Your inbox has probably carried at least one pitch for a GEO package this year. Short answer first. GEO is not replacing SEO. Google’s AI answers are rooted in the same core Search ranking and quality systems its ordinary results run on. That’s Google’s own phrasing, and the hedge in it is deliberate. Other assistants retrieve differently, and I’ll cover that. The practices getting cited tend to be the ones whose sites were already built so a machine can lift a clean answer out of them.

That’s most of the geo vs seo question, at least for Google, where the documentation settles it rather than my opinion doing the work. The rest of this post is about the part that genuinely did change, which is smaller than the pitch decks claim and more annoying than they admit.

01What GEO Actually Means, and What Is GEO in SEO

GEO stands for generative engine optimization. It’s the practice of making your content likely to be quoted by an AI system, rather than clicked on in a list of blue links. AEO, answer engine optimization, is the same idea under a different acronym.

Here’s the part the acronym hides. Google defines both terms in its own guidance, then says that from Google Search’s perspective, optimizing for generative AI search is optimizing for the search experience, and that’s still SEO.

That isn’t marketing spin from a search engine defending its turf. It’s a description of how the plumbing works. Google’s generative features run on retrieval-augmented generation, which means the AI answer is grounded in pages pulled from the ordinary Search index. There’s also query fan-out, where the model fires off several related searches at once instead of one. Both mechanisms start at the same place. The index. The one your SEO already feeds.

Now the caveat that most posts on this skip. That description is Google’s, about Google. ChatGPT, Perplexity and Copilot each retrieve differently. Copilot is grounded substantially in Bing’s index, per contemporary reporting on how Microsoft’s generative experience is assembled. OpenAI’s crawler documentation names its own robots, including OAI-SearchBot for ChatGPT search and ChatGPT-User for user-triggered fetches. Perplexity’s crawler documentation says its agents gather and index information from the internet. All three publish crawler and publisher guidance. None publishes a ranking or citation formula, and none is obliged to behave like Google.

What travels across all of them is the input. Every one of these systems needs a page it can fetch, read, and quote accurately. That part is platform-independent, and it’s the part you can actually work on.

So when someone sells you GEO as a separate discipline with a separate invoice, the honest question is which part of it isn’t just SEO with a 2026 sticker on it.

There is a real answer to that question. It’s just narrower than the pitch.

More on how AI Overviews and ChatGPT pick the practices they name.

02What Actually Changed, Measured

I’d rather show you this on a site I’m allowed to be rude about, so I’m using ours.

In August 2026 we pulled ninety days of Search Console and Analytics data for transference.studio, looking for a content problem. We found a measurement problem instead, and underneath it a demand shift.

The commercial terms this business was built on are shrinking. Checked against DataForSEO on August 21, 2026, our core service keywords are down between 18 and 89 percent year over year. The AI-search terms sitting right next to them are up between 26 and 285 percent.

Read those two together, because separately they mislead.

My reading is that demand relocated rather than disappeared. People who used to search “therapist website design” appear to be asking an assistant instead. But two keyword trends moving in opposite directions don’t prove the same people moved between them, so treat that as interpretation, not measurement.

The trap underneath it is real whether or not my interpretation is. Search Console impressions can fall while visibility elsewhere rises, because Google’s reporting has never covered what other assistants do.

Then something happened while I was checking those figures, and it’s more useful than the figures themselves.

I first pulled this keyword set on August 11. When I re-pulled the identical ten keywords ten days later to verify them, eight came back byte for byte the same. Two had moved. “GEO vs SEO”, the term this very article targets, read up 53 percent year over year on the first pull and down 17 on the second. “Therapist SEO” read down 61, then down 18.

The tempting write-up is that keyword data is unreliable. It isn’t, and the sequence is worth laying out properly, because the explanation is the actual lesson.

Read the bottom row first. The June number never moved. Nothing was restated, corrected or revised, on either term. What moved was which month the headline figure referred to.

These tools report search volume on a lag and add one month at a time. On August 11 the newest month available was June, so every figure in the set compared June 2026 against June 2025. In the ten days that followed, July landed, but only for two of the ten keywords. Those two quietly started reporting July 2026 against July 2025 instead. Same label, same column heading, a different comparison underneath it.

Which means the eight that held steady hadn’t held steady at all. They simply hadn’t refreshed yet. My ten-keyword set is now a mix: two terms on July, eight still on June, sitting in one table under one heading as though they were measured the same way.

Only one of the two genuinely reversed, either. “GEO vs SEO” flipped sign. “Therapist SEO” stayed negative and got less negative, which is a decline easing rather than a trend inverting. Reading those two as the same event would be a second mistake stacked on the first.

So the fragile thing here isn’t the data. It’s the label. A year-over-year percentage belongs to one specific month, and quoted without that month it changes meaning the next time the source refreshes, with nothing on screen to tell you it happened.

That gives you a sharper question to ask anyone showing you a keyword trend, including me. Not just when was this pulled, but which month is each row measuring. If different rows are measuring different months, it isn’t one table. That applies to every number in this post.

We also found something more embarrassing, and it’s relevant here. Across all 32 landing pages, over ninety days, we could attribute exactly zero conversions, because no key events had ever been configured in our own GA4. We sell measurement. We hadn’t measured ourselves.

I’m including it because it’s the thing I most often find on practice sites. Before anyone debates GEO strategy, check whether your analytics can prove anything at all works. In my experience, a lot of them can’t.

For a wider dataset than one practice’s numbers, I keep a sourced set of SEO statistics for therapists.

03Does SEO Still Work in 2026?

Yes, and the evidence is unusually direct for once.

In May 2026 Google published a guide on optimizing for generative AI features, and included a section titled “Mythbusting generative AI search: what you don’t need to do.” It lists five things. Two of them are worth a private practice owner’s attention, because both are being actively sold to clinicians right now.

You don’t need special AI files. Google states you don’t need to create machine readable files, AI text files, markup, or Markdown to appear in Google Search, including its generative AI capabilities, because Google Search doesn’t use them. That includes llms.txt, the file currently being packaged as “AI readiness.”

The independent data is blunter. In 2026, Ahrefs checked server logs across 137,210 domains and found that 97 percent of published llms.txt files received zero requests in a month (Ahrefs, “We analyzed 137K sites: 97% of llms.txt files never get read”, June 2026). More telling still, AI bots never requested the file on domains where it didn’t exist. A crawler that used it would go looking and log a 404. None of them do.

You don’t need to rewrite everything for robots. This is the same argument I make about chasing algorithms generally. Google says AI systems understand synonyms and general meaning, so there’s no need to chase long-tail phrasings, and no requirement to chop your pages into tiny chunks.

Notice what that leaves standing. If the special files do nothing and the AI-specific rewriting does nothing, then for Google at least, the GEO delta largely comes down to whether your content is good and extractable. Google names other factors too, including crawlability, indexing and page experience. It’s the same job SEO always was, with a stricter grader. The other assistants publish crawler documentation but no equivalent ranking formula, so treat anyone quoting precise rules for them as guessing.

The longer version of this argument is in Is Traditional SEO Dead?.

04The Four Things a Practice Site Needs That a Normal SEO Checklist Misses

A standard SEO audit checks titles, speed, mobile, and links. Useful, and not sufficient. These four are where practice sites actually lose AI citations.

Entity clarity

An AI system has to work out who you are before it can recommend you. Your name, credentials, licence jurisdictions, specialisms and location need stating in text a machine can read, not implied by a logo or buried in an image.

I run into the same failure repeatedly on therapy sites. The About page says “I’ve worked with anxiety for fifteen years.” Nowhere does the site plainly state “Dr. X is a licensed clinical psychologist practising in Illinois, specialising in anxiety and autistic adults.” The first is warmer. The second is the one that gets extracted.

Write both. Put the warm version in the flow and the flat version somewhere findable, usually your about page. The same details need to match across your local listings and Google Business Profile.

Passage-level answerability

This has the clearest data behind it, and it’s why the FAQ story got so confusing.

SE Ranking analysed 129,000 domains across 216,524 pages to find what correlates with ChatGPT citations (SE Ranking, “How to optimize for ChatGPT”, published November 24, 2025). Sections of 120 to 180 words between headings averaged 4.6 citations. Sections under 50 words averaged 2.7.

That’s the shape of an extractable answer. Long enough to be complete, short enough to lift whole. If someone must read your previous three paragraphs for an answer to make sense, it can’t be quoted.

The same applies to what you publish on the blog. Go through your service pages and ask one question per section. Could a stranger paste this block elsewhere and have it still make sense? If not, it isn’t answering anything.

Schema completeness

Here’s where I have to correct something the industry keeps repeating, including things I’ve probably implied myself.

Schema markup did not move AI citations for the pages that have actually been tested. In 2026, Ahrefs ran a controlled test on 1,885 pages that added JSON-LD schema for the first time, matched against 4,000 comparable control pages, measuring citations 30 days before and after (Ahrefs, “We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved.”, May 11, 2026). Google AI Overviews moved down 4.6 percent. AI Mode moved up 2.4 percent and ChatGPT up 2.2 percent, both indistinguishable from noise. Their conclusion was that adding schema produced no major uplift in citations on any platform.

Why does everyone believe otherwise? Because in a plain observational look at 6 million URLs, pages cited by AI were roughly three times more likely to carry schema, 53 percent against 18 percent. That correlation is real, but it doesn’t establish cause. Sites that bother with schema also build links, publish decent content, and maintain their pages. The schema is a marker of a site doing everything else properly.

So why is schema on this list at all? Two reasons. Fabrice Canel, a principal product manager on Bing, said publicly at SMX Munich in March 2025 that schema markup helps Microsoft’s LLMs understand content, referring to Copilot. And schema still earns rich results in ordinary Google Search, which for most practices I work with is still the largest single source of organic traffic. Check your own numbers rather than taking mine.

Do it because it’s correct and cheap. Don’t pay a premium for it as an AI service.

One specific fix while you’re in there. If a previous designer added star ratings to your practice’s own LocalBusiness markup, remove them. Google’s standing guideline is that review and aggregateRating on LocalBusiness are only for sites capturing reviews about other businesses. Self-serving ratings on your own markup aren’t eligible. I find this on clinician sites regularly. The ethical way to build reviews doesn’t involve marking up your own.

Off-site mentions

The uncomfortable one, because you can’t fix it on your website.

In the SE Ranking data, referring domains were the single strongest predictor of ChatGPT citations. Practices with heavy brand presence on Quora and Reddit averaged 7.0 citations against 1.7 for minimal presence. Presence on review platforms averaged between 4.6 and 6.3, against 1.8 for absence.

Before anyone panics about ethics codes, note what that does not mean. Not soliciting client testimonials. It means your practice existing, accurately and consistently, where your field is discussed. Directory listings, professional association pages, a podcast appearance, a quoted comment in local news.

Read that alongside what your ethics code permits and ACA Code C.3.b on testimonials.

05What You Can Stop Paying For

Two line items, effective immediately.

FAQ schema as an SEO deliverable. Google switched off the FAQ rich result on May 7, 2026 and deleted the documentation on June 15, 2026. Back in August 2023 the feature had been narrowed to government and health sites, which is why a lot of therapy practices kept the markup and why some agencies sold it as a health-vertical advantage. That exemption no longer exists for anyone.

Now the important half, because people get this backwards and do real damage. Do not delete your FAQ content. The markup lost its payoff. The content never lost its purpose.

The same SE Ranking analysis is worth reading carefully here. Pages carrying FAQ schema markup averaged slightly fewer ChatGPT citations than pages without it. But the researchers found the model treated the absence of visible FAQ sections as a negative signal, once you account for FAQ sections tending to appear on smaller pages that earn fewer citations anyway. Their recommendation is to add FAQ or common-questions sections to major articles and use question-style subheadings naturally.

That’s the argument of this post in one dataset, with one qualifier they state themselves: it’s correlation. The markup stopped paying while pages carrying visible FAQ content kept getting cited. It doesn’t prove that bolting an FAQ onto a thin page earns citations.

llms.txt files. Covered above. Pressed on why Chrome’s Lighthouse tool checks for one when Search ignores it, John Mueller, a Search Advocate at Google, called it a temporary crutch for AI coding tools reading developer documentation, not something non-developer sites need to think about. Your practice site has a dozen pages and no API docs, so there’s no demonstrated search-visibility benefit and not much for the file to do.

If your platform generates one automatically, glance at what’s in it occasionally. The largest research crawler hitting these files in the Ahrefs dataset identified itself as prompt-injection-survey/1.0, which tells you what some people are studying the format for.

06How to Measure GEO When Your Analytics Cannot See It

This is the part most competing posts skip, because the honest answer is partly “you can’t, yet.”

Here’s what actually exists.

Search Console has a Generative AI performance report, but read the small print. Google announced it on June 3, 2026. It shows impressions inside generative AI features across Search and Discover, broken out by page, country, device and date.

Several limits matter more than the announcement. Google says it went to a subset of websites, and contemporary reporting adds that the subset was UK-first, with a further rollout stated but undated. Many practices won’t have it yet. It also carries no click data, so it tells you how often you appeared, not whether anyone came.

Two things to know before reading the numbers. Data starts May 18, 2026 with no backfill, so there’s nothing earlier to compare against. And a logging error has depressed reported impressions from August 13, 2026 onward, which Google has said is a reporting problem rather than a real drop. If you see a dip this month, that’s likely why. Don’t let anyone bill you for reporting out of a panel they can’t access.

Bing Webmaster Tools has an AI Performance report, and it’s free. Microsoft announced it on February 10, 2026. It shows total citations, average cited pages, page-level citation activity, and grounding queries, which are the phrases the AI generated internally when it went looking for content. It covers Microsoft Copilot, AI summaries in Bing, and selected partner integrations. Microsoft expanded it on June 16, 2026 with intents, topics, citation share, and comparison.

Grounding queries are the most useful thing on this list and almost nobody in healthcare marketing uses them. They tell you the question the machine asked, not the one the human typed. The gap between those is where your next five blog posts live.

Then do the manual thing. Once a month, ask ChatGPT, Perplexity, Gemini and Copilot the five questions a prospective client would ask before booking with someone like you. “Who treats autistic adults in Chicago?” Note whether you appear, who does, and what the assistant says about you. It takes twenty minutes and it’s the simplest no-cost way to see the actual output. Paid tools will track this automatically if you’d rather not do it by hand.

Set a reminder and keep the answers somewhere. The trend over six months beats any single check.

The numbers behind what earns a citation live in my backlink and AI citation statistics.

07Where This Advice Runs Out

Three honest limits.

The Ahrefs schema study only looked at pages already receiving 100 or more AI citations. It can tell you schema won’t push a well-cited page higher. It can’t tell you whether schema helps a page no AI system sees at all.

The SE Ranking work is correlation, not a controlled experiment, and its effects are tangled together.

And this moves fast enough that a confident post from 2026 will look silly by 2027. Google shipped its generative AI guide on May 15, 2026, clarified llms.txt on June 15, and updated the page again on July 10. Anyone selling a permanent playbook is selling a snapshot.

What I’d bet on holding is the mechanic underneath. Machines quote things that are specific, sourced, self-contained and attributable to somebody real. The 2024 academic work found that by testing edits to content. The 2026 data found it by measuring 129,000 domains. Two different methods, two years apart, same answer.

So the practical instruction is boringly old-fashioned. Write about what you know, be specific enough to be checkable, cite the code or study you mean, and make each section stand on its own. None of that is GEO as a separate discipline. It’s good writing, graded by something new.

The practices cited in 2027 won’t be the ones buying AI packages this year. They’ll be the ones who wrote useful pages in 2024 and kept them current.

Frequently asked questions

Sources