Entity consistency
One set of facts about the business, stated identically in every place a model might read them. Old trading names, superseded addresses and dead phone numbers all get cleared.
When a buyer asks an assistant for a supplier instead of searching for one, a shortlist gets made without a results page ever appearing. This is the work of being on that shortlist.
An assistant does not rank pages and hand you ten links. It retrieves what it can, reconciles it, and writes one answer naming a small number of things it is confident about.
Confidence is the operative word. Where the facts about a business are inconsistent, the safe output is a general answer that names nobody.
So the work is not persuasion. It is removing every reason a model has to hedge.
The starting point is a recorded set of answers to the questions your buyers would actually type, captured on a stated date. Most businesses discover they are not mentioned at all. That finding is useful rather than discouraging.
Legal name, trading name, company number, address, phone number and service definitions are made to agree everywhere they appear. This is dull work and it is the single biggest lever on whether a model will commit to naming you.
A question asked as a heading with a direct answer underneath can be lifted cleanly; the same information buried mid-paragraph often cannot. Structured data then states the same facts again in a form that needs no interpretation.
A claim that appears only on your own website is a claim with one source behind it. Independent places repeating the same facts are what turn an assertion into something a model treats as established.
The same questions are put to the same assistants again later. The answers are compared to the recorded baseline. That comparison is the only honest measurement available in this channel today.
Fig. 3
AI assistant referral sessions, 90 days to 19 July 2026
SOURCEDGA4, 90 days to 19 July 2026. Pulled 20 July 2026.
1,434 sessions arrived from AI assistants in ninety days, and roughly nine in ten of them came from one assistant. A site in my own portfolio, shown by sector rather than by name. Results vary.
Two sources were deliberately left out rather than counted up. An SEO tool that crawls with an AI-shaped user agent is not an assistant sending a reader. Including it would have flattered the total.
One set of facts about the business, stated identically in every place a model might read them. Old trading names, superseded addresses and dead phone numbers all get cleared.
Pages are built so that any given fact can be lifted out in one piece without inference. Schema then restates the same facts in a machine-readable form, matching the visible copy exactly.
Answers are recorded verbatim with the date and the assistant named, then re-asked later. Without a dated baseline there is no way to tell improvement from noise.
It assembles an answer from sources it can retrieve and reconcile. It therefore favours businesses whose facts are stated consistently in several independent places. A firm described three different ways across the web is a firm the model has reason to hedge about rather than name.
It overlaps heavily and it is not identical. Ranking rewards a page that satisfies a query. Citation rewards a business that is unambiguous, well corroborated and structured so that a specific sentence can be lifted out and trusted.
No. Nobody controls what a model outputs and anyone selling guaranteed AI recommendations is selling something they cannot deliver. What is controllable is whether the underlying facts about you are consistent, retrievable and clear enough to be used.
It sends real sessions. The chart on this page is from a site in my own portfolio rather than a client. Volumes vary enormously by sector and the channel is currently far smaller than Google for most businesses.
Yes when your buyers are procurement people researching suppliers. That is exactly the kind of shortlisting task assistants get used for. Businesses whose enquiries come almost entirely from local emergency searches should still give Google the bulk of the attention.
The Sprint exists so you can answer this question cheaply before deciding whether it is worth ongoing spend. It is fixed price and it ends with a dated document.