Pulse Method — how Pulse is measured

VERSION 1.1 · QUESTION LIBRARY PQL 1.0 · EFFECTIVE 24 SEPTEMBER 2026

Pulse runs on the same controlled infrastructure as the Pitot Index, but it is a different measurement, with its own questions and its own version. This page sets out how it works, what it does not do, and what is withheld. It is complete enough to show the controls and deliberately not detailed enough to copy. See also the Index methodology.

1. What a Pulse measures

A Pulse is a single-snapshot health-check of how AI systems describe a named brand, in a market the buyer chooses. The brand is named in every question, so a Pulse is a prompted assessment: it shows how the systems characterise the brand when asked about it directly.

2. Pulse and the Pitot Index are different measurements

Pulse runs on the ARES-C infrastructure: the same pinned model registry, the same capture of the model actually served on each call, the same extractive scoring discipline and the same isolation rules. It does not use the Pitot Index benchmark questions and it is not an Index measurement.

The Index is unprompted: the brand is never named, and the systems surface and rank brands on their own every week.
Pulse is prompted and framed in one market, on one day.

The two answer different questions, so a Pulse result is never compared with an Index rating. Pulse measurements are kept in their own lineage. They never enter the Index record, and a Pulse run cannot read from or write to the Index vault.

3. The question library

Every Pulse asks the same fixed library of questions (PQL 1.0).

Four pillars. Authority, Reputation, Expertise visibility and Contextual presence. Each pillar question asks the system to commit to one label from a fixed set and then justify it in a sentence.
Self-perception battery. Open questions on what the brand stands for, its perceived strengths and weaknesses, its closest competitors and the gap between intended and received identity. These feed the intended-versus-received result.

Only three things change from one order to the next: the brand name, the sector and the market. Nothing else a buyer supplies enters a question.

The library is fingerprinted. Each run recomputes a cryptographic fingerprint over everything that decides what is asked and how it is scored: the questions, the market frame, the label parser, the band thresholds and the market list. If the fingerprint does not match the published version, the run is refused. A change of any kind therefore produces a new version and a changelog entry here. It never happens silently.

4. The market frame

The buyer chooses one market:

a country or territory — 242 in all, identified by ISO 3166-1 code (uninhabited territories excluded); or
a region — Worldwide, Europe, North America, Central America and the Caribbean, South America, Latin America, Middle East and North Africa, Sub-Saharan Africa, Asia-Pacific, East Asia, Southeast Asia, South Asia, Central Asia, Oceania.

Every question is asked from the perspective of a typical consumer of the category in that market, and the sector is stated with the market attached. Systems are asked to say so when a brand is little known or unavailable there, rather than describe its standing elsewhere. Worldwide uses the same international, no-allegiance frame the Index uses.

The questions are asked in English. A system can answer differently when asked in a local language. Every report says so.

5. The systems

Buyers choose AI systems by sphere: Western (Base or All models), Eastern, Middle East & Africa, or all of them. The version recorded for each system is the one that answered: the model identifier its provider returns with every response, never the name we asked for. If more than one version answers within one order, the report lists each with its share of answers. Before any measurement starts, every system (and the judge) must answer a short check call; if one does not, the order stops and nothing is measured. Systems are named in one place only, the report's Method & versions page.

6. Scoring

Pillar and overall results. Each committed label is parsed extractively (the label the system chose, not an interpretation of its prose) and placed on a fixed internal scale. A pillar's result is the mean across its questions and systems. The overall result is the mean of the four pillars. Results are reported as one of five bands: Dark · Faint · Uneven · Established · Strong, which are equal-width bands on that internal scale. The internal number is never published, so a Pulse band cannot be mistaken for an Index rating.
Agreement across systems. For each question, the share of systems that chose the most common label, averaged over the pillar: High (three quarters or more), Mixed (half or more), Low (below half).
Intended versus received. The buyer's one-sentence positioning is compared with each system's self-perception answers by a separate judge system, working at zero temperature against a fixed ten-point rubric with five classes (Strong, Good, Partial, Weak, Misaligned). A result whose score and class disagree is flagged for review.
Competitor read (optional). Up to three competitors, measured in the same run, with the same questions and the same systems. It is reported in bands, never as a league table.
Failures. Failed calls and answers without a committed label are excluded and counted. The count is part of the order's quality record.

7. Analysis and review

The written sections (summary, pillar findings, intended versus received, root-cause hypotheses) are drafted from the scored results only. No raw answer is quoted. A Pitot analyst reviews every report before issue against a stop-ship checklist, which covers model names outside the Method page, internal identifiers, and statements unsupported by the results.

8. Integrity of the intake

Text a buyer types must never steer a system or the judge. So:

the intended positioning is one sentence of at most 25 words;
brand, sector and competitor names have length and character limits, and markup, brackets, quotation-mark runs and web addresses are refused;
instruction-like wording is refused, and borderline wording holds the order until a person has read it;
inside the drafting step, everything the buyer typed is marked as data, never instructions.

None of these rules can change a question, a label set or a score. They only decide whether an order is accepted.

9. What a Pulse does not do

No trend. A Pulse is one pass on one day. Only repeated measurement can show direction or stability.
No divergence measurement. Pulse does not measure how live-search AI departs from what a model has learned.
No Index rating. A Pulse band has no place on the Index.
No local-language measurement. Questions are asked in English (§4).

10. Reproducibility and versioning

Every report is stamped with the method version, the question-library version and fingerprint, the market, the option, the systems as served, the date and the order reference. Two Pulses are comparable only when the method version, market and option match, or when the changelog declares two method versions comparable. The raw answers are kept in the order's audit trail and are not distributed. Every finished Pulse is sealed: its files are hashed, catalogued and kept write-once for seven years in a Pulse archive that is separate from the Pitot Index record. A monthly Pulse joins into one trend only checks measured under comparable method versions, in the same market and with the same option.

11. What is withheld

Question wording, question counts, label sets and their mappings, the persona wording, the judge procedure in detail and dispatch mechanics are withheld to protect the instrument. They are available to bona fide researchers and auditors under agreement.

Changelog

1.1 · PQL 1.0 — 24 September 2026. Renamed from “Pitot Pulse Method” to “Pulse Method”. Name only: no question, frame, label, band, market or scoring rule changed, so results under PPM 1.0 and Pulse Method 1.1 are directly comparable. The fingerprint changes because it includes the method version.

PPM 1.0 · PQL 1.0 — 21 September 2026. First published version. Introduces the buyer-chosen market frame (242 countries and territories, 14 regions), the fingerprint check, the intake integrity rules and the judge consistency check. It supersedes a pre-release run profile that had no market frame and was used only for internal tests. That profile was never sold.