Marketing teams are unusually bad at this, because years of positioning work trains people to describe the product the way the company wants it described. A prompt set written by the people who wrote the positioning tends to measure the positioning rather than the market.
The Blocks Nobody Chose Most blocking discovered during audits was never a decision. A disallow copied from a template. A staging rule that survived a migration. A security plugin with an aggressive default. A content delivery network setting labelled bot protection with a switch nobody has looked at since launch.
The Argument Against Waiting The usual counterargument is that assistant traffic is still small in most categories, which is often true. But the audit is not primarily about capturing that traffic. It is about finding out whether you are mechanically invisible, whether your identity is coherent, and which third party pages your category's answers are built from.
Measure Position Change in the Prompt Set This is the closest thing to an output metric that you can genuinely audit, because you own the instrument. Run a fixed prompt set on a fixed schedule under fixed conditions, and track four things:
The fix is straightforward if slightly humbling. Pull the language from sales call notes, support tickets and the search queries in Search Console, then have somebody outside marketing read the prompt set and flag anything that sounds like a brochure.
The missing skill is the reflex to ask for the sample size and the publisher before repeating a figure, and to attribute it when using it. Teams that skip this end up presenting a vendor's marketing to their own board as market data, which is a difficult position to recover from.
This means a single answer is a sample. Being absent once is not evidence of a problem and being named once is not evidence of success, and treating either as a result is the most common analytical error in this field.
Blocking these is therefore not one decision. Turning away a training crawler is a defensible editorial position. Turning away the agent that fetches pages at answer time removes you from answers entirely, and the two are frequently confused.
One presentational point makes this considerably easier to defend. Put the limitations on the first page rather than in a footnote. A report that opens by stating what cannot be measured is read as careful, while the same information discovered later is read as something that was concealed, and the difference determines how the numbers around it are treated.
The most citable content most businesses could publish already exists, unwritten, in sales calls and support tickets. It is the set of questions people actually ask, with the answers your team gives verbally every week and has never put on a page.
One organisational habit makes this sustainable. Give the sales and support teams a single place to drop questions as they hear them, with no process attached beyond writing down the question in the customer's words. Anything more elaborate stops being used within a month, and a shared document with fifty verbatim questions in it is worth more than a formal intake process nobody completes.
The honest position is that attribution in this channel is harder than in any other you are currently running, and the field has responded to that difficulty mostly by inventing numbers. Confident figures circulate widely, and a surprising share of them trace back to a vendor's own sample or to a study far smaller than the claim implies.
Get the Basics Right Before Anything Clever Once access is confirmed, check that content actually exists for a crawler to read. Load your important pages with JavaScript disabled. If your specifications, pricing, service areas or contact details vanish, they are effectively absent from this channel regardless of how permissive your robots file is.
A frequently quoted comparison showing assistant referrals converting several times better than search came from a vendor selling the service, across 312 business to business brands. A widely shared claim about explosive referral growth rested on nineteen analytics properties. Both are legitimate observations and neither supports the confident generalisation usually attached to them.
Why Real Questions Beat Generated Ones Questions produced by keyword tools are smoothed. They use category vocabulary, they avoid awkward specifics, and they tend to be the questions everyone has already answered.
If you must change the prompt set, add new prompts as a separate cohort and keep the original series running unchanged. Editing the instrument retrospectively destroys the comparison you have been building.
This entire area usually amounts to a day of work. It is routinely the difference between a brand that appears in answers and one that does not, and it is worth doing before anybody writes a single word of new content. ai visibility Agency
Then listen for language. When prospects begin describing your business using phrasing you did not write and your competitors do not use, that phrasing came from somewhere, and generated answers are an increasingly likely source. It is anecdotal, it is not a number, and it is often the earliest indication that anything is working.