We Fact-Checked the SEO Stats Everyone Quotes

We traced 22 of the most widely quoted statistics in AI search and SEO back to their sources. Four of them were invented — no study exists, anywhere. Six more started as real research and got mangled somewhere between the study and the sales pitch. That's 45% of the numbers this industry runs on, wrong in some form. This post is the receipts.

What started it: one of those numbers was sitting in our own playbook. It fooled us too, and finding it changed how we write every page for our clients. Here's the whole thread, pulled.

Where did the fake statistic come from?

The number that started this claims content updated within the last 30 days earns an 82% citation rate on Perplexity, versus 37% for older content. If that were true, every business would need to re-touch every page every month, forever — which is exactly the kind of work agencies love to sell. So before we sold it, we went looking for the study.

There is no study. Every version of the claim we could find dead-ends: the blogs quoting it either cite a statistics roundup that doesn't contain the number, cite a large industry study that doesn't contain it either, or cite nothing at all — and two of the copies define the 37% differently, which tells you they weren't drawn from the same well. The closest thing to an origin we found was a Reddit thread.

Meanwhile, the real research says something close to the opposite. Ahrefs analyzed 16.9 million URLs cited by AI assistants and found the average cited page is about 2.9 years old. Fresher than what wins ordinary Google results, yes. A 30-day cliff where citations collapse? Nothing like that exists in any dataset we could verify. And the engine the fake stat names, Perplexity, actually cites the oldest content of the major assistants in that same Ahrefs study. We keep the fake number in our playbook now as a warning, not a fact.

How common is this?

Answer: of 22 widely repeated statistics we ran fully to ground, four were invented and six more were mangled. That's 45% wrong in some form. We didn't cherry-pick. We pulled the quantitative claims circulating in our corner of the industry, AI search and local SEO, and traced each one until we hit a primary study or ran out of road.

The four with no source behind them included a precise-looking table of "citation ranking factor weightings" that sums to a suspiciously clean 100%, and a claim that being in position 1 on Google gives you a 58% chance of an AI citation. We could not find either finding published anywhere, and neither could the pages quoting them. They are numbers shaped like research.

The six garbled ones bothered us more, because each one started as real work and got broken in transit:

A famous review study gets quoted as "businesses with 50+ reviews earn 35% more revenue." The actual Womply study found the 35% belongs to businesses that respond to reviews. There is no 50-review threshold anywhere in it.
"Every review star is worth 5 to 15% more revenue" traces to a Harvard Business School paper that found 5 to 9%, for independent restaurants, on Yelp. The finding is real. The version in circulation stretched the number and dropped every condition.
A widely shared claim that 86% of AI citations come from business listings misreads Yext's actual study of 6.8 million citations. The 86% covers listings plus the brand's own website, and the website is the single largest source at 44%. The real finding argues for investing in your own site. The circulated version argues for buying listing services. That inversion is not a small error.

Step 1

A real study publishes

A genuine finding, with conditions: sample, market, timeframe.

Step 2

It gets compressed

The number becomes a punchier sentence. The conditions fall off.

Step 3

Three blogs later

It means something its authors never found — and nobody clicks through.

How a real finding becomes a fake statistic · Greenlight Systems

That's the pattern: a real study gets compressed into a punchier sentence, the conditions fall off, the number drifts, and three blogs later it means something its authors never found. Almost no one clicks through. We know, because clicking through was all it took.

What does the real research actually say?

Answer: the verified findings are less dramatic and more useful. Once we threw out what wouldn't trace, what survived tells a consistent story, and every item below links to the primary source.

AI-cited content skews recent at a year scale, not a 30-day scale. About 75% of pages cited by AI assistants were updated within the past year. Updating your key pages a few times a year is the right dose. Panic-refreshing weekly is not.
Roughly 60% of the sources AI Overviews cite do not rank in Google's top 20 for the question asked, a figure corroborated by Ahrefs' independent analysis. Read that again: most of what AI recommends never won the old ranking game. The door is open to businesses that never had the budget to win at classic SEO.
Pages whose headings closely match the question asked get cited 41% of the time versus 29% for loose matches. Wording matters. Specifically, the customer's wording.
44.2% of ChatGPT's citations pull from the first 30% of a page. Answer first, explain after. (You may have noticed this post does exactly that.)

Notice what those four have in common. None of them reward tricks. They reward saying what your customers are actually asking, in the words your customers actually use, near the top of the page, and keeping it current.

What happens when you build on the verified findings instead?

Answer: page-one rankings and AI citations you can screenshot — from small local businesses, not enterprise budgets. We've been building on the verified version of this research since January. Here is what that looks like six months in, with real clients in our portfolio, and every number pulled from Google's own reporting this week.

A portable-restroom company in San Jose. Before writing their pages, we read 138 real public statements from real buyers in their category — reviews, forums, community threads — and matched the site's language to the words buyers actually use. Their site has been live in its current form since early July. Four weeks later: on the live results page this week, their homepage sits #1 for "portable restrooms in capitola," #1 for "portable restrooms in santa cruz," #2 in San Jose, and #3 for both the Gilroy and Bay Area phrasings — ahead of national brands on their home turf. On the Santa Cruz search, Google's AI Overview names our client first in its answer and cites their site as its #1 source. And the before-and-after is measurable: comparing June to July in Google's own Search Console, the number of searches where they appear on Google at all jumped 54% (231 to 356 distinct queries), searches where they hold a top-3 position with real volume grew 41% (29 to 41), and total search impressions more than doubled, up 115%. One month. In the same month, a real wedding lead arrived on their contact form tagged utm_source=chatgpt.com. Not a demo, not a test: a customer whose path to a local porta potty company ran through ChatGPT. That is what the 60%-of-citations-don't-rank-top-20 finding looks like when it lands in your favor.

A healthcare staffing agency. Their new site is built answer-first top to bottom: question headings in the words therapists actually search, the real number in the first breath of every page. The week it launched, Google's AI Overview was already citing them by name as the anchor source for the national travel-therapist pay answer. A staffing agency most people have never heard of, named by Google's AI as the source — because their page answers the question and the incumbents' pages dance around it.

A commercial real estate broker in Silicon Valley. This one is the next phase, already here. An inquiry landed in his inbox that read like it was written by an AI assistant acting on behalf of a company — because it was. He followed it anyway and scheduled the walkthrough. The company was real, they showed up, and the deal is now moving forward on one of his listings. Now the math on how hard that is. Silicon Valley office vacancy closed Q2 2026 at 16.0% on 137.8 million square feet of inventory — roughly 22 million square feet of empty office space competing for every active tenant — and vacancy is falling for the fourth straight quarter, which means more brokers fighting over fewer available tenants every month. In that market, a buyer's AI did the shortlisting, and out of all of it, it walked into an independent broker's listing — because his was the one it could actually read. AI agents aren't just answering questions anymore. They are picking who gets the meeting.

None of these companies has an enterprise budget. What they have is content written from this year's customer language and the verified research above, while most of their competitors are still optimizing to statistics that don't survive a click-through. We have been building on the verified findings since January. The first movers are collecting right now, and the window where it's this uncontested will not stay open.

How do you protect yourself?

Answer: ask one question of every number, and one question of every page. You don't need our tooling for either.

For every statistic in a proposal, a pitch deck, or a blog post you're about to trust: "who measured that, and can I read it?" A real finding has a source you can click, a sample size, and a date. "Studies show" is not a source. If the person quoting the number can't produce the study, treat the number as decoration. This applies to us too. If we ever hand you a figure and can't show you where it came from, hold us to it.

For your own website: "is this what my customers actually say?" Read your last twenty reviews and your competitors' last twenty. Read what people ask in your inbox and your contact forms. If your pages are named in industry vocabulary and your customers speak plainly, every one of those mismatches is a page AI will skip and a customer who won't feel understood. Your customers have already written the outline of your best content. Most businesses never read it.

If you want to see where you stand today, our free scorecard checks whether AI engines can see your business at all, and this guide explains what AI visibility means in plain terms. And if a marketer hands you an impressive statistic this week, you know what to ask.

We Fact-Checked the Marketing Statistics Everyone Quotes. Nearly Half Didn't Survive.