciling.
Two free first party reports exist and almost nobody has switched them on. For the assistants that report nothing at all, the honest instrument is a fixed query panel run the same way every month. Also why referral traffic undercounts, and why a single composite visibility score cannot move for a reason you can act on.
A page is invisible to an AI crawler when the first response carries no body copy, which is measurable in bytes rather than arguable. The method, the shell baseline measured on the same domain the same day, and the raw counts for every route type we fetched with the documented GPTBot string.
Every number this site publishes, what instrument produced it, what it does not cover, and where we have no data at all. The failing rows from our own audit are left in, because a results page with no failures in it is a marketing page wearing a lab coat.
The table every other page here defers to. Each load bearing claim appears once with a confidence label, its source, the date it was last verified and the date it stops being trusted. Rows are ordered by recheck date, soonest first, so the claim closest to expiry is the first one you read.
InboundPros and Outbound Pros are the same group, so recommending the agency without applying our own labels to it would make the labels decorative. Here is the group agency run through the same confidence rules as everyone else, including the claims we will not stand behind and the buyers it fits badly.