Skip to main content
The Heralded Score is an index from 0 to 100. It measures how often AI answers to your tracked questions name and recommend your brand. It is not a share of all buyers, and it is not traffic. The score uses the AI side only for now. Search has no score until its scale keeps the score at or below its own Mentioned and Recommended. Search visibility keeps those shares and each keyword’s results. The API serves the Heralded Score for both ai and all. Figures in the API describes the sides. The score’s info button keeps its definition and range. This page holds the formula, engine weights and calibration method.

The raw score

On the AI side, each answer earns points: A Weak answer earns the same 50 as a Mentioned one, even when it names you to criticise you. The score measures whether the engines put you in front of buyers. What they say about you is how AI describes you, shown next to the score and not part of it. Averaged the same way as Mentioned and Recommended, the points come to the mean of the two, so the raw score is (Mentioned + Recommended) ÷ 2.

Engine usage weights

Each question weighs equally. Within it, each engine that answered weighs its share of usage, rescaled over the engines measured on that question. An engine your plan does not read, or one that returned nothing, drops out. These weights use the share of adults who use each tool in Pew Research’s Americans and AI 2026, published on 17 June 2026. Pew’s share who read AI summaries is split between AI Overviews and AI Mode using Google’s monthly-user figures, 2.5 billion and 1 billion. Pew does not ask about Perplexity, so its weight is estimated relative to Claude using Similarweb’s May 2026 web-visit shares. Adult usage, global monthly users and web visits measure different things. The latter two are estimates for the split and the missing tool. The same weights apply in every market. Updating them creates a new score version.

The Search side

Each result in Google’s top 10 takes a share of the clicks by its position: For one keyword’s results page:
  • A result on your own site counts toward Mentioned without reading the page. It never counts toward Recommended.
  • A result on a competitor’s site counts toward that competitor’s Mentioned.
  • Any other page counts toward Mentioned when Heralded’s reading of the page says it names you, and toward Recommended when the reading says it recommends you.
  • A page Heralded has not read, or could not read, is unknown. Its share stays in the total and counts as not naming you, so Mentioned and Recommended are a lower bound. The unknown share is what the top of the range adds.
  • For a competitor, a page whose reading did not cover that competitor is unknown too. A competitor added after the pages were read is unknown on them until they are read again, and one no reading covered is not measured.
  • A results page that could not be fetched is wholly unknown.
An unread page can lower a figure or widen its range. It can never raise it. Each keyword keeps its Mentioned and Recommended shares, with no score or band. One keyword is a results page and not a sample of keywords, so its range runs from the share Heralded read to that share plus the unknown share. Its shares are shown once a check has fetched its results. See How a keyword is measured. Each keyword is checked weekly in every country of its market, within the measurement limits. New keywords are checked on the next daily update. Each weekly report counts one observation per keyword, the mean of its checks that week. Checks that share page readings are not independent. Within each country, keywords are weighted by their latest known monthly search volume. In Europe, countries keep their fixed market shares regardless of how much search volume each country has. When a report combines Europe with single-country or Global topics, the two groups split the weight by keyword count. Keyword research keeps researched phrases with demand and preserves keywords you add or restore without volume. A keyword without positive volume keeps a small fixed weight. A country without measured keywords is left out when research has finished with no demand. A completed research entry in any language counts, even from an earlier month. If any tracked topic has never finished research there, the country keeps its weight as unknown. A country with measured keywords is measured by those rows. The report keeps each keyword’s volume and weight as they were that week, so a later volume update never changes a past figure. A period of several weekly reports weighs each report the same, each on its own weights. A keyword with no successful check stays in the total at zero, all of it unknown. Keywords that name your brand are left out. The range treats keywords as the sampled units. It uses the scatter of their values around the side’s value and a Student t interval, and its top adds the unknown share. The Search side is shown only when at most 20% of its weighted results are unknown and at least three keywords have a successful check. Otherwise it is not measured for that report. It is never recomputed over only the keywords that were read. On the Search side, change compares two periods only when the side is shown in both, and pairs the keywords checked in both. It uses the same country weighting over paired keywords, with the later report’s keyword weights for both periods, so a volume update never shows as movement. Countries without rows have nothing to pair. Its range widens by the paired keywords’ unknown share in each period. A leaderboard uses the same keyword weights in its Search measurements. Published leaderboard scores currently use the AI side while Search coverage is being checked.

All

All’s Heralded Score, Mentioned and Recommended equal the AI side’s. Their ranges, change and trends use AI answers too. This applies to every stored report at read time, so history stays comparable without a new break.

Why the score is calibrated

Raw scores bunch up at the bottom. Most brands are named in a small share of answers, so a raw score of 15 can already put a brand among the real contenders in its category. Shown as-is, the scale would make nearly every brand look like a failure and hide real progress. Heralded places the raw score on a calibrated scale, where each number marks a measured position: Between these points the scale is a straight line, with one addition at the bottom: a single named answer always reads at least 1. So 0 means that across the questions that don’t name you, no measured answer named you.

How the scale was measured

On 7 October 2026 Heralded refit the scale from 68 published category reads. The sample contains 7,921 answers and 3,208 brand and category observations. Brands are ranked within each category using the usage weights. Those runs read ChatGPT, Gemini, Google AI Overviews and Google AI Mode only. Their weights are rescaled over those four engines. The fit does not measure how adding Copilot, Claude or Perplexity changes a category’s scores.

The score bands

Next to the score, Heralded names the band it falls in. Five bands cover the scale: A Strong score sits between a typical top-three brand and a typical category leader. The band names a range of the score. It is not Recommended, the share of answers that recommend you. The API serves the 50 to 69 band as Recommended in heralded_score_band, score_bands and the band of a score, as it always has, and the names never change within v1. The topics read uses Strong for a topic’s band, its filter and its menu. The app, the emails and the share cards show it as Strong.

Score versions

A refit of the scale, or any change to how the score is computed, becomes a new score version. The current version is v9.1. It weights Search keywords within their countries and keeps unfinished countries unknown. New weekly reports with keyword checks and leaderboard runs with Search measurements use v9.1. Stored v9.0 reports keep their frozen weights and version, so comparisons break between v9.0 and v9.1. Reads without Search measurements stay on v8.0. A leaderboard’s AI history continues on v8.0. The AI side of stored Command Center reads, leaderboard months and issued Snapshot reports is recomputed from its frozen per-answer facts, so comparable history uses the same weights and scale. A read without those facts keeps its issued score and version. Emails and share images already sent keep the number they carried. Versioning covers the API. For a period made up of older reads without per-answer facts, the score is the mean of the measured issued weekly scores. It has no pooled range or paired change.

The range and the answer count

Next to the score is its 95% range. Open the score’s info button to see its range and how many answers the score was read from, counting the measured answers to questions that don’t name your brand. More answers mean a narrower range.

When there is no score

The score needs both Mentioned and Recommended. If either couldn’t be measured, the score is left empty. See When a figure is empty.