The system behind a mandate.

How the supply side works

The supply role is not a marketplace of freelancers waiting for buyers. It receives a funded work package from an approved mandate, for example: verify the technical product information for one product family, improve six existing buying pages, and implement a quote flow, within this allowance. It then compares doing the work internally with using a vetted specialist, checks availability and competence, gets a scoped commitment, routes the output through review, and records the delivery quality and the real cost. The pool of specialists can be maintained in advance; bespoke commitments follow a qualified mandate, never precede it.

Third-party coverage is handled separately from your own pages. We can commission original research, pitch legitimate stories to editors, and buy clearly disclosed sponsorship where it makes sense. We cannot sell you control over an independent editorial decision or an AI citation, and nobody who claims to can deliver it honestly.

The design decisions matter more than the agent count

Ten agents is a diagram, not an argument. Four decisions underneath them are what make the output worth acting on, and each one costs something — which is why they are the first things a cheaper pipeline drops.

• The prompt set is frozen per client. Once a corpus is agreed it does not change between runs. Swapping questions is the easiest way to make this metric say whatever you want, and it is undetectable from outside. The cost is that a frozen set slowly ages against a moving category, so we add questions in a separate, separately-reported tranche rather than editing the baseline.
• Every prompt is asked more than once. Generative engines sample; they do not look up. The same question asked three times will often return three shortlists. Repetition is most of the compute bill and all of the reason a number can be compared to itself next quarter.
• Raw transcripts are kept, addressed by run id. Every figure we publish carries the identifier of the run it came from, so any of it can be re-derived rather than taken on trust. The two runs on our numbers print theirs.
• Inconclusive is recorded as inconclusive. When the competitor-discovery pass cannot read a site — JavaScript-only rendering, or a bot block — that candidate is dropped rather than guessed at from the model's description of it. The alternative is a dataset that quietly contains the model's memory instead of the web.

What the stack cannot do

It cannot reliably tell you whether an engine read your page or merely linked it. Those are different events, and a citation count collapses them — we pulled them apart on eight answers and they inverted, which is written up in naming your URL is not reading your page.

It cannot make an engine cite you. No system can, because the answer is generated rather than retrieved from a ranked list, and anyone selling a guaranteed placement is selling something they do not control. What the pipeline does is measure the gap precisely enough to argue about, and then do the retrieval and distribution work that changes the inputs.

It also cannot sample an engine that refuses. Blocks happen and they are not predictable: Google AI Overviews blocked our own August run and Perplexity partially blocked the July one. Where that happens the run records no reading for that engine rather than an estimate or a zero. This is the main reason a single-engine dashboard flatters itself — if you only ever ask the engine that always answers, you never see a gap in your own data.

And it cannot tell you how long a fix takes. There is no primary evidence for time-to-first-citation that we could find, so the pipeline reports what changed and when, and declines to forecast. That grading sits with everything else on the evidence register.

The parts that are deliberately not automated

A human reviews and tags every prompt before a corpus runs, because a corpus full of prompts containing your own brand name will find your own brand and prove nothing. Distribution drafts queue for approval rather than posting, because automated participation in communities is both against their rules and a reliable way to damage the reputation you hired someone to improve. Even the photography on this site passes a human gate after the licence filter, which is documented down to the individual image on the credits page.

If you want the shape of an engagement rather than the architecture, the services pages set out each stage and its price, and the methodology covers how discovery, buyer behaviour and commercial outcomes are measured.