Skip to main content
Figures vary from week to week because the engines’ answers do. A drop of three points might be a real loss or a different set of replies. Heralded separates the two with a paired test, and marks a change real only when the test says the figure moved.

Observations in your week

Home’s timeline lists what Heralded observed during the week, such as an engine beginning to cite your site or a competitor changing a page. It states the observation without a diagnosis or a claim about its cause. Results keep a link to the work they describe and the states of that work’s verification. An event in the timeline and a real change in a figure are different facts. A new citation can appear in the timeline while the figure’s change stays steady. The test below decides whether the measured figure moved.

The period

A change compares the reports in a period with the reports before it. A period is the latest weekly report, the last four, the last twelve or all of them. Home always reads the latest weekly report, so its figures and score use the same change as the other pages and the API. The weekly report email, the topic rows and the manager Portfolio state the same change for the latest weekly report. A topic’s change uses that topic’s answers. The Period filter on the pages that have one and the API’s period choose another. A period counts reports, not days, so a paused week stretches it over more calendar weeks. The figures count each answer once across the selected reports. A change compares that period with the adjacent earlier window of the same length, using comparable reports and stopping at the first non-comparable report. For a single-country brand, Latest compares with the previous weekly report. A pooled Europe report covers fourteen days, so it compares with the fourteen days before those. Its first change appears once that earlier window holds answers from before the current window starts. An earlier report whose answers all fall inside the current window supplies no baseline, so the figure stays Calibrating. The initial report added during onboarding does not supply a full earlier window. A figure’s change is its shown value minus the baseline’s value. Both windows use the same estimator. Count deltas round to whole numbers; other figure deltas round to one decimal place. A move that rounds to zero has a steady badge. The badge is coloured only when the paired test calls the move real and the move has the same sign. Otherwise it is grey and carries its signed delta. A zero move reads No change since {date} when the comparison date is available, using the last day of the baseline in your time zone. A baseline with no value reads First report on a brand’s first report. Home, reports and engine rows use Calibrating when the brand has earlier reports but no baseline value. A newly added competitor keeps its new label. Without an adjacent earlier window, there is no change. The response states the baseline’s dates, report count and unpaired cells. Figures in the API describes these fields. Each weekly trend point shows that report’s own window, including answers another point also holds. Citation counts compare per answer, scaling the earlier count to the current window’s number of answers. A larger sample alone does not count as an improvement.

The paired test

Heralded compares the two sides question by question and engine by engine, using only the cells both sides measured. A cell measured on only one side drops out of the comparison instead of skewing it. Country and language also match, so a Romanian answer pairs with another Romanian answer in the same country even when the topic’s language policy changes. For each shared cell it takes the difference between the two sides and weights it the same way as all four figures. It sums the differences and puts a 95% interval around the total. If that interval excludes zero, the change is real. For the Heralded Score, the test runs on Mentioned and Recommended and the result is placed on the calibrated scale. The test decides only whether the move is real. It does not size the move, so the badge never contradicts the trend beside it, even when some measured cells are present in only one report. The interval reflects how much the questions’ own differences scatter. Fewer than three units leave the test without an interval, so the badge stays grey and keeps its signed delta. Each question’s difference averages its shared engines using their rescaled usage weights. A figure for one question uses its engines as the units. The test uses a Student t interval, which is wider with fewer units. It needs at least three questions, or three engines for a single question. Identical differences have no scatter, so their interval has zero width. This approach assumes the units’ differences are independent. Its calibration is provisional while Heralded checks how answers move together across weeks. Share of voice keeps its own pooled-slot interval. On the Search side the units are keywords, and All uses the AI changes. Reputation weighs Understood and Praised across their AI and web parts. The web part’s units are pages, paired between the periods that read them.

When the engines change

Adding, pausing or removing an engine changes what a figure pools, so the chart can move for a reason that is not the brand’s. Heralded keeps calling the change and sizes it on the engines both periods asked. It pools each period over those engines the same way as a trend point, and the change is the difference. The paired test already pairs only the cells both periods measured, and it decides whether the badge is coloured. A period of several reports counts an engine as asked when any of its reports asked it. The figure, its chart and its engine boxes keep every engine asked. A new engine shows its own value from its first report, and its change reads New. The trend chart marks the report where the engines changed, such as “Copilot added”. When the figure includes an engine the size leaves out, the change pill has an info button that names the engines compared. The API serves them as compared_engines, and null when the size is the move the chart draws. An engine whose calls were all skipped or paused is not asked, so it leaves the comparison instead of reading as a fall. The Search side asks Google only, so engines do not touch it.

Reading a change

A change pill is coloured only for a real move. New baseline (not_comparable in the API) means there is no valid comparison, so the move is not called. A headline omits the change clause when the figure starts a new baseline. At 95%, about one figure in twenty that didn’t really move will still show as real. One real change is a signal. The same move over several reports is a trend.

When there is no change

A figure has no change when:
  • no comparable weekly report comes before the period;
  • Heralded changed a market definition, the score version or the way answers are read;
  • the brand moved to a new domain. Change returns after two reports on the new domain.
The trend chart labels changes such as “Copilot added” or “Romanian answers are now asked in English and Romanian”. Changing a topic’s languages within the same country keeps the comparison over the cells measured in both periods. Adding or removing an engine keeps it too, on the engines both periods asked. The same mix of languages is never a change, however it is stored: a rotation of English, Romanian, Romanian and English one third, Romanian two thirds are one policy. A different mix is a labelled change. A change of country or pooled market definition breaks comparison. Rotating through countries and languages under the same policy does not. Global’s English-only series continues unchanged. Within one weekly report, each topic counts only the answers asked on the report’s own market. Daily answers asked before the market changed sit out of that report, so a market change starts a new baseline and blanks the change, while the figures stay. Sparklines, rank charts and Visibility’s comparison chart join measured points across a new baseline with a dashed line. Hover over the join to see “Measured with a different set of questions”. Your line within one baseline stays solid. Questions are written for who the brand sells to, businesses or consumers, so changing it starts a new baseline on every figure, whether its questions were redrafted or kept. Turning Ecommerce on or off changes no question, so every series continues. The API names the two settings sells_to and ecommerce, and only a change of sells_to starts a new baseline. Where both reads record how many passes they used, a change in that count also breaks comparison. A read with no recorded pass count does not break the series on that basis. Your first report that reads Understood part by part starts a new measure on Understood alone. Understood and its claims show no change across it, and its trend point carries understood_measure in its break. Every other figure’s series continues. When no cells can be paired, low and high are null and real is false; delta still states the move between the two values. Figures in the API describes its states. Adding or archiving a question or a keyword, or adding a topic, doesn’t break the comparison. Cells and keywords with nothing to pair with sit out. A change to the keywords a report checked is recorded in its instrument key as keyword_set_hash and doesn’t break the series. A Search change compares with the baseline report’s drawn value: the figure the trend draws for that report, over the same keywords, topic, audience and market as the current value, so a filter never moves the baseline. Search reads the keywords you track today across every report, as AI figures read the questions you track today, so archiving a keyword removes it from past points too. Each check keeps the weight and volume it froze; only the population follows today’s tracking. A report whose tracked keywords fall short of the coverage rule draws as not measured. If the trend draws the baseline as not measured and the figure is measured now, the change reads First report, with no delta and no date. Otherwise it states the move between the two drawn points and is dated from the baseline report. On All or AI, a change between the AI part and the blend remains a new baseline.