Lying With Statistics: When Good Numbers Tell a Misleading Business Story

Mark Twain is often credited with saying there are three kinds of lies: lies, damned lies, and statistics. Whether he actually said it or not, the line has stuck with us.

The bad news is that most of us are not very good at spotting lies. Research by Bella DePaulo, an authority on lying, found that in a review of 253 samples, in which participants were shown equal numbers of truths and lies, people correctly detected lying and deceit just 53% of the time. That’s barely better than a coin toss.

Which brings us to statistics, and why the lies they can tell are particularly insidious.

Numbers carry an air of authority. Put a percentage in a headline, an average in an earnings presentation, or a chart in an investor deck, and the argument suddenly looks objective. That’s exactly what makes statistics such a useful tool for anyone who wants to win an argument without actually being right.

The trick isn’t to make up the numbers (although that does still happen). More often, it’s choosing which number to show, what to compare them with, and what context to leave out.

Whether you’re reading a corporate earnings report or a pitch for a seemingly lucrative investment opportunity, the same techniques show up again and again. Here’s how to spot them.

Exhibit A: Mean versus Median

Let’s start with Statistics 101. When someone tells you that the “average” salary, return, sale price, or revenue was a certain amount, what does average actually mean?

There can be a significant difference between the mean (add all the values together and divide by the number of observations) and the median (line up all the values and take the one in the middle). To be clear, “mean” means average; they are synonyms.

This matters whenever a business is selling its investors, employees, customers, or buyers on what is “typical.” It also matters in countless other ways:

• Housing markets
  • Mean: Pulled way up by a few multi-million dollar mansions.
  • Median: Shows what a normal, everyday house actually costs.
• Company salaries
  • Mean: Inflated by huge executive bonuses and stock options.
  • Median: Reflects what a typical employee takes home.
• Recovery time from a particular surgery
  • Mean: Lengthened by rare patients who stay in the hospital for months.
  • Median: The typical number of days most people need to heal.

We’ve previously discussed this in the context of franchise disclosures. For that, you can read: Is Investing in a Franchise Worth It? A Buyer’s Guide to Due Diligence. When a franchisor reports the average (i.e., mean) unit revenue or profit that its stores generate, but a handful of unusually successful locations can pull that average upward.

That’s why the more revealing question for a prospective franchise buyer is what the median franchise unit earns after expenses. The median is the value sitting in the middle when all the numbers are lined up in order. For a franchise model with a wide mix of successful and ailing franchises, the median can provide a far more useful picture of what a typical franchise earns.

Whoever chooses which measure to show you is also choosing which story they want you to hear.

When an average sounds surprisingly good, ask for the median, the range, and the distribution behind it.

Exhibit B: The “Cherry-Picked” Comparative Timeframe

Suppose a company’s most recent quarterly revenue figures were $100 million, $50 million, and then $75 million. How did it perform in the latest quarter?

Its earnings report may state that revenue was up 50% in the latest quarter– an impressive figure. And yet, compared with two quarters ago, revenue is still down 25%. Both statements are completely accurate.

Financial performance is inherently cyclical, which makes the choice of baseline dates one of the easiest ways to manipulate a statistical story. If a quarter looks terrible compared with the same period last year, a company can emphasize quarter-over-quarter growth instead. If the sequential comparison looks weak, it can reach for year-over-year growth, a year-to-date figure, a multi-year comparison, or a pre-pandemic baseline.

A timely example of this comes from Anthropic. According to Bloomberg, the AI company is telling prospective investors that its Q2 revenue jumped at least 14-fold compared to the same period last year, from $787 million to more than $11.5 billion. That’s quite the headline, especially as the company is preparing for its mega-IPO later this year.

But Anthropic generated $4.73 billion in Q1 of 2026. Compare Q2 with Q1 instead, and revenue increased by roughly 143%. Still extraordinary, but a little less compared to “14-fold.”

There are other numbers worth keeping in mind, too. Bloomberg reports that Anthropic saw positive adjusted operating income in Q2, without providing a figure. So while the revenue numbers tell us Anthropic is growing extraordinarily quickly, they don’t tell us nearly as much about how profitable that growth is. That’s crucial in the context of frontier AI companies, where enormous computing and infrastructure costs have made the path to sustained profitability one of the biggest questions hanging over the industry.

All of which makes it especially important to pay attention to the numbers these companies are reporting. The statistic they choose to disclose, and the timeframe they use, shape the story they want investors to hear.

Exhibit C: Where Did the Good News Come From?

A company reports that profits have jumped 80%. Great news. But before you take action based on the report, there’s a question worth asking: Where did the extra profit actually come from?

A big increase in a headline number doesn’t necessarily mean the underlying business improved by the same amount. Profit can be boosted by investment gains, asset sales, tax benefits, accounting adjustments, or other items that have relatively little to do with selling more products or services (and even doing that can be mired with questions– look up the term “channel stuffing” if you don’t already know it).

Consider Alphabet’s Q2 2026 earnings. The company reported a record $112.1 billion in net income, up 298% from the prior year. At first glance, one might assume Google had just had one of the most extraordinarily profitable quarters in corporate history.

Its underlying business did have a very good quarter, but that wasn’t where most of the increase came from. Alphabet booked approximately $99 billion in realized and unrealized gains on equity securities, largely driven by the soaring valuations of its investments in Anthropic and SpaceX. Those investment gains added roughly $77.1 billion to net income after taxes and accounted for $6.26 of Alphabet’s $9.11 in earnings per share.

Alphabet’s actual operations were performing strongly in their own right: revenue increased 24% to $119.8 billion. But the headline numbers were driven largely by paper gains in investment, rather than by Google suddenly making three times as much money from ads, cloud services, and subscriptions.

To be clear, Alphabet disclosed all of this in its earnings report. Our point is about reading beyond the headline number to understand what actually drove it. A surge in net income caused largely by investment gains tells you something very different about a company than a surge driven by its core operations.

Exhibit D: Correlation ? Causation

You’ve probably heard the phrase “correlation does not equal causation.” In business, thinking it does is an especially easy mistake to make.

Consider online advertising. Many consumer companies run paid search ads for their products. Many will assume that search ads drive sales, since the numbers show that people who click on the ads go on to purchase items from the company.

But is there really a causation? An old eBay experiment says no.

Researchers switched off paid search advertising in some markets while leaving it running in others, and they found that much of the traffic simply moved to unpaid channels. In other words, many of the customers clicking the ads were going to shop on eBay anyway. The correlation between ad clicks and purchases had made the ads look much more effective than they actually were.

So how can you avoid making the same mistake? When you see a claim that X caused Y, ask a few questions:

• Does the relationship hold up over time and across different data? A one-off correlation could simply be a coincidence.
• Could something else explain both? In the eBay example, a customer’s existing intention to shop on eBay could explain both clicking the ad and making a purchase.
• Is there evidence from experiments or studies? Observing two things happening together is much weaker evidence than a controlled experiment.
• Is there a logical reason why X would cause Y? If nobody can explain the mechanism connecting the two, be wary of assuming causation.

Oftentimes in business, you may see claims that X is associated with Y. That claim can very well be true, but don’t assume one caused the other. Ask what else could explain the relationship, and what other evidence would actually prove it.

Exhibit E: Ask the Right People, Get the Right Answer

Companies love customer satisfaction scores. But an impressive score is only as reliable as the customers it represents.

Suppose a company announces that 90% of its customers are satisfied. Sounds impressive. But 90% of which customers? The ones who responded to the survey? What about those who ignored it, or former customers who were so unhappy they had already left?

That’s selection bias. If the people represented in your sample differ systematically from the wider group you’re trying to describe, the resulting statistic can be perfectly calculated and still give a misleading picture.

A company’s ‘net promoter score’ (NPS) offers a good example. The number purports to be a measure of customer loyalty. Customers are asked how likely they are to recommend the company on a scale of 0 to 10. Their answer groups them into either promoters (9–10), passives (7–8), or detractors (0–6). The NPS is then calculated as the percentage of detractors subtracted from the percentage of promoters, and can range from -100 to +100.

A company with a 20% customer response rate could calculate an impressive NPS of +50. But if the 80% who didn’t respond are considerably less enthusiastic, the true customer mix could produce an NPS of -22.

So, what sample size is actually statistically valid? According to SurveyMonkey, that number depends on factors including the size of the population you’re studying, your desired confidence level, and your margin of error. As a general rule, a proper sample of 400 respondents is sufficient to achieve a 95% confidence level with a margin of error of approximately ±5%, for any population size over 100,000 people.

A few caveats here: a sample of 400 respondents means 400 completed responses. Respondents should also be appropriately selected to limit selection bias. Depending on the type of survey, it may require using a probability sampling method, such as random sampling. For convenience, businesses will often choose less reliable sampling methods, such as sampling from customers who are subscribed to a mailing list, which naturally introduces more bias.

But even with sufficient numbers, there’s also deletion bias– where the customer’s answer is removed from the dataset. In a 2026 analysis of more than 2.6 million NPS responses, customer feedback platform Retently found that negative “detractors” on their platform were 2.4 times more likely to have their responses deleted than positive “promoters.” In e-commerce accounts, “detractors” were deleted at 28 times the rate of “promoters.”

Of course, legitimate data cleaning does happen. And Retently stressed that this behavior was concentrated among a small number of accounts rather than being widespread across its platform. But it illustrates that even an accurately calculated statistic can be misleading if the underlying dataset has been skewed.

Whenever you encounter a survey statistic, ask whether there’s a more concrete number you could be looking at instead. In the case of NPS, this could mean that actual sales, revenue, customer retention, repeat purchases, and cancellation rates can tell you far more about how customers really behave.

A Statistic Is Only Part of the Story

One point is worth reasserting: a statistic doesn’t have to be false to be misleading. The numbers, on their own, often do not lie. What matters is the context around it, and the choices that were made before it reached you.

So, whenever a number is being used to make a point, especially if it’s one that’s making the news headlines, get into the habit of asking questions. The ability to discern accurate takeaways from what the purveyor of a statistic being offered wants you to take away is one of the more useful skills you can bring to any negotiation, boardroom, due diligence process, purchase, or investment decision.