Vercatus Intelligence

Vercatus Intelligence is the part of Vercatus that knows what no single customer could. It is the Learn step under the loop, and it learns three things:

What works: which kinds of play move which engines, from every frozen verdict across sites.
What moved answers across the web: natural experiments, from every page change Vercatus observes.
How much demand there is: AI demand estimates, calibrated by first-party data from connected sites.

You see its output inside your topics, competitors and plays, marked with the Intelligence badge, and on the Intelligence destination.

Play priors

Every verdict frozen at day 45 is stored with the kind of play, the engine, the vertical and the market. Pooled across sites, they become priors: how often plays of that kind won, and their typical effect.

Who contributes. Only real, final verdicts of grade A to C. Each site contributes at most a few verdicts per segment per quarter, so no one site dominates a segment.
Pulled toward the whole. A segment with few verdicts is pulled toward its broader group (all engines, all verticals) until it has enough of its own.
Where you see it. Once a segment has enough verdicts, it sets the expected gain of every new play of that kind, and the play says so: “23 plays like this on 9 sites: won 61%, median +2.4 pp” (illustrative). The play’s page shows Plays like this, the record of its kind. Otherwise the gain falls back to what is observed on the topic’s own answers. A kind of play with a long record of no effect is ranked lower, and says so.

Natural experiments

Vercatus reads far more pages than any customer owns: the pages engines cite, competitors’ pages, everyone’s pages in a topic. Each read is kept as a version. When a page changes, for example it newly names a brand, drops one, or gains a comparison table or a price, and the answers that cite it then move, that is a natural experiment.

It is measured like a verdict: the brand’s Answer Share on the topics that cite the page, before and after, against the same brand on topics where the page is not cited, with the engines that never cited it as the check that should stay flat. Only clear reads at grade A or B are recorded. They appear on a competitor’s page under what changed, for example “Fabrikam was added to a widely cited list on 12 Sept; its share on the topic rose 9 points by 30 Sept (grade B)” (illustrative), and they feed the priors at a lower weight than customer verdicts.

This is how Vercatus learns from the whole market, many times faster than from its customers’ plays alone. Natural experiments use public pages and public answers only.

AI demand, calibrated

Every topic’s AI demand is modelled from public adoption and search data, as a range with a grade. Connected sites make it better. For every connected site, Vercatus compares the first-party AI exposure of the pages that serve a topic (Search Console’s AI impressions, Bing’s AI citations, assistant fetches from crawler logs) with modelled demand times the site’s Answer Share there. The ratio, per engine, market and kind of topic, is pooled across sites and scales the model, and the range narrows as more sites connect. AI-assistant sessions from Analytics are shown as a floor and never scale the estimate.

Your site’s data never leaves your workspace: only the ratios are pooled.

Engine playbooks and engine changes

The Intelligence destination answers three questions:

What works: per engine and vertical, the kinds of play with the best record, as plain sentences with their counts, filtered to your verticals and engines.
How much demand: your market’s topics by AI demand, with ranges, grades, calibration, and the coverage of demand.
What changed: confirmed engine changes in plain words, with the topics of yours they touched, for example “an engine changed what it cites; on 4 of your topics the cited pages changed” (illustrative).

It shows no vendor-side counts. Every figure on it answers a question about your market.

Where engines look: forums and communities

Beside your own Forums and communities, Intelligence pools the same count across every site in a market: the share of answers that cite any forum, Q&A site or community, per engine, and which platforms they are (Reddit, Quora, a country’s own forums). It sits under your own figure on Sources, so a site with little data of its own still sees where its market’s engines look. It is shown from three sites and thirty answers in a market, a platform or an engine from three sites, and it names platforms, never a subreddit, a thread, a site or a workspace.

The knowledge base learns

Each rule of the knowledge base that a kind of play rests on gains an observed line: supported by verdicts, not observed, or contradicted, with how many verdicts and sites stand behind it. A rule the verdicts contradict is demoted from play generation.

Privacy floors

Aggregates only. Priors, playbooks, calibration ratios, observed lines and the market’s share of answers citing communities hold counts and effects, never a workspace, a site or a question.
At least three sites. Nothing is shown, and nothing sets a play’s gain, until at least three sites contribute to it. Below that, a line reads “not observed”.
Few from each. Each site contributes at most a few verdicts per segment per quarter.
Public data for natural experiments. They are built from public pages and public answers.
Simulated reads never count. Demonstration data is labelled and never enters a prior.
Published only on purpose. Learnings appear on vercatus.com only above the same floors, and only once we choose to publish them.

What is collected, and what never leaves a workspace, is on Security and data handling.

Ask Vercatus Intelligence

The assistant in the workspace answers questions about your sites and how Vercatus works. See Ask Vercatus Intelligence.