Pew Research Center tracked the browsing behavior of 900 U.S. adults across roughly 69,000 Google searches in March 2025. When an AI summary appeared, users clicked a traditional search result in just 8% of visits. Without an AI summary, that number was 15%.

Clicks on the links cited inside the AI summary? One percent.

Now, let’s not do the usual marketing thing and turn one study into the death of civilization. Only 18% of the searches Pew tracked produced an AI summary.

But the interesting part is which searches produced one.

Only 8% of one- and two-word searches triggered an AI summary. For searches of ten words or more, the number jumped to 53%. For searches phrased as questions, it hit 60%.

In other words, exactly the kinds of searches people make when they’re trying to understand a problem.

Nobody wakes up, types “contract lifecycle management,” slaps down a credit card and calls it a day.

They ask something more like:

“How do we stop legal from becoming the bottleneck on every renewal?”

And increasingly, the answer appears before the traditional search results, already assembled, summarized, and decorated with somebody else’s name.

We call this phantom rank.

Your dashboard says you’re number three. Or four.

Congratulations! But. Number three or four may now be sitting outside the answer box most buyers are paying attention to.

It’s like horse racing: Do you really want your horse to run a good race but still always come in out of the money?

Ranking and getting cited aren’t the same thing anymore

Traditional SEO gave us a wonderfully familiar measurement system: a list.

You’re number one. They’re number two. Somebody you’ve never heard of is number three. Everyone goes into the Monday meeting with a chart.

Generative search makes that much messier.

The Princeton researchers who helped define Generative Engine Optimization made this point in ⁠their KDD 2024 paper. Ranking works nicely as a visibility metric when a search engine returns an ordered list of sources. Generative engines don’t necessarily do that. They synthesize information into an answer, then weave sources and citations into it.

So visibility is no longer one number.

Did you appear?

Where did you appear?

How much of the answer came from you?

Did the engine actually attribute any of it to you?

Those are four different outcomes.

And two of them are unpleasant for marketers.

You can be retrieved but not cited. The engine finds your page, learns from it, constructs an answer, and puts somebody else’s name on the finished product.

You can also be cited but not clicked.

(See that 1% number above.)

The Princeton boffins also found that old-school tactics such as increasing keyword density had little impact on visibility, while content associated with authority and specificity performed much better. Adding statistics, credible citations and authoritative quotations improved visibility by as much as 40% in their benchmark.

That “as much as” is doing some heavy lifting there.

Half the internet seems to have converted 40% into a magical GEO uplift number, because apparently we learned absolutely zilch from two decades of SEO case studies.

It isn’t.

Forty percent was the upper end of improvement on a particular visibility metric in a controlled benchmark. The strongest techniques landed roughly in the 30–40% range. Useful finding? Absolutely.

Is it a lock that your traffic will increase 40% if you add a few statistics? No.

What the research does tell us is more interesting anyway: specific claims are citeable.

“Our platform streamlines workflows and empowers teams to drive better business outcomes” is corporate oatmeal. There’s nothing an answer engine can grab.

“Legal teams spend X hours per month doing Y, according to Z research” is a claim. It has edges. It can be extracted, evaluated and attributed.

If you want machines to quote you, give them something worth quoting.

“But Google says we don’t need schema for AI”

Somebody on your team is going to say this. They may already have the link ready to go.

And technically, they’re right.

⁠Google’s documentation says there’s no special schema.org structured data required to appear in its AI features. You don’t need some magical new MakeMeFamousInAI markup or a secret file tucked into the root directory.

Great. Now keep going…

The same documentation tells site owners to make sure crawling is allowed through robots.txt, CDN and hosting controls; that important content can be found through internal links; that useful information is available as text; and that structured data accurately reflects what’s visible on the page.

Google isn’t saying structure doesn’t matter.

It’s saying there isn’t an AI-specific markup shortcut.

Those are very different things.

Structured data still helps machines resolve what you are, what you sell and how the pieces of your site relate to one another. And that matters because one of the most common problems we find isn’t really a schema problem.

It’s an entity problem.

The open web doesn’t clearly agree on what the company is. So the machine fills in the blanks.

Sometimes from directories. Sometimes from third-party sites. Sometimes from some forgotten field in a CMS that nobody has touched since Obama was president.

We discovered our own company was being identified as “MSMC” for a while, courtesy of an ancient Yoast field.

Nothing like discovering that your sophisticated digital identity strategy is being run by a form somebody filled out in 2019.

Half of this is a settings problem

This is why an AEO/GEO visibility audit shouldn’t begin with a content calendar.

It should begin with a much less glamorous question:

Can the machines actually reach you?

⁠OpenAI publishes separate controls for its crawlers. OAI-SearchBot is used to surface sites in ChatGPT search. GPTBot is associated with content that may be used to improve foundation models. ChatGPT-User handles certain user-initiated requests.

Those controls are separate. Which creates a wonderfully stupid way to disappear.

Someone forwards an alarming article about AI companies scraping websites. Security decides to be proactive and blocks every OpenAI user-agent in one commit.

Marketing notices nothing.

But eight months later, somebody asks why the company never appears in ChatGPT search and commissions a content initiative to solve a robots.txt problem.

I wish that were an absurd hypothetical.

The same thing happens with JavaScript-dependent content, CDN rules that block datacenter IPs, painfully slow pages and important copy that technically exists but can’t reliably be rendered by crawlers.

Those aren’t messaging failures. They’re plumbing failures.

And before you spend $50,000 remodeling the kitchen, it’s worth checking whether somebody turned off the water.

Then there’s the ceiling we forget about

Here’s the more annoying part.

You can fix everything on your own site and still lose.

Pew found that Wikipedia, YouTube and Reddit were among the most frequently cited sources in Google’s AI summaries, collectively accounting for 15% of cited links. Government sites were also disproportionately represented, making up 6% of AI-summary sources compared with 2% of standard search results.

There’s a reason. Machines like corroboration.

For B2B companies, that means review platforms, analyst coverage, trade publications, customer references, press mentions and other independent sources that confirm you are what you say you are.

This is where smaller companies often hit a wall.

The problem isn’t the homepage. It’s corroboration density.

Suppose G2 has a category page for the thing you sell and your company has four reviews.

You can write the most beautifully optimized category page since the invention of the <title> tag. The engine still has a pile of third-party evidence sitting somewhere else.

Guess which source looks safer to cite?

Now you’re being described in somebody else’s words, on somebody else’s website, according to somebody else’s timetable.

That’s an uncomfortable finding to put in an audit because the solution isn’t “publish six blog posts.”

It’s PR. Customer marketing. Reviews. Analyst relations. Partnerships.

And, everybody’s favorite growth tactic, patience.

But telling a company to build another pillar page when its actual problem is eleven customer reviews would be craptent malpractice.

So what does an AEO/GEO visibility audit actually score?

When we do one, we look at six levers. Each gets a score from one to five, for a total of 30 points.

Machine Accessibility. Can crawlers and agents actually get to the page, render it and read it?

Structured Data Coverage. Is the schema there, is it valid, and does it agree with what a human sees?

Entity Definition. Does the open web clearly and consistently understand who you are, what you do and where you fit?

Answer Extractability. Have you written something an engine can pull out and use as a clean answer, or buried the useful part under 600 words of corporate malaise?

Corroboration Density. Do independent sources give an engine enough evidence to trust what you’re saying?

Conversion Bridge. If an AI-referred buyer actually lands on the site, is there an obvious next move, or have we successfully optimized our way into a dead end?

The composite score is useful. By itself, it’s also close to meaningless.

A 20 out of 30 caused by a Machine Accessibility problem might be fixed on Tuesday.

A 20 out of 30 caused by a Conversion Bridge problem could require a quarter of work, three departments and a completely different budget owner.

That’s why the lever scores matter more than the big number.

We also build a query inventory around the questions buyers actually ask, then test what shows up in traditional search and across AI engines.

Who gets mentioned?

Who gets cited?

Who doesn’t exist?

Then we run the same six-lever analysis against the competitors the engines are choosing instead of you.

That’s usually the slide where the “I’m not sure this AI search thing matters yet” conversation gets considerably shorter.

Why bother doing this now?

We’ve been doing versions of AEO/GEO work since before everyone agreed what to call it.

Not because we had a crystal ball. Clients working in AI were already seeing buyers use these tools differently, and the visibility problems were showing up before the terminology caught up.

For Evisort, blog traffic increased 316% in roughly a month after the first implementation.

And before anybody puts that number in 48-point type on a LinkedIn carousel, it came off a low starting base.

Still a real number. Still worth examining.

But we’ll show you the starting point because percentages without denominators are how perfectly respectable marketing departments accidentally become infomercials.

The bigger reason to audit first is simpler:

Content is the most expensive possible way to fix a robots.txt file.

Before you commission 30 articles, rebuild your resource center or start stuffing “best practices” into another 2,500-word guide nobody asked for, find out what’s actually broken.

Run the queries your buyers really use.

See whose name comes back.

Find out why.

We’ll do the first pass for free. And if the answer turns out to be that you’re already the source AI engines trust in your category, we’ll tell you.

Then you can spend the money fixing something that actually needs fixing.

This is a standard paragraph for your post.  Follow the same rules described above.  Try to avoid having large blocks of run-on text, which hurt readability, stickiness, and visitor loyalty.

Use callouts drawn from the copy to emphasize key points.