Warning: This is very technical, but extremely important to everyone who has a website.
Hint: If you don’t want to read through all the data, copy the article and feed it into your favorite AI to give you a summary of the most important points. You can even ask it to give it to you in easy to read bullet points. But come on back for Part 2 & 3. Not as technical but very important!
Google’s AI Overviews cut organic click-through rates by 61%.
You’ve seen that number. It turns up in every roundup of AI SEO statistics published this year.
So I went looking for where it came from.
That one turned out to be real, and I’ll show you the study.
But while I was checking it, I started pulling on the other numbers sitting right next to it.
- The 317% boost from adding images and video.
- The 3.2x from ticking one box in an SEO plugin.
The one that supposedly proves the single biggest factor in AI search is whether a paragraph makes sense on its own.
Two clicks in, I stopped being able to check anything at all.
Because nearly every one of those numbers traces back to a single web page. That page contradicts itself repeatedly.
Nobody’s name is on it.
Its biggest number came from asking an AI…
…and the website it lived on no longer exists.
Or it never did exist ever!
Author that has “no name”, used an AI Hallucination which was then copied and used by others!
Here’s what survived, what didn’t, and the one-minute method I used. You can run it on any article you read.
Including this one — I dare you.
What generative engine optimization actually means
**Generative engine optimization (GEO) is the practice of getting your content quoted inside an AI-generated answer, instead of ranking in the list of links underneath it.**
That’s the whole shift, and it’s a big one. Traditional SEO fights for a position on the page. GEO fights to be the thing the AI actually says.
One quick bit of housekeeping, because almost nobody separates these properly:
- AI Overviews — the summary box that appears above your normal Google results
- AI Mode — a separate conversational tab where you chat with search like a chatbot
Nearly all the research below measures AI Overviews. Most advice online is written as though it covers both. Tuck that away… it matters later.
Another thing worth noting is the terminology. SEO, GEO, AEO and AIO. These all grew out of the AI era. If this keeps up we’re going to need a special AI dictionary. 🤭
The numbers that hold up
Three of the most-quoted numbers do survive checking: the 61% traffic drop, the 35% citation bonus, and the 40% visibility boost.
So let’s start with the good news — I’m not here to tell you the whole field is nonsense. Some of this research is genuinely solid.
Here’s where each number actually comes from, and what the write-ups leave out.
The 61% is real. It comes from Seer Interactive, who tracked 3,119 search terms across 42 client organizations and 25.1 million impressions, from June 2024 to September 2025.
Click-through on searches with an AI Overview fell from 1.76% to 0.61%.
One caveat that matters, and almost nobody repeats it. Seer deliberately studied “informational and educational queries, most vulnerable to AIOs.” They picked the queries most likely to get hit — sensibly, because that’s what they were investigating.
So 61% is not “what happens to your traffic.” It’s what happens to the most exposed slice of it. If your posts answer straightforward questions, that’s you. If they don’t, your drop is smaller.
And it’s not the only number out there. Search Engine Land rounded up several:
- Ahrefs — a 34.5% drop for the top result, across 300,000 keywords. That was April 2025; their 2026 re-run puts it at 58%
- Pew Research — tracked 900 real people’s actual browsing for a month. They clicked a result 8% of the time when an AI summary appeared, 15% when it didn’t. Only 1% clicked a link inside the summary itself
- Amsive — 700,000 keywords across ten sites, narrowed to the 10,000 that actually triggered an AI Overview. Average drop: 15.49%. But branded searches went the other way — up 18.68%
That last one deserves a second look. When people search for you by name, AI Overviews help. It’s when they search for a topic that you lose the click.
Which is an argument for becoming a brand people look for, rather than a page they stumble onto.
So the honest answer is a range, roughly 15% to 60%, depending what you measure and which queries you look at. Anyone quoting one figure as the figure is just picking their favorite.
The strongest of the lot is different in kind. Two academics — Saharsh Agarwal at the Indian School of Business and Ananya Sen at Carnegie Mellon — built a Chrome extension and randomly assigned 1,065 people to see AI Overviews or not, for two weeks each.
They registered the study design publicly before collecting any data, so they couldn’t quietly move the goalposts afterward.
Their finding: a 38% drop in clicks through to publishers.
That one carries more weight than everything above it, because a randomized experiment can show cause. Every other study here shows two things happening at the same time and leaves you to guess at the link.
The famous “40% boost” is real too. It comes from an actual peer-reviewed paper — GEO: Generative Engine Optimization, presented at a major research conference in 2024. The researchers tested around 10,000 searches and found three edits that reliably improved visibility by 30–40%:
- Add statistics instead of vague description
- Add quotes from credible named people
- Cite your sources and link to them
Notice how boring and old-fashioned those are. Hold that thought.
The sentence nobody prints
Here’s the bit of the Seer study that keeps getting left out.
Everyone quotes “cited pages get 35% more clicks!” Perfectly true. Here are all three numbers from that same table:
| Situation | Click-through rate |
| No AI Overview at all | 1.45% |
| AI Overview, you’re cited | 0.70% |
| AI Overview, you’re not cited | 0.52% |
Being cited is 35% better than not being cited. It is also less than half of what you were getting before AI Overviews existed.
Same study. Same table. Guess which number makes it into the marketing?
You don’t have to rank to get cited — the part nobody says out loud
Almost a third of the pages Google’s AI cites don’t rank in the top hundred results at all.
Now this is the genuinely encouraging bit — and it’s the best-sourced number in this whole article. Almost nobody mentions it.
Ahrefs checked 863,000 keywords and 4 million AI Overview links to see where cited pages actually rank.
| Where the cited page ranks | Share of citations |
| Top 10 | 38% — down from 76% in July 2025 |
| Positions 11–100 | 31.2% |
| Beyond position 100 | 31.0% |
Read that last row again.
Almost a third of AI citations go to pages that don’t rank in the top hundred. Pages that, in old-fashioned search, may as well not have existed.
One honest note on that 76%, since I’ve spent this whole article complaining about people skipping the fine print.
Ahrefs say they “improved our parsing methodology” between the two studies, so they can now spot more citations than before. They don’t say how much of the change is better measurement and how much is a real shift.
So take the direction as solid and the exact size with a pinch of salt. It’s still the best data anyone has on this.
Why? Something called query fan-out. Google quietly takes your question, splits it into a cluster of smaller ones, runs them all, and cites whoever answers the pieces best. Your overall ranking stops being the bouncer at the door.
Blue-link search rewarded whoever already ranked. AI search digs much deeper into the pile.
If you run a small site with genuinely first-hand material, that is an opening, not a threat. It’s the most hopeful finding in AI search right now and I almost never see anyone say it out loud.
One thing that looks like a contradiction, in case you bump into it: SE Ranking found 92.36% of AI Overviews include at least one top-10 page.
Both are true!
Most answer boxes contain one top-ranking page — but most individual citations inside them aren’t top-ranking pages.
Different question, different answer.
STOPPED HERE0000000000
GEO vs traditional SEO: what actually changed
Before we get to the fun part, here’s the honest comparison. Not everything flipped.
| Traditional SEO | Generative engine optimization | |
| The goal | Rank in the list | Get quoted in the answer |
| What wins | The whole page | One paragraph it can lift out |
| Ranking position | Everything | Matters much less — 31% of citations rank past 100 |
| The payoff | A click | A citation, and maybe a click |
| Who it favors | Established, high-authority sites | Sites with genuinely original material |
| What still counts | Clear writing, real expertise, sources | …exactly the same things |
Notice that last row. The fundamentals didn’t change. Anyone selling you a totally new discipline is selling you something.
The Two-Click Test
Right. Here’s the method. It takes about a minute and it works on anything.
The Two-Click Test: Take any statistic and click the source it credits. You’ll usually land on another article rather than a study — so find the source that article credits, and click again. If two clicks haven’t taken you to an actual methodology — a sample size, an explanation of how the thing was measured — it’s folklore, not research.
Aka maybe an article take from AI and reposted without verifying. What is called AI slop.
Two clicks is the magic number because real research survives it. A proper study has a methods section. Click, click, and there it is: how many, over what period, measured how.
Folklore doesn’t survive it.
Every link leads to another blog post, which links to another blog post, and somewhere in that chain the trail just… stops.
Try it on the next SEO article you read. Try it on this one — I’ve linked everything on purpose.
So I ran it: where these numbers actually come from
I picked the biggest claim in the pile. In the guides it’s written like this:
Semantic completeness is the number one ranking factor, with a correlation of 0.87.
Two bits of jargon in one sentence, so let’s get rid of both.
“Semantic completeness” just means a paragraph makes sense on its own, without needing the rest of the page. That’s it. That’s the whole idea, dressed up.
And a correlation is just a score for how strongly two things move together. It runs from 0 (no connection at all) to 1 (perfectly locked together). So 0.87 is nearly perfect — about what you’d get comparing people’s height to their weight. For messy real-world web data, that would be extraordinary.
Suspiciously extraordinary, as it turns out.
Click one. The article I was reading credited a source called AI Mode Boost. So I checked what else pointed there. The correlation for images and video. The one for fact-checking. The one that supposedly proves domain authority is dead. The “4.8x” for entities. The 317%.
All of it. One source underneath the entire numeric foundation of modern AI SEO advice.
Click two. The website doesn’t load.
Not slow. Not down for maintenance. I checked it a second way in case the problem was on my end — nothing. The domain isn’t registered to anybody anymore. The company is simply gone.
What the archived copy says
The Internet Archive caught the page once, on 8 September 2025. Reading it is quite the experience. See for yourself – AI Mode Boost
It claims to cover 63 industries. Then it lists its industry breakdown: Technology 23%, Health 19%, Finance 15%, E-commerce 43%. That’s 100%… across four. The other fifty-nine apparently contributed nothing whatsoever.
It advertises “12 months of continuous data collection.” Its own methodology section says June to December 2024. That’s seven months. It’s also a “2025 study” built on 2024 data, still being used to give 2026 advice.
It declares the perfect passage length to be 134–167 words — the number that has since hardened into a rule people repeat with total confidence. Scroll down the same page to its list of recommendations and it tells you to target 127–156 words. It contradicts its own headline finding without leaving the page!
And then there’s the line that explains everything.
“Content Quality: Automated assessment of factual accuracy, citation quality, and semantic completeness using LLM evaluation.”
Read that slowly. They used an AI to score the content for completeness. Then they compared those AI-generated scores against what Google’s AI chose to quote.
So that famous 0.87 largely measures one thing: two AIs tend to agree with each other. It tells you nothing about how Google ranks anything. And it explains the impossibly high number, because genuine studies of this kind usually land somewhere between 0.1 and 0.4.
That’s not my accusation, by the way. It’s their own description of their own method, apparently unread by everyone who has quoted it since.
There’s no author name anywhere on the page. No university, no company research team, nobody. The real methodology sat behind a “download the full 47-page report” form that now leads nowhere.
And the company sold AI Overview optimisation services — so the research concluded that readers urgently needed the exact thing the researchers were selling. Funny how that works.
One honest limit. I’m working from a single archived snapshot. The page might have changed before the site went dark, and I have no way to check. I’d rather tell you that than pretend to be more certain than I am.
What I can say is what’s demonstrable: the source contradicts itself, its headline finding is circular, and nobody can verify any of it ever again.
Meanwhile: one of the articles citing it was last updated July 2026, and it still credits every one of those numbers to a website you can no longer open.
Before You Pay Anyone for AI Search Visibility
There’s a practical use for all this beyond feeling clever about statistics.
There are companies charging real money to get you cited in AI Overviews. Plenty are perfectly good.
One of them was AI Mode Boost. Now out of business, if the website has anything to do with it. It could’ve changed it’s name though. Or maybe bought out by another company. Which is scary. Did they sell their business which bad data? hmmmm
So if you are thinking of hiring a service, before you hand anything over, run the Two-Click Test on their own marketing data. Take the most impressive number on their homepage and follow it.
If it leads to a study with a methods section, that’s a decent sign. If it leads to their own blog post citing their own unpublished research — or to a page that doesn’t exist anymore — you’ve got your answer for the price of sixty seconds.
And you saved yourself a bunch of money!
While you’re there, check who wrote it. A named person with a background you can look up is worth a great deal more than “our research team.”
It won’t tell you whether they’re any good at their job. It will tell you whether they’re honest about evidence, which is a reasonable proxy when you can’t judge the rest.
Three more AI SEO claims that don’t survive checking
The fan-out statistic, the FAQ schema multiplier, and the llms.txt file. All three fall apart under two clicks, and once you’ve got the method this goes fast.
The 161% has moved three times
There’s a real study here. Surfer SEO took the top ten pages for 10,000 keywords, used Gemini to pull out 33,000 fan-out sub-questions, and checked who got cited. Solid work, and they show you how they did it.
What they found: pages that rank for fan-out questions are 161% more likely to be cited.
I have now seen that number attached to three different claims. The study says ranking for fan-out questions.
One guide attached it to writing self-contained paragraphs. Another attached it to naming your headings after fan-out questions.
That last one is my favorite. Ranking for a question is an outcome. Renaming a heading is an action.
You can’t get from one to the other. That’s like noticing marathon finishers all own running shoes and concluding that buying shoes finishes marathons.
FAQ schema doesn’t cause citations
Quick explainer first, because this one sounds more technical than it is.
FAQ schema is invisible code that sits in your page and tells search engines this part is a question, this part is the answer.
Readers never see it. And you don’t write it — your SEO plugin does, from a setting.
If you use Rank Math or Yoast, you almost certainly have schema on your site already and have never thought about it.
The claim is that switching FAQ schema on makes you 3.2 times more likely to be cited. Sometimes it’s quoted as 2.3x. That wobble is your first clue.
Ahrefs actually tested it. They tracked 1,885 pages that added the code and measured what happened.
Almost nothing happened.
Pages with FAQ code do get cited more — but the code isn’t why. Pages that add it tend to be pages that also add clear questions and clear answers, and that is what gets quoted. The code is a passenger, not a driver.
Add it anyway. It’s free. Just don’t expect it to do your writing for you.
The llms.txt file does nothing for Google
You’ll see llms.txt described as “the new robots.txt” — a little file you upload to tell AI models how to read your site.
Google added a line to its official documentation on 15 June 2026 confirming the file “will neither harm nor help your site’s visibility or rankings in Google Search.” Because Google Search ignores it completely.
Google’s John Mueller responded directly to a question about it in a Reddit thread /TechSEO. His answer — “AFAIK” being as far as I know:
“AFAIK none of the AI services have said they’re using LLMs.TXT (and you can tell when you look at your server logs that they don’t even check for it). To me, it’s comparable to the keywords meta tag – this is what a site-owner claims their site is about … (Is the site really like that? well, you can check it. At that point, why not just check the site directly?)”
— John Mueller, Google Search Advocate
The keywords meta tag is the self-reported signal search engines abandoned over a decade ago, for exactly the reason he gives: anyone can claim anything about their own site. So nobody trusts it.
One analysis backs him up: 97% of these files are never fetched by anything at all.
Other AI tools might use it one day. Google doesn’t, and says so plainly.
The fake expert in the guide about verifying experts
This is my favorite thing I found, and I saved it for you.
One of these guides explains that you should build authority by naming real people with verifiable credentials, so Google can confirm you’re a genuine, trustworthy entity.
It gives two examples. One is Sundar Pichai, the CEO of Google.
The other is “Dr. Sarah Chen.”
Dr. Sarah Chen is not a person!
She’s one of the most thoroughly documented fake names on the internet. A placeholder expert that AI models reach for whenever they need somebody credible-sounding.
She shows up as a researcher, a psychologist, a data scientist, an educator, across totally unrelated fields, always distinguished, never real.
Educator Michael G Wagner wrote a whole piece about her, noting that AI models keep producing “the same characters… featuring the same names and occupying the same professional roles across entirely different contexts.”
And I’ll be straight with you: this is especially a Claude habit, and I use Claude every single day. This isn’t me pointing at somebody else’s AI.
So there it is. A guide about proving your experts are real… illustrated with an expert invented by an AI… published without anyone checking.
If you write with AI — and most of us now do — that’s the lesson, and it’s much cheaper to learn it from someone else’s article than from your own.
AI will hand you a confident, plausible, completely fictional source, and it will never once sound unsure.
- Check the names.
- Check the numbers.
- Check the studies.
Frequently asked questions
Is SEO dead?
No. But one version of it is in real trouble — the version built on ranking for simple informational questions and collecting ad money from the traffic. AI answers those questions on the results page now. Sites that sell something, build an email list, or give people a reason to come back are in far better shape.
Do I need to rank on page one to get cited?
No, and this is the most encouraging finding in the whole field. Only 38% of cited pages rank in the top 10, and 31% don’t rank in the top 100 at all. Google splits your question into smaller ones and cites whoever answers the pieces well, wherever they happen to sit.
Does FAQ schema help me get cited?
Barely, on its own. A controlled test of 1,885 pages found the markup itself made almost no difference. What works is the actual questions and answers on the page — clearly written, in plain language. Switch it on anyway, it’s a setting in your SEO plugin and takes two minutes, just don’t expect magic.
Should I add an llms.txt file to my site?
Not for Google’s sake. Google’s own documentation says it ignores the file entirely, and 97% of them are never fetched by anything. If you want one for other AI tools, fine — but it won’t move you in Google.
How long does generative engine optimization take to work?
Nobody credible can tell you, and anyone offering a specific number is guessing. What you can control is publishing something genuinely original, sourcing your facts, and writing sections that make sense on their own. Those worked before AI Overviews and they’ll outlast whatever comes next.
So what do you actually do?
Most AI SEO statistics churn constantly. Every few months brings a new acronym, a new rule, and a fresh batch of numbers with suspiciously precise decimal places.
The underlying stuff barely moves. Here’s your whole checklist:
- Write something only you could write. Your screenshots, your results, the thing that broke on step four.
- Say where you got your facts, and link to them.
- Make each section make sense on its own, so it can be lifted out and still work.
- Put a name on it, and use the same one everywhere. It doesn’t have to be the name on your driver’s license — plenty of good writers use a pen name. It has to be consistent, and it has to be attached to work someone can actually look at.
- Ignore the precise-sounding numbers. The more exact a claim looks, the more you should want to see the study behind it.
That was good advice before anyone said “generative engine optimization” out loud, and it’ll still be good advice when we’re all calling it something else again.
And keep the Two-Click Test in your back pocket.
Follow the citation, follow the next one, and see whether you land on a real methodology or somebody’s marketing blog. It costs you sixty seconds and it’ll stop you repeating a number that quietly stopped being true — or never was.
Have you run into a stat that fell apart when you checked it? Or one you’re not sure about? Drop it in the comments below and I’ll take a look… I’ve clearly got a taste for this now.
Next in this series: how to use AI to check your own article before you hit publish — the six prompts I actually use, and the three things AI genuinely can’t do here. Because “get an AI to score your content” is precisely how we ended up with that 0.87 in the first place.
Definition – GEO represents a new era of organic search. It has introduced new platforms (LLMs) that aggregate information from all over the web, including your site content, your PR, your social presence, your reviews. HEADLESS CONTENT? https://www.google.com/search?q=what+is+headleass+content%3F&rlz=1C1VDKB_enUS1113US1113&oq=what+is+headleass+content%3F&gs_lcrp=EgZjaHJvbWUyBggAEEUYOTIJCAEQABgNGIAEMgkIAhAAGA0YgAQyCAgDEAAYFhgeMggIBBAAGBYYHjINCAUQABiGAxiABBiKBTINCAYQABiGAxiABBiKBTIHCAcQABjvBTIHCAgQABjvBdIBCDUxNTJqMGo3qAIAsAIA&sourceid=chrome&source=chrome.ob&ie=UTF-8
Leave a Reply