AI got the weekly report wrong? Check percent change, percentage points and overall conversion

Separate the absolute gap from change relative to the starting value

ONS guidance on percentages distinguishes differences in percentage points from relative percentage changes. The numbers here are our own arithmetic examples, not site traffic or product performance.

Suppose 20 of 100 participants pass in period A and 50 of 200 pass in B. Rates are 20% and 25%. You can report a five-percentage-point rate increase and 30 additional passing participants. They describe different measures and cannot replace each other.

The participant count changed. Growth in passing people and improvement in the pass rate therefore answer different questions. Define the event, denominator and period first. These two rows alone do not establish that a training program caused the change or that it is statistically significant.

Require raw inputs and formulas before prose

Do not provide only two percentages and ask for highlights. Supply counts and definitions, request a review table, then write prose after confirmation. This exposes changed denominators, reversed periods and count changes mislabelled as rate changes.

For PDF-derived inputs, confirm each number’s column first. Use the missing-versus-zero review workflow when values are absent. Do not fill unknown values with zero simply to make a calculation run.

Fictional data: A has 20 passing participants out of 100; B has 50 out of 200.
List counts, denominators, rates, the percentage-point gap and relative change using A as the baseline. Show each formula.
Distinguish counts from rates and percent from percentage points. Do not infer causes, significance or future effects.
For zero or missing baselines, or incompatible periods/definitions, explain why a direct comparison is unavailable.
Finish with two accurate summary sentences retaining periods and measure names.

Calculate a sample manually and inspect invalid cases

1. 1
   Copy counts and denominators into a separate table and calculate each rate. Ensure the denominator matches the event definition rather than another period or an annual total.
2. 2
   Subtract the earlier rate from the later rate for the percentage-point gap. Divide that gap by the earlier rate for relative change. Keep formulas before choosing display precision.
3. 3
   Check zero baselines and missing values. A zero starting value gives a zero denominator: report absolute change and explain that this relative-growth formula cannot be used.
4. 4
   Compare the AI’s words such as increase, decrease and double with the calculations. Each statement needs a formula and source row.

For spreadsheet formulas, use sample rows to audit AI formulas. Confirm the formula answers the intended measure, not merely that it executes.

Do not automatically average group percentages

Consider another original example: group X has 1/2 passing, and group Y has 9/90. Their rates are 50% and 10%, whose simple average is 30%. For the overall participant rate, combine counts instead: 10/92, approximately 10.87%. The question determines the appropriate aggregation.

An equal-weight group average can be appropriate if that is your intended measure, but label it distinctly from the overall individual rate. The AI should not choose this definition silently. Check overlapping groups, matching periods and duplicate denominators before combining data.

Keep unrounded calculations and round for display. If only rounded published rates are available, explain that recalculated relative changes may reflect rounding. Important analysis needs original data and suitable methods, not an assessment based on fluent wording.

Keep evidence another person can recalculate

Retain raw counts, denominator definitions, periods, formulas, rounding rules and source locations. Reverse a case: a move from 25% down to 20% is a 20% relative decrease, not 25%. This catches a mistaken fixed phrase reused for both directions.

This editorial audit works with many AI writing tools; it does not claim accuracy for a particular model. Check account access, retention and privacy settings separately. The responsible data owner should review decision-making reports.

How can channel mix change the overall direction?

This fictional acquisition case was designed by the editors. It is neither a test output from an AI model nor KitNelo traffic data. Both periods use qualified leads divided by visits with consistent scope and complete periods. Most baseline visits come from the higher-converting channel, while most current visits come from the lower-converting channel.

Baseline qualified leads are 180+10=190 from 900+100=1,000 visits, or 19%. Current leads are 30+135=165 from 100+900=1,000 visits, or 16.5%. Sum the numerator, sum the denominator, then divide. Do not average or add channel growth rates, or attribute the change to ads, a page or AI from four rows alone.

Both channels improve: what should the complete summary say?

Direct inquiry rises from 20% to 30% and Content reading from 10% to 15%. But Direct inquiry supplied 90% of baseline visits while Content reading supplies 90% of current visits. Changing channel weights makes the overall fall from 19% to 16.5% consistent with both improvements. Present channel performance alongside the visit mix rather than selecting one attractive growth figure. The pack includes these four records and overall reference answers.

An editor-designed faulty draft might say “Overall conversion improved significantly, proving the redesign worked.” Its overall direction is wrong, significance has not been tested and causation lacks a comparison design. A reference summary is: “Each period has 1,000 visits under the same definition. Qualified leads fall from 190 to 165; overall conversion falls from 19% to 16.5%, down 2.5 percentage points or approximately 13.16% relatively. Channel rates rise by 10 and 5 percentage points, but more visits shift to the lower-converting channel. These counts alone do not establish a cause.”

Baseline overall: =SUM(D2:D3)/SUM(C2:C3)
Current overall: =SUM(D4:D5)/SUM(C4:C5)
Percentage-point difference: =(SUM(D4:D5)/SUM(C4:C5)-SUM(D2:D3)/SUM(C2:C3))*100
Relative overall change: =(SUM(D4:D5)/SUM(C4:C5))/(SUM(D2:D3)/SUM(C2:C3))-1
Expected: 19%; 16.5%; -2.5 percentage points; approximately -13.16%

Leave room for an answer that cannot be calculated

For an ordinary positive baseline, percent change is (current−baseline)/baseline. A zero baseline makes the denominator zero: report the actual values and absolute increase. A missing baseline requires source repair. A negative baseline requires explaining the metric before presenting a percentage as healthy growth. Turning every error into 0% with IFERROR conflates no change with an undefined comparison.

For sales, Period Report can sum consistently defined amounts over two periods. This conversion example requires separately aggregating numerator and denominator; a single amount-change rate cannot substitute for it. Keep channel detail alongside the overall result so each sentence can be traced back to the four input records.