Limits of the method

Vercatus is designed to be quotable to a board. That means being precise about what each number is, and what it is not.

The measurement

Answer Share is a sample of answers, not a count of people. The poll asks questions that stand for your topics, in proportion to their demand, and reports the share of answers that name or cite you, with its range. It does not say how many people saw those answers. Engines answer differently by time, place, account and conversation; the poll asks from your market and records the date of every answer.

The range is the honest part. A share with few answers, or few distinct questions, has a wide range, and a change inside the range is “no clear change”. Read the range before the point.

AI demand is modelled. Nobody who runs an assistant publishes how often a question is asked. Demand is estimated from adoption, engine share and search volumes, always as a range with a grade, and labelled “modelled from adoption and search data, not observed prompts”. Calibration from connected sites narrows it; it never turns it into a count.

ChatGPT is read through its app where it can be. On days the app cannot be read, its developer interface answers instead, and the engine strip says which. The two answer differently.

The reasons

The lenses show associations. Brands on the pages engines read, or in the results of the searches they run, are named far more often; that does not prove a given page caused a given answer. The Search lens checks the index each engine is known to read, which for some engines is only part of what they search.

Engine Briefs are perception. What an engine says about itself often differs from what moves its answers. A brief is shown as what the engine says, and its claims become plays only when the evidence agrees.

Plays and verdicts

An expected gain is an expectation. It is measured from verdicts on similar plays across sites, observed on the topic’s own answers (correlational, and labelled so), or given as a range with no point claimed. It is never a promise.

A shipped play is a live change, not a result. The acceptance check says the change exists; the verdict says whether it moved your share.

Only grades A and B say “caused”. A verdict compares the play’s topics with topics left alone over the same days, from the day the engines could see the change. With nothing left alone, or other changes in the window, the verdict is graded down and never called a win. A single-topic play on a brand that is already named often needs many answers before it can be called either way; the verdict then says how many more it needs.

Natural experiments are observational. They are graded like verdicts and recorded only when clear, but nobody assigned the change, so they inform priors at a lower weight than your own verdicts.

The site audit

A sample is a sample. Findings scoped to a template or the site hold for pages outside the sample; page-level findings do not. A run on a large site is a statement about templates and access posture, not an inventory audit.

A user-agent probe is not the real crawler. Requests carry a vendor’s published user agent from Vercatus’s network. A refusal proves the site refuses that identity from that origin. Sites that verify crawlers by IP or reverse DNS will treat genuine traffic differently. Confirm against firewall and bot-management logs, or connect crawler traffic, before acting on an access finding.

The simulation is a model. It reproduces the shape of a retrieval pipeline over the site’s own passages. Real engines add ranking, freshness and source-credibility signals it does not have. Its verdicts are predictions; the poll measures the real engines.

Lab timings are indicative. LCP, CLS and TTFB come from single headless-browser loads from one location. Confirm with field data before prioritising performance work.

The knowledge base

Hypotheses have grades. A rule grounded in a grade-A source (official documentation) and one grounded in grade-C industry data are not equally certain, and the grade is shown. Entries marked reported were supplied second-hand and never located at source, and their figures are never quoted as fact. Each rule a play rests on also carries what our verdicts observed.

The knowledge base has a date. Every entry carries a revalidation date. The field changes quarterly; so does the base.