MethodologyAction plan
How we decide what to recommend
Knowing your share of AI recommendations is only useful if you know what to do about it. The Action plan turns a store's measured visibility into a prioritized list of opportunities, and this chapter explains exactly how a recommendation earns its place on that list.
Evidence, proposals, judgment
A run has three stages. The store's measured data goes in; a prioritized, evidence-anchored list comes out. No stage trusts the previous one blindly.
Evidence
Everything AI Visibility measured for the store is assembled into one evidence pack: the raw 30-day corpus of AI answers, share of voice per category, the sources engines actually read, competitor standings, and the technical readiness picture. Recommendations start from what the engines really said, not from generic best practice.
Proposals
Three models from three different vendors read the same evidence independently and each proposes opportunities. They also receive our researched, evidence-graded knowledge base of how AI engines select sources, so proposals follow verified mechanisms rather than SEO folklore. The models genuinely diverge; each catches things the others miss.
Judgment
A separate judging pass, currently run on one of the same models, merges the three proposals into one prioritized list, recording for every action how many models arrived at it independently. A sound idea only one model found is kept, and labeled as such.
Models choose, code counts
The single most important rule in this pipeline: a language model is never the source of the evidence numbers you see.
When a model wants to ground a recommendation, it does not write statistics. It selects anchors: the names of real categories, real tracked prompts, real competitors, real source domains. Code then verifies every anchor against the store's live data; an anchor that does not resolve to something real is discarded. Only after verification does code compute the evidence you see, how many prompts are affected, the gap in percentage points against each named competitor, which cited sources are involved. If a model fabricates a reference, it is dropped at this boundary and never becomes evidence.
What one recommendation carries
An illustrative action, anonymized; the fields are exactly what the live product stores.
content_signal:comparison-pages
Publish comparison pages for your five weakest categories
In these categories the engines cite comparison and review content far more often than store pages, and your domain is absent from every cited comparison.
Anchored to the affected category, its tracked prompts, and the cited comparison sources, all verified against live data.
Support says how many of the three models proposed this independently. Confidence and horizon come from the judge. The evidence anchors are verified by code against the store's data before anything is shown. Actions keep a stable identity between runs, so when a new run confirms an existing recommendation, your planning state on it, planned, in progress, done, is preserved rather than reset.
What keeps it honest
Models never write the numbers
Recommendations reference real things: categories, prompts, competitors, sources. Every reference is verified against the store's live data and fabricated ones are discarded. Every count and gap is computed by code from the verified references.
Agreement is shown, not hidden
Each action carries its support, the judge's own record of how many models arrived at it independently. A recommendation only one model found is kept when it is sound, and labeled as exactly that.
The ordering is a prioritization, not a promise
Actions are ranked by a fixed weighted blend: confidence 45%, cross-model support 30%, evidence strength 15%, time to impact 10%. It is a defensible ordering of where to start, and it is never presented as a measured business result.
No causal victory laps
AI engines re-read the web on unknown schedules, and answers from model memory can take months to move. Measurement is framed as observation of the affected area over time, clearly labeled a signal, never proof.
Common questions
Why three models instead of one?
Because they genuinely disagree, and the disagreement is information. Each model catches opportunities the others miss, and when all three arrive at the same recommendation independently from the same evidence, that agreement is the strongest signal we have. Every action shows its support, for example 3/3, so you can see it.
What stops the AI from inventing evidence?
A hard boundary in code. The models select references to real things: category names, prompt fingerprints, competitor labels, source domains. Every reference is verified against the site's actual data, and a fabricated one is discarded. Every number you see, affected prompts, gaps versus competitors, cited sources, is computed by code from the live data. The models never write a number.
Will you tell me whether an action worked?
Carefully, and honestly. AI engines re-read the web on unknown schedules, and some answers come from model memory that may not move for months regardless of what you change. A before/after chart that ignores this would flatter us and mislead you. So measurement is framed as observation, not attribution: we monitor how your presence develops in the affected area and label it as a signal to watch, never as proof an action caused it.
Can I use this today?
The Action plan is in closed beta while we tune it against real business data, and runs are triggered manually rather than on a schedule. The AI readiness and AI Visibility chapters describe what is live for every account today; if you want early access to the Action plan, get in touch.
Previous chapter
How we measure AI Visibility
The Action plan is built on AI Visibility data; that chapter explains where every number in the evidence comes from.
Start with the measurement
The Action plan is in closed beta, but the measurement it builds on is live: see who AI recommends in your categories today.