How I Verify Statistics Before I Cite Them

I’ve become much more suspicious of statistics that look current because the article quoting them is new. Before a number makes it into anything I publish, I trace it back to the organization that collected the data.

Many a time, I have spent an embarrassing amount of time chasing statistics around the internet. You find a number that fits what you’re researching, the article looks recent, and there is a source attached to the claim, so you assume you’re probably one click away from the original research. Then you click.

The “source” is another article.

You open that one, find its citation, and click again. Sometimes there is another article after that. And another. By the time you eventually reach the organization that actually produced the number, a few different things may have happened:

• the statistic is several years older than the page you found it on;
• the original study measured a much narrower group than the article now suggests;
• the wording has shifted from something like “tried” to “regularly uses”;
• the original methodology adds context that changes how impressive the number looks; or
• you get to the supposed source and the statistic is simply not there.

These days, finding a number is just the lead. Then I open the original research, check what it actually measured, and decide whether the figure belongs in the article I’m writing.

Original graphic. How context can drift as a statistic moves from the original research through secondary coverage.

A number can stay the same while its meaning changes

A lot of statistics online are probably not completely made up. What I run into more often is context getting shaved off as a number moves from one page to another.

Imagine an original survey finds that 58% of 700 marketing executives at large U.S. companies use a particular technology. Another article summarizes that as “58% of marketers.” Someone else cites that article and turns it into “58% of businesses.” Another roundup repeats the claim. Eventually the number appears in a page titled “2026 statistics,” even though the original data may be years old.

The percentage has survived perfectly. The meaning hasn’t.

That is also why I’m careful with phrases that sound close enough to be interchangeable. Depending on the study, any of these could describe different things:

• people who have tried an AI tool;
• people who use one at least occasionally;
• people who use one every day;
• people who use AI specifically for search;
• people who say AI has replaced some of their traditional search behavior.

If the original research measured the first one, I don’t want my article quietly turning it into the fifth one because that makes for a better headline.

I check the publication date and the data date separately

A 2026 report can use a survey fielded in late 2025. Both dates belong in my notes: when the data was collected and when the analysis was published.

A good example is McKinsey’s Global Tech Agenda 2026. McKinsey published it on February 9, 2026, but the methodology says the online survey was in the field from September 29 to November 10, 2025. The survey had 632 respondents across 69 countries and 24 industries and subindustries.

McKinsey Global Tech Agenda 2026

Published February 9, 2026

Fieldwork September 29–November 10, 2025

Sample 632 respondents

Scope 69 countries, 24 industries/subindustries

Source: McKinsey & Company, Global Tech Agenda 2026. The report was published in 2026; the survey responses were collected in 2025.

That McKinsey report is a useful example of the date distinction. Its publication year is 2026, but the responses came from 2025. For my current-year statistics roundups, I look for studies that actually collected their data in 2026.

It sounds like a tiny distinction until you start comparing studies, tracking changes over time, or building an annual reference page where recency is part of the promise.

For the statistics pages I’m building now, my rule is stricter

We’re heading toward 2027, and I want a reader opening my 2026 statistics pages to find evidence collected in 2026. A newer headline on an older survey doesn’t help someone looking for the year’s actual changes.

An older study can still make a useful historical comparison. I’ll label its fieldwork year and keep it separate from the current-year findings.

A real example: one statistic, three different dates

The Editorial Freelancers Association’s 2026 rate chart is a good example of why I keep these dates separate. The chart was released in 2026. The survey ran from November 2025 into January 2026. Respondents described the rates they charged during the 2025 calendar year.

If I’m using the chart to discuss freelance writing costs, I can describe it as the EFA’s 2026 published rate reference, based on work billed in 2025. Those details tell me exactly what period the prices represent. For a research page limited to activity that happened in 2026, however, I’d look for figures measuring that year’s activity directly.

A second example: how one word changes a finding

Clutch and Conductor surveyed 459 content marketers in January 2026. They found that 87% planned to increase their content budgets that year. The words “planned to” are part of the finding. They tell me the survey captured a budget intention at the beginning of the year, which is what I’ll put in the published heading and explanation.

That small source record saves me from chasing the same statistic again when I refresh an article next year. It also makes writing the headline easier, because the population, period and measurement are already in front of me.

What I actually mean by a primary source

I start with the organization that collected the data.

So if Pew Research Center ran the survey, I want the Pew report. If a company analyzed data from its own platform, I want that original analysis. If an academic team conducted the study, I want the paper. If a government agency collected the dataset, I want the agency’s release or database rather than an article summarizing what the agency supposedly found.

Secondary sources are still useful for discovery. Sometimes a great article is how I learn that a study exists in the first place. I just don’t want my trail to end there.

My quick primary-source test

Did this organization produce the evidence behind the number, or are they telling me what somebody else found?

If they produced it, I start checking the methodology. If they are quoting somebody else, I follow the citation and keep going.

Company research can still be primary research

I also don’t think “primary source” automatically means government or academia. A company analyzing a large dataset it actually owns can be the primary source for that analysis. The same goes for a research firm conducting its own survey or a software platform publishing aggregate data from its product.

I check the sample, methods and the company’s stake in the findings. A SaaS platform analyzing its own customers has useful first-party evidence, especially when it explains what it measured and how.

My verification process before a statistic makes the cut

Once I find a promising number, I basically start interrogating it. I don’t need every article to turn into an academic literature review, but there are a few things I want to know before I’m comfortable putting the statistic on a page someone else might cite.

• Who actually produced the number? I keep following citations until I reach the original report, dataset, study, survey, filing, research page, or official release.
• When was it published? For the current statistics library, that means 2026.
• When was the data collected? I look for fieldwork dates, measurement windows, crawl dates, query dates, or whatever period makes sense for the research.
• Who or what was measured? Consumers, B2B buyers, marketers, executives, websites, queries, sessions, domains, and citations are all different populations.
• How large was the dataset? Sample size does not decide quality by itself, but it tells me a lot about what kind of claim the study can support.
• What exactly does the percentage measure? This is usually where superficially similar statistics stop being comparable.
• What does the methodology say that the headline doesn’t? Weighting, geography, exclusions, definitions and study design can change how I interpret the result.

Original graphic. My seven-step check before I cite a statistic.

Scope is where I think a lot of stats posts go wrong

A dataset of millions of consumer searches might explain broad search behavior. To understand how B2B SaaS buyers compare vendors, I’d look for research involving software buyers, sales conversations or the relevant search queries.

This is something I’m watching closely in the AI search and generative engine optimization research I’m working on now. There are plenty of interesting numbers about ChatGPT, Google AI Overviews, Perplexity, AI referral traffic and AI citations. I label the sector and respondent base of each study so readers can identify which findings cover B2B SaaS specifically.

If the source studied the broader web, I’d rather say that clearly and then explain why the finding might matter to B2B SaaS than quietly change the population because the narrower framing fits my article better.

And similar-sounding metrics can still be completely different

This becomes especially messy with AI search because there are several layers that people sometimes collapse into one conversation:

• adoption: whether people or organizations use the technology at all;
• usage frequency: how often they use it;
• visibility: whether a brand or domain appears in an AI-generated result;
• citations: whether the system explicitly cites a source;
• referral traffic: whether somebody actually clicks through to the site; and
• conversion: what those visitors do after they arrive.

You can have strong growth in one layer and barely any movement in another. That is why I don’t want to put two percentages beside each other simply because both contain the words “AI search.”

Sometimes the correct result of the research is deleting the number

This is probably the biggest change in how I approach statistics content. I used to think a better statistics page meant collecting more statistics. Now I care much more about how many of those numbers I can actually defend.

I’ll usually drop a statistic when:

• I cannot locate the original research;
• the supposed source exists, but the statistic is not actually there;
• the underlying data is too old for the page I’m building;
• I cannot work out who or what the number describes;
• a key definition has disappeared somewhere along the citation chain;
• the methodology makes the number unsuitable for the claim I wanted to make; or
• dozens of sites repeat the statistic, but every citation eventually loops back to another secondary page.

It’s remarkable how credible a number can look once enough websites repeat it. I can open five citations and end up following the same chain of secondhand articles, with no original report in sight.

For a substantial statistics reference, I’m targeting at least 30 useful findings, each traceable to the organization that collected the evidence. I’d rather invest time finding another credible study than pad the total with recycled claims.

The analysis gets better once I know what I’m actually comparing

This is the part I care about beyond just having clean citations. Verifying the source, scope and methodology lets me compare the evidence without accidentally comparing different things.

Say one 2026 dataset shows AI referral traffic growing by several hundred percent, while another shows AI referrals still account for a tiny share of total website traffic. I don’t immediately see a contradiction there. Something can grow very quickly from a tiny base and still remain small in absolute terms.

The same thing applies to adoption and usage. One study could show a lot of people have experimented with AI search while another shows much lower daily usage. Again, both can be true. Trying something and building it into your normal behavior are different measurements.

I prefer to analyze clusters of evidence

For the actual statistics pages, I want the numbers to stay easy to find. Somebody who landed on the page because they searched for one statistic should not have to dig through a mini essay after every bullet.

So the structure I’m moving toward is fairly simple:

• put the strongest statistics near the top;
• give readers a master table they can scan quickly;
• group related statistics into useful sections;
• then step back after each cluster and look at what the group of findings actually suggests.

That is where I can start asking the questions I actually find interesting: Do independent datasets point in the same direction? Are two studies disagreeing because they measured different populations? Is the giant growth percentage mostly a base-effect story? Does visibility turn into citations? Do citations turn into clicks? Do clicks turn into anything useful?

And if the evidence cannot answer one of those questions yet, saying that is useful too.

There is still a line between inference and making up a story

If 22% of respondents did something, I can say roughly one in five did it. I can say about 78% did not. If another genuinely comparable group is at 44%, I can quantify the difference between them.

If the remaining respondents weren’t asked why they made a different choice, I leave their motives out of the analysis. I can compare the reported responses and look for another study that examines the reasons.

When two measures move together, I look for studies that explain the relationship and what else may influence it. That produces more interesting research questions than an unsupported explanation.

More useful statistics, with the original evidence attached

I still want these pages to be very easy to use. If someone Googles a statistic, lands on one of my pages, and only needs the number and the source, they should be able to get both quickly. The table should work. The attribution should be obvious. The original research should be one click away.

Behind the simple table, I keep the research notes: source link, fieldwork date, respondent group, exact metric and anything needed to interpret the figure.

Sometimes that process gives me a better statistic. Sometimes it gives me a more interesting comparison. And sometimes, after opening eight tabs and spending 20 minutes trying to verify one sentence, it tells me to delete the damn number.

That counts as useful research too.

For examples of what this process looks like in published reference work, see my 2026 AI search and GEO statistics, SEO traffic and click-through benchmarks, and content marketing pricing research. Each page separates the reported figures from my interpretation and includes the original source links.