GEO vs SEO: what the hell is the difference?

A client asked you about ChatGPT this year. They'd typed their own category in and got three competitors back, or their traffic slid and somebody told them AI was why. Either way it landed on you.

So you did the sensible thing. Schema, headings, speed, the content refresh you'd been meaning to get to anyway. And the answer didn't change.

Then in June, Google updated its own "Do you need an SEO?" guide to say that optimizing for generative AI is optimizing for the search experience — so it's still SEO. Tidy answer. Everybody quoted it, and I understand why. It's a relief.

I'd already spent a few months measuring the thing itself, so let me give you the other kind of answer.

I asked five AI engines fifty real buyer questions each about ten national brands — Gymshark, YETI, Hoka, Helix, AG1, Sonos, Patagonia, Peak Design, OLIPOP and Caraway. 2,500 live answers, then 21,213 citations traced to 8,247 pages, and I went and opened 2,357 of those pages to see what was on them.

Here's what the engines read. Reddit, 1,578 citations — three times the next one. YouTube, 533. Amazon. Forbes. Walmart. Trustpilot.

Not one brand's own website made that list. And ChatGPT, the one your client is actually using, cites the brand's own site in 3.6% of its answers.

The reading list is overwhelmingly about you, and overwhelmingly not by you.

Which is the difference, in one line. SEO is a discipline you apply to a website you control. Almost nothing the engines read is that website.

Counting citations tells you very little, so I classified the 1,884 of those pages that were readable enough to sort.

The biggest slice is somebody's ranked list. "Best cookware brands," "top ten coolers," your client sitting at position four in a round-up they didn't write. That's 27.8%. Another 22.7% is forum threads, buyers talking to each other. Nearly 20% is video, mostly YouTube reviews. Actual editorial, a journalist writing about the space, is 14.9%.

Read that list again as an SEO. Which of those can you optimize?

You can't put your client into a Reddit thread without getting them mauled. You can't add schema to a YouTube review somebody else filmed. You can't edit the round-up that ranked you fourth, and Amazon isn't taking your suggestions.

One brand broke the pattern. OLIPOP is the only one of the ten whose reading list carries more editorial than ranked lists, and that mix is genuinely earnable. Just not with on-page work.

Before any of that reading list matters, the engine has to go and look. I recorded whether it did, from each engine's own disclosures rather than guessing from the answer.

ChatGPT searched on 166 of its 500 answers. Two in three came from memory.

And it quit hardest exactly where buyers decide. On "X vs Y" questions it went to the web 4 times out of 80. On "best thing for my situation," 17 out of 120. So the best page you'll ever write for that client can't win a search that never happened.

The others do look. Perplexity searched on all 500, Gemini on 494, Claude on 445. And I checked whether the difference was the brands rather than the engines — it isn't. 11 points of spread across the ten brands, 66 across the engines.

Which is a harder truth than the citation data. Whatever you publish, however good it is, you don't get to earn the search. That decision was made before your client's question arrived.

Here's the half Google's right about, and it's the best-value thing on this page. I found it by accident. When I sent a reader after every cited page, a lot of them wouldn't open.

About 1 in 6 refused outright, which is just the web being the web. But the group that failed hardest is the one nobody would guess. The brands' own websites. 54% of their own cited pages wouldn't open, and most of those were the brand's own bot protection turning me away. Review platforms were worse — 14 pages cited, 14 blocked.

Meanwhile Reddit let me in every time. YouTube too. The open layer stayed open. The wall was the commercial layer, the part brands pay somebody to control.

And it's a setting, not physics. OLIPOP — the same brand with the editorial reading list — let the reader through on all 15 of its cited pages. AG1's blocked all 22. Same day, same reader. One of those brands decided something and the other one let a vendor default decide it.

That's my reader being turned away, not necessarily theirs. I can't see what the engines get. But bot protection and robots.txt are SEO work, they're checkable in minutes, and at a lot of brands nobody has looked at them since the day they were switched on.

That guidance gets quoted as if it settled something, so let's name the interest behind it.

Google isn't a neutral referee. Every question a buyer types into ChatGPT or Claude or Perplexity instead of Google is market share walking out of the building, and "nothing new to learn, keep doing what you were doing" is a wonderfully convenient position for the incumbent. It keeps the industry's attention on Google's surface and off the ones eating its lunch.

Their guidance also covers exactly one product. It has no standing over what Claude cites or how Perplexity picks a source, and the engines are nowhere near each other. Reddit is 14.9% of Perplexity's diet and 11.8% of Google's own AI Overview. On Claude it's a measured zero — not one Reddit citation in 7,260. There's no single "AI diet" to feed, and a source strategy built by watching one engine is a strategy for one engine out of five.

Of the cited pages carrying a readable date, 53.9% are more than a year old and 31.9% are more than two.

The age sits in the forums. Nearly half the forum threads the engines cite are over two years old, against one in six ranked lists, because list publishers refresh for a living and nobody refreshes a thread.

So the conversation layer — the part where buyers talk to each other, the part you can't edit — is also the part most likely to be describing a product your client has since fixed, repriced or discontinued. Somebody's two-year-old complaint about shipping is still out there answering questions on their behalf.

None of this matters if the target won't hold still, so I asked all 500 questions again inside twenty-four hours. Same engines, same wording, and I compared everything.

The aggregate barely moved. Brands were mentioned in 69.0% of answers the first time and 68.3% the second. Underneath it the engines rewrote roughly half their own searches, and which pages came back split hard: Perplexity returned to the same URLs 90.3% of the time, Claude 66.8%, ChatGPT 51.7%, Gemini 33.9%. Ask Gemini twice and you get an answer built from a different reading list.

The surfaces churn. The aggregates hold. Both halves of that matter, and they tell you exactly what goes in the deck.

The aggregates holding is what makes this worth measuring at all — the overall picture is stable enough to work on and report on month to month. The surfaces churning is why you never promise a client a position in an AI answer, and never hand them a single query as a result. There'll be a different one on Thursday and you'll be the one explaining it. Where recurrence is high, a page you win stays won. Where it's low, visibility is a monthly reading, not a trophy.

Two jobs, and only one of them is the one agencies already know how to do.

Job one: make the site machine-readable. Check the front door the way an AI crawler sees it — curl -I -A "PerplexityBot" yourdomain.com — and look at the status code. A surprising number of brands watch their own site answer with an error. Then the rest of the hygiene: content in the HTML rather than assembled in the browser, the answer near the top, schema so nothing has to guess who you are. This is SEO, it's cheap, and it's necessary. Google earned that half of the argument.

Job two: go and earn a place on somebody else's page. This is where agencies go quiet, because there's no checklist. So here's what it actually looks like in the week.

That round-up your client sits at position four in — somebody publishes the next year's version every December, because refreshing it is what keeps it ranking. There's a window a few weeks wide when that writer is deciding who's in the new one, and turning up inside it means you're helping with a job they'd already scheduled rather than asking a stranger for a favour.

That Reddit thread you can't post in — you can still read it. If forty people are saying the same thing about the shipping, that's not a marketing problem to spin. It's a product complaint sitting in the exact place the engines go to read, and fixing the actual thing is the visibility work now.

That YouTube review from two years ago that keeps getting cited — somebody already did the hard part. The only question is whether anyone at your agency has ever sent that person the new model.

Ranked lists, forum threads, video reviews, the occasional piece of real editorial. Digital PR, community presence and review-platform work, and I've written about how a small shop actually staffs that.

The first job gets you eligible. The second one gets you cited. Calling the whole thing SEO tells an agency the work they already do is enough, and the reading list says plainly that it isn't.

And while everyone argues about what to call it, the ground is moving underneath.

Two thirds of Google searches now end without anyone clicking anything. When an AI Overview shows up, the page ranking first loses more than half its clicks. DMG Media told the UK regulator the Daily Mail's desktop click-through went from 25% to under 3%.

So the surface you can optimize sends less every quarter, and the surface you can't optimize decides more. No acronym fixes that.

The study numbers here — the citations, the search rates, the pages that wouldn't open — came out of the thing I build, so weigh them accordingly. The click-loss and zero-click figures are other people's, and they're linked.

AI Visibility Pulse asks the engines your client's real buyer questions, then goes and reads what sits behind every answer — the Reddit threads whole, the round-ups, the videos with their full transcripts — and tells you which pages are carrying your client and which ones are handing the answer to a competitor. Every run is versioned, so the second pulse reports what moved since the first. That's how the 69.0% became 68.3%.

You don't need any of it to start, though. The reading list is public. That's the whole point of this article.

Pick the account that matters most. Search their category on Reddit and read what comes back — not the ranking, the conversation. Count how many times a competitor gets named and your client doesn't.

Then run curl -I -A "PerplexityBot" theirdomain.com and see whether their own front door even opens.

That's an afternoon, and I'd bet it annoys you. But it'll tell you more than the last twelve ranking reports did, and you'll know by teatime whether you've got a problem worth selling a fix for.

What's the difference between GEO and SEO?

SEO is work you do to a website you control. In a field study StyleForge ran across ten national brands — Gymshark, YETI, Hoka, Helix, AG1, Sonos, Patagonia, Peak Design, OLIPOP and Caraway — five AI engines produced 2,500 live answers and 21,213 citations, and the pages they actually read were Reddit, YouTube, Amazon, Forbes, Walmart and Trustpilot. No brand's own website made that list. The technical hygiene is shared between SEO and GEO; the part that decides whether a brand gets cited happens on pages it doesn't own.

Why does AI cite Reddit so much?

Because Reddit is where buyers discuss products openly, at length, and without a marketing department in the way. In StyleForge's field study of 21,213 AI citations across ten national brands, Reddit was cited 1,578 times — three times more than any other website, and it appeared for all ten brands. It is very uneven by engine, though: Reddit is 14.9% of Perplexity's citations and 11.8% of Google's AI Overview, but only 0.2% of ChatGPT's and exactly zero of Claude's 7,260.

Does ChatGPT read my website before answering?

Usually it doesn't read any website. In StyleForge's field study, ChatGPT went out to the live web on 166 of 500 answers about ten national brands — and on "X vs Y" comparison questions, just 4 times in 80. Claude searched on 445 of 500, Gemini 494, and Perplexity on all 500. Whether an engine searches at all is a property of the engine, not of your brand, your fame, or your content.

Google says optimizing for AI is still SEO — is that wrong?

Half right. Google's June 2026 update to its "Do you need an SEO?" guide is accurate about technical hygiene, and it covers exactly one surface: Google's own. It has no standing over what Claude cites or how Perplexity selects sources. It's also worth remembering that Google competes with those assistants for the same questions, so "nothing new to learn" is a convenient position for it to hold.

What's the fastest AI visibility check I can run today?

Whether an automated reader can open the site at all. In StyleForge's field study, 53.8% of the brands' own cited pages wouldn't open for a machine reader — 70 of those failures were the brand's own bot protection. Run curl -I -A "PerplexityBot" yourdomain.com and read the status code.

Can I promise a client a position in AI answers?

No. StyleForge re-asked all 500 questions inside twenty-four hours and found the aggregate barely moved — brands were mentioned in 69.0% of answers, then 68.3% — while the pages underneath churned hard: Perplexity returned to the same URLs 90.3% of the time, Gemini only 33.9%. Report share of voice across many buyer questions and its movement between runs. Never report a single query as a result.

• Part One · The Retrieval Study — do the engines even search before answering, and what do they type when they do.
• Part Two · The Source Study — every page behind the answers, what it is, how old it is, and whether it opens.
• You're not buying links, you're buying your way out of PR — how the off-site half actually gets done.