The Race You’re Not In

Updated October 2, 2026: corrected the accessibility figure to match WebAIM’s March 2026 report, and added a quick summary at top.

In the space of one month, my inbox told me:

• to abandon old-school SEO for GEO
• that Reddit was where AI found all its answers
• that Reddit was losing its grip on AI answers
• that visitors from AI search convert 23 times better than organic. Also 31% better. Also worse than Google traffic.

Most of them persuasive. All confident. All with something to sell.

I know some of you are getting the same emails, because you’ve forwarded them to me.

Back in May I promised to write up a study on AI search that pushed back on the panic. Then summer happened. When I finally sat down to do it, I found the research had gone in three directions at once, which turned out to be the more useful story.

Why the advice keeps contradicting itself

Yext looked at 155.5 million citations across 1,623 brands in the first quarter of this year. They found roughly 80% of what AI answers cite comes from sources a business controls, about half from its own website and about a third from its listings. That was the study I’d been so pleased with in May.

Omniscient Digital looked at 23,387 citations from branded queries in January and found close to the reverse. Content on the business’s own domain accounted for 23%. Other people’s sites accounted for 48%.

Neil Patel, in a webinar in June, landed near the second one. If the rest of the web doesn’t trust you, he said, AI won’t trust you either, no matter how good your website is.

So which is it? All of them, as it turns out.

Yext asked local questions, the kind a nearby customer asks. Omniscient asked evaluation questions: comparisons, pricing, reviews. Patel was describing the stage before either, where people research across a dozen places and only arrive at your website once they’ve mostly decided.

Different questions, different answers. Nobody’s lying.

They’re measuring different moments and reporting them as though each were the whole picture.

One more thing about the questions, because it surprised me: When somebody types something into ChatGPT, the model doesn’t search for what they typed. It expands it into several searches first, and the words it adds most often are “best,” “reviews,” and the current year. (This is called a query fan-out, if it matters to you.)

So a plain question about finding a counsellor nearby becomes a search for the best counsellors, with reviews, in 2026, before anything gets looked up.

Nobody chose that.

Everyone here is selling something

Yext sells listings management and found that listings drive a third of citations. Patel sells search strategy and found you need a broader search strategy. Omniscient sells content and found that content off your own site is what gets cited.

That doesn’t make the numbers false. The samples are large and the methods look sound. But there’s very little disinterested research in this field, and the studies with the best distribution are the ones with a product attached.

Which is why the most useful number I found came from the one organization with nothing to sell.

The Pew Research Center took 900 American adults who agreed to let researchers see their actual browsing, pulled every Google search they ran in March 2025, and looked at what those people did next. Nearly 69,000 searches. When an AI summary appeared, 1% of visits ended with someone clicking a source link inside it. One percent.

An entire industry is currently selling you citations. The only neutral measurement of what a citation is worth in traffic says: almost nothing.

Being cited still matters for whether people recognize your name, and I think it does matter. But if someone is selling you AI visibility on the promise of clicks, that’s the number to ask them about.

The race you’re not in

Here’s where most of this advice goes wrong for a business like yours.

You’ll see reports saying 50 domains account for most of what AI cites, with Reddit and Wikipedia at the top. Those reports are accurate. They’re also describing a race you were never entered in.

No small business is going to outrank Reddit. Not you, not me, not anyone reading this. Reddit has 20 years of people arguing about everything and the domain authority to match. It will win “how do I know if I need therapy” and “is it worth hiring a web designer” every single time.

It has nothing to say about whether there’s a family counsellor within a 30-minute drive who can see your kid in October.

That’s the whole thing. Reddit wins the general question. You win the specific one, the local one, the one with your name or your town in it. Losing a race you’re not running is not a problem you need to solve.

Pew’s numbers back this up from another angle. An AI summary appeared on 18% of all the Google searches they tracked, and they clustered on long, question-shaped searches: 53% of searches of 10 words or more, 60% of ones starting with “who” or “what.” Short, specific searches mostly didn’t trigger one at all.

While I’m here, a correction. In May I passed along a line about not worrying about Reddit, on the strength of forums sitting at 2% of citations. References to Reddit inside ChatGPT’s expanded searches have since climbed from around 0.15% to 3.68% over five months. Still small, but also about 25 times bigger than it was, and deliberate rather than accidental. The reassurance I offered was true when I offered it and is aging quickly.

What is actually moving

I don’t want to hand you a comfort that expires the same way my Reddit line did.

Semrush went through 600,000 US keywords between November and April. AI Overviews on commercial searches, the comparing-and-shortlisting kind, grew 71% over those months. In finance it was 231%. Google Ads and AI Overviews now turn up together about twice as often as a year ago, which tells you Google is no longer keeping AI answers away from the queries it makes money on.

Transactional searches, the ready-to-book kind, went the other way and dropped 5%.

So the shortlisting stage is being absorbed fastest, and the moment somebody is ready to actually hire is holding. For now.

That “for now” is doing real work, and anyone who tells you where that line will be next spring is guessing.

What your website is actually for now

Its job changed, and not in the direction people assume.

These days, I talk a lot less about your website being where people find out about you anymore. It’s where they decide whether to believe what they heard, and where a machine checks whether you’re real. Which makes the most valuable thing on it the plain sentence most websites are too polished to include: what you do, who you do it for, and where you are.

Then there are your listings, which are duller and matter more than anyone tells you. Your Google Business Profile, the professional directory your association runs, the membership page nobody has looked at since 2019.

And then the housekeeping, which is duller still and considerably more irritating.

A lot of directories never got their information from you in the first place. They pulled it from somewhere older, and they’ve kept it. Legacy sources like the Yellow Pages data seed dozens of smaller sites, and those sites feed each other, so a business name you stopped using years ago keeps circulating on pages you have never visited, next to a phone number you gave up and a category somebody typed in wrong at the start.

Ask me how I know: I’m in the middle of this myself, on two fronts. Bright Webs became Clarity Web Design Studio in 2023, and I changed my own surname this spring. I still turn up listings carrying the old business name, the old personal name, or some mix of the two, on sites I have no memory of ever joining and many I’ve never even heard of before.

Some of them you can claim and correct in about 4 minutes. Some have no login at all, just a contact form that goes nowhere in particular. And a few will show you exactly what’s wrong, then ask for a paid subscription before they’ll let you touch it, which is a business model I find downright obnoxious.

You won’t catch all of them. I haven’t. Fix the ones people actually land on, keep your name, address and phone identical everywhere you do have control, and make peace with the tail. I wrote in July about the gap between the version of you people find and the current one. This is the form-filling half of closing it, and it is a job that does not seem to ever finish. It’s closer to lawnmowing.

Two jobs, moving in opposite directions

Before anybody reaches you, the other stuff does the heavy lifting. Listings, directory profiles, other people talking about you. That’s what all of this research keeps pointing at, and it’s the part almost nobody has a plan for.

BrightLocal surveyed 1,002 American consumers in February:

• 97% read reviews for local businesses.
• Nearly half won’t consider one with fewer than 20 reviews.
• Three quarters weight the last 3 months most heavily, so a wall of glowing reviews from 2022 is doing less than you think.
• 68% now want to see at least 4 stars, up from 55% a year earlier.
• 80% prefer a business that replies to reviews, and half of them dismiss replies that read as templated.

And two more from that survey worth putting in a fresh list:

• Google’s share as a place people read reviews fell from 83% to 71%, while ChatGPT and similar tools jumped to 45%, now the third most common.
• And 82% read AI-generated review summaries, with 23% making the call on the summary alone.

It also means the businesses whose professional bodies prevent them from soliciting reviews are carrying a real structural cost, not an inconvenience. If that’s you, the answer isn’t a workaround. It’s building the same signal somewhere else: professional directories, referral partners, associations, and a Google profile you actually keep current.

All of that is the discovery half. If your work comes by referral, a good number of your buyers skip it entirely.

Somebody gets your name from a person they trust, usually with your link attached, and goes straight there. No search. No AI summary. No results page setting you beside four alternatives. They arrive on your website cold, holding one recommendation, deciding whether to believe it.

Which makes your website more important, not less. It’s the only place you control the story. It’s the only version of you that you can keep current on purpose. And it’s the only surface where you decide what somebody sees first, second and third, rather than a model deciding for you.

So there are two jobs here, and they’ve moved in opposite directions. Being found now leans heavily on things that live outside your site. Being believed, once somebody arrives, sits almost entirely on it.

And whatever you build, check that it can be read at all. More than 95.9% fail basic accessibility testing. A page a screen reader can’t work through is often a page a machine can’t summarize.

Good design is good for everybody, including the robots, which I did not expect to be writing in 2026.

None of this is exciting. Most of it is the same work I was recommending in 2019, with different reasons attached.

That was the point of the May piece too. What looks like a sudden result is almost always somebody’s seventh year of showing up. The thing standing between you and the person looking for you keeps changing shape. Being findable, reachable, and unmistakably yourself has not changed at all.

My tagline is Building websites for humans, by humans. The machines seem to prefer those too, when they can find them.