The most expensive number in a PR report is usually correct. It is the sentence around it that is wrong. 

Take share of voice going from 12 percent to 15 percent. Call that a 25 percent increase and you are not lying, exactly. The CMO still hears growth that never happened. It moved three points. One subtraction, two ways to write it up, and only one of them describes the quarter. 

Five arithmetic errors account for most of what I send back when I review a monthly or quarterly report. None of them require statistics. All of them survive review because the math is easy enough that nobody checks it, and because the wrong version almost always flatters the campaign. 

Here they are with the numbers worked out, plus the reason AI-assisted reporting is making all five more common rather than less. 

1. Percentage points are not percent

This is the one that reaches the boardroom most often. A percentage that moves from 12 to 15 has moved three percentage points. Expressed as relative change, it moved 25 percent. Both are accurate. They are answers to different questions. 

The asymmetry is what catches people. A slide back down from 15 to 12 is the same three points, but only a 20 percent relative decline, because the base changed. Report the relative figure in both directions and your chart appears to gain more than it loses while the underlying position is flat. 

Anything already expressed as a percentage is exposed to this: share of voice, sentiment splits, message pull-through, quote inclusion rate. 

The fix Use “points” for the difference between two percentages and “percent” for relative change, and never let one report use “percent” for both. If a stakeholder only remembers one number, make it the one in points. 

2. Averaging monthly share of voice

A quarterly share of voice figure is often produced by averaging three monthly percentages. That gives a quiet month the same weight as a loud one, and category volume is rarely stable across a quarter. 

Here is the shape of a quarter I see often. January brings 5 brand mentions against 50 in the category, so 10 percent. February is quiet, 12 out of 40, which lands at 30 percent. Then March goes wide: 30 mentions out of 500, or 6 percent. 

Month	Brand mentions	Category total	Monthly SOV
January	5	50	10.0%
February	12	40	30.0%
March	30	500	6.0%
Quarter (pooled)	47	590	7.97%

Average the three monthly percentages and you get 15.33 percent. Pool the quarter properly, 47 mentions against 590, and you get 7.97 percent. The reported number is nearly double the real one. 

February did the damage. Forty category mentions is a dead month, so a handful of hits produced a spike that means very little, and the unweighted average treated that spike as equal to March, where 500 mentions were in play. 

The fix Percentages are not quantities you can average unless the denominators match. Add the numerators, add the denominators, divide once. If your dashboard exports monthly percentages only, ask for the raw counts. 

3. “Up 200 percent” on a base of two

Coverage went from 2 articles to 6. That is a 200 percent increase, and it is four articles. 

Elsewhere in the same report, a program that went from 40 placements to 48 shows up as 20 percent growth. It produced eight more articles, twice the absolute gain, and it looks like the weaker line on the chart. 

Small bases make percentages behave badly, and the effect runs the other way too. Lose one placement from a base of two and you are down 50 percent, which will read as a collapse in a program that never had scale to begin with. 

The fix Print the absolute change next to every relative one. Set a floor below which the report shows counts only. There is no industry standard for that floor, so pick one, write it into the template, and apply it to good news and bad news alike. 

4. Summed audiences are not reach

Three outlets cover the launch. Each reports around 2 million monthly unique visitors. The report says the campaign reached 6 million people. 

It did not. Anyone who reads two of those outlets is counted twice in that sum. Put any plausible overlap on it and the gap shows up fast. 

The arithmetic Say 300,000 people overlap between each pair of outlets, which is 15 percent of one outlet’s audience, and 50,000 read all three. Unique reach is 6,000,000 minus 900,000 plus 50,000, or 5,150,000. The summed figure is 850,000 people too generous. 

Those overlap numbers are illustrative, and that is exactly the problem. Without panel data you cannot know the true overlap, which is the reason the sum is not a number you should present as reach at all. 

Impressions are different and it is worth being precise about why. An impression is an event, duplication is built into the definition, and impressions can legitimately be added. Unique reach is a set of people. Sets do not add. 

The same table holds a related trap, and it is the one that catches me most often. A single outlet can be so much bigger than the rest that the average stops describing anything real. 

The outlier effect Five outlets at 20,000, 35,000, 50,000, 60,000 and 12 million monthly visitors. Mean audience: 2,433,000. Median: 50,000. Report the average outlet size and you have described the outlier, not the campaign. 

The fix Label summed figures as potential impressions and never as reach. When a distribution is skewed by one large outlet, report the median alongside the mean, or drop the average and list the top outlets by name. 

5. A multiplier applied to an invalid number

Advertising value equivalency has been ruled out by the industry’s own measurement body for years. The Barcelona Principles 4.0 put it plainly in principle five: AMEC’s position is that invalid measures such as AVEs should not be used, and that contribution should be measured by outcome and impact instead. 

It survives anyway, usually rebadged as earned media value with a multiplier attached for the credibility of editorial coverage. That multiplier is where the arithmetic goes from questionable to indefensible. A $48,000 base at a 3x multiplier reports $144,000. A 20 percent error in the underlying rate card, which is common, turns into $28,800 of invented value once it is tripled. 

The multiplier is also almost never sourced. In most reports I have inherited, nobody in the room could say where the 3 came from. 

The fix Replace it with outcome measures tied to the objective you set at the brief. If a parent company mandates an EMV line, report the base rate and the multiplier as separate visible fields so the assumption can be challenged rather than buried in one dollar figure. 

Why AI assistants make these five worse

Most teams now hand some part of the reporting cycle to a chatbot. Paste the export, ask for a summary, get a paragraph back. The trouble is that this workflow lands squarely on the one weakness these models have not shaken. 

Researchers at NYU Abu Dhabi built GSM-Ranges, a generator that takes standard math word problems and systematically scales the numbers up, then measures where models break. Two findings matter for anyone building a report. 

First, arithmetic accuracy falls sharply once a calculation is embedded in a word problem rather than presented on its own. Models that handle a bare division reliably start getting it wrong when the same division is wrapped in a sentence about a situation. 

Second, error rates climb as the numbers get bigger. Across their perturbation levels the gap reached 14 absolute percentage points for the weakest model tested. 

A PR report hits both conditions at once. Impressions run into the millions, and every calculation sits inside a sentence about a campaign. The output comes back fluent and unhedged too, so a wrong figure reads exactly like a right one. 

So split the two jobs. Let the assistant draft prose if that saves you time, but lift every calculation out of the paragraph and check it as bare arithmetic. I put the pooled share of voice and any reach figure through a step-by-step math solver and read the working rather than the final answer, since a number can match and still have come from the wrong operation. No account, no cost, which matters when the alternative is rebuilding a spreadsheet nobody wants to touch. 

Treat that as a second opinion rather than an authority. It is another model, its answers are not independently verified, and nothing you check is saved anywhere for you to revisit. Two systems agreeing on the steps is meaningfully better evidence than one confident paragraph, and that is the whole benefit. 

The words around the numbers need a separate check

The same drafting shortcut carries a media relations cost that has nothing to do with arithmetic. 

One PR services company asked journalists about this directly in its State of the Media report, and the answer was not subtle. Fifty-three percent object to receiving AI-generated pitches and press releases, on grounds of accuracy and personalization. 

The sample 1,899 journalists across North America, EMEA and APAC, surveyed in January and February 2026, working in digital, print, broadcast and emerging media. 

The same journalists are leaning on PR material more than ever. Two thirds of them use it for story ideas. Nearly three quarters said under a quarter of the pitches reaching their inbox are relevant at all. 

Read those together. Journalists are more dependent on your material than ever and more hostile to the version a model wrote. Generic phrasing is not a style complaint in that context, it is a deliverability problem. 

Before a draft goes to a reporter, it is worth running it through a detector to see whether it reads as machine-written. TextToHuman gives a free AI probability score without a signup, and there are several others that do the same job. 

One caution before you lean on that score. Detectors flag plenty of writing that people actually wrote, and they do it most on formal, structured copy. That is most of what a PR team produces. So read a high score as a reason to reread, then fix the draft the only way that works. Put in the detail no model could invent. A figure from your own data. The reason this reporter’s last piece makes them the right person to send it to. 

Before the report goes out 

• Every percentage change is labeled as points or percent, consistently in both directions 
• Quarterly share of voice is pooled from raw counts, not averaged from monthly percentages 
• Absolute change appears next to every relative change 
• Summed audience figures are labeled potential impressions, never reach 
• Averages are checked against the median wherever one outlet dominates 
• Any EMV line shows the base rate and the multiplier as separate fields 
• Each calculation has been verified outside the paragraph it appears in 
• Every number in the narrative traces back to a row in the source export 

The bottom line

None of this is advanced math, which is precisely why it slips through. A report full of correct arithmetic described in the wrong units will pass every review in the building and still mislead the person it was built for. 

The five errors above share a useful property. Every one of them, left uncorrected, makes the campaign look better than it was. That is worth remembering the next time a quarterly number comes in higher than expected and nobody feels the urge to check it.