We Traced Every Viral GEO Statistic to Its Source. Most Don't Survive.

We traced the ten most-repeated statistics in AI search marketing back to their primary sources. Several are misquoted, one appears to be invented outright, and the true figures are usually less dramatic than the versions in circulation. Two of the wrong ones were on this site until last week. Here is what the evidence actually supports, and how to check the next number someone quotes at you.

A young discipline attracts confident numbers, because nobody has been around long enough to check them. GEO is at exactly that stage. Statistics get repeated until repetition becomes the evidence, and the sources they trace back to are frequently a blog post quoting a blog post.

This matters commercially, not just intellectually. If you are budgeting against a conversion multiple nobody can reproduce, you will over-invest and then conclude the channel failed. Below, each claim with what the primary source actually says.

1. “AI traffic converts 4-6x better”

Verdict: the range appears to be invented.

I went looking for the origin of “4 to 6 times” and could not find one. What exists is a spread of different measures that get conflated. Siege Media’s study, the largest independent one, screened over 120 properties and analysed the 78 with usable channel data between January and May 2026: the median AI-to-organic conversion ratio was 1.26x. Semrush published a 4.4x figure, but that is visitor value, not conversion rate. Similarweb published 7.1%, which is an absolute conversion rate and not a multiple of anything. Search Engine Land’s analysis of ChatGPT e-commerce referrals found them converting 31% better than non-branded organic. And a large academic study across hundreds of stores and millions of transactions found ChatGPT traffic converting worse than most channels.

So the honest position: usually better, contested, nothing like 4-6x, and one credible study says the opposite.

This one was on my site. Four of my own posts said AI traffic converts “roughly twice the rate,” sourced to enterprise data. That was defensible when written and is no longer the best available evidence. I have rewritten those posts, and the main one now carries a dated correction note at the top rather than a silent edit.

2. “Over 90% of AI Overview citations come from the top 10”

Verdict: outdated by a wide margin.

Ahrefs analysed 863,000 keyword result pages and around 4 million AI Overview URLs in March 2026. Restricting to organic blue links, 37.1% of citations came from the top 10. Its previous study put it near 76%.

Two caveats that most people repeating either number skip. Ahrefs improved its citation parsing between the two studies, so part of the drop is better measurement rather than changed behaviour. And figures above 90%, which appear in plenty of agency material, do not match either study.

This one was also on my site, in eight posts, cited to the very URL that now reports 37.1%. All eight are corrected. The underlying argument survives: SEO is still the foundation. What changed is that ranking is no longer the gate, and I was overstating how much it does for you.

3. “Gartner says AI will take 25% of search by 2026”

Verdict: misquote of a real forecast.

Gartner’s actual projection was that traditional search engine volume would drop 25%. That is a statement about one channel shrinking, not about AI capturing a quarter of all search. The difference matters: the real forecast is compatible with search simply becoming smaller, while the misquote implies a specific new channel you can go and capture. The misquoted version is now more common than the original.

4. “llms.txt improves AI visibility”

Verdict: not for Google Search, and no evidence elsewhere.

Google’s documentation states plainly that Search ignores these files, and that maintaining one “will neither harm nor help your site’s visibility or rankings.” An Ahrefs study of 137,210 domains found 97% of published llms.txt files received zero requests in May 2026, with SEO audit tools the heaviest requester of the remainder. I covered this in full in does llms.txt work, including why I removed it from my own service list.

Scope note, because the opposite overclaim is also circulating: Google’s statement covers Google Search. It says nothing about ChatGPT, Claude or Perplexity, and some coding agents do read these files.

5. “Fresh content gets cited 4.3x more”

Verdict: untraceable.

This figure appears in a great deal of GEO marketing and I could not trace it to any primary study. The real, sourced finding on freshness is more modest: cited content averaged around 1,064 days old against 1,432 for organic results, roughly 26% fresher. Freshness matters. It does not appear to matter 4.3 times.

6. “AI Overviews cut click-through by 34.5%”

Verdict: real number, wrong scope.

That figure came from an Ahrefs study measuring position-one desktop results only. Quoted as a general effect on all traffic, it overstates the case. The stronger evidence is a randomised controlled trial with 1,065 users and over 68,000 searches, which found a 39.8% reduction and, importantly, no difference in dwell time, which undercuts the argument that AI Overviews simply filter out low-quality clicks.

Here the correctly-sourced number is worse than the misquoted one. Rigour does not always favour the comfortable answer.

7. “Press releases drive AI citations”

Verdict: barely measurable.

An analysis of roughly 4 million citations found press releases accounted for 0.04%. If a PR package is being sold to you primarily as an AI-visibility tactic, that is the number to weigh it against.

8. “Your AI visibility score is X”

Verdict: mostly measuring noise.

SparkToro ran repeated prompts with around 600 volunteers across roughly 3,000 runs and found that fewer than 1% of repeated prompts returned an identical list of brands. Generative answers are probabilistic and personalised. A dashboard reporting your visibility as 34.7 this week and 31.2 last week, based on a handful of checks, is mostly reporting variance.

Tracked consistently across many prompts over months, these tools are genuinely useful for direction. Read as a weekly score, they are close to meaningless. More on this in being cited isn’t being recommended.

9. “43% of AI answers citing you recommend a competitor”

Verdict: real finding, over-generalised. Including by me, nearly.

The Ahrefs experiment behind this looked at pages promoting a brand-new conference: 43% of answers citing those pages never mentioned the conference. For an established product from the same publisher, the figure was 11%. It is a vendor self-experiment across two brands and 34 pages.

The mechanism is real and worth knowing. The 43% is not a law of nature, and I nearly published it as one before checking the sample.

10. “The GEO market is worth $1.1-1.5 billion, growing 40% a year”

Verdict: weak sourcing, treat as unusable.

These figures come from second-tier report mills. No tier-one analyst firm has sized this category. That does not mean the market isn’t growing, but the specific numbers should not appear in your business case. If you need to size the opportunity, work bottom-up from what services actually cost in your market.

How to check the next one

Three questions handle most of it.

Who published it, and what do they sell? Vendor data isn’t worthless, but a number that supports the publisher’s product deserves a look at the methodology. Several figures above are vendor self-experiments, which is fine when labelled and a problem when quoted as universal.

What was the sample and method, and is it disclosed? A widely-quoted Hong Kong AI tool ranking turns out to measure organic search traffic to those tools’ websites, a fact disclosed only in the Chinese version of the page. It is presented everywhere else as usage data.

Has it been repeated into truth? The 4-6x conversion claim and the 4.3x freshness claim both have the texture of numbers everyone assumes someone else verified. When you cannot find the study, that is usually because there isn’t one.

Nobody has clean answers in this field yet, and anyone who says otherwise is selling something. That includes me: the honest version of my pitch is that the mechanism is well-evidenced, the magnitudes are not, and the way to know your own numbers is to measure your own site. If you want to see what the AI engines actually say about your business rather than what a deck says they say, the free check does exactly that.

Frequently asked questions

Is it true that AI traffic converts 4-6x better? No. That range appears invented. The largest independent study found a median of 1.26x; vendor figures run higher; one academic analysis found AI referrals converting worse than most channels.

Do 90%+ of AI Overview citations come from the top 10? Not any more. Ahrefs found 37.1% in March 2026, against roughly 76% previously, with part of the fall attributable to improved measurement.

Did Gartner predict AI would take 25% of search? No. Gartner projected traditional search volume would drop 25%, which is a different claim with different implications.

Does llms.txt improve AI visibility? Not for Google Search, which says it ignores these files. Across 137,210 domains, 97% of published files received zero requests in a month.

How should I judge a GEO statistic? Ask who published it and what they sell, whether the sample and method are disclosed, and whether it has simply been repeated until it sounded true.