GEO vs SEO: What Actually Changed, and What Is Just a New Word
Three acronyms arrived in roughly eighteen months. GEO, for generative engine optimization. AEO, for answer engine optimization. LLM SEO. Most agency decks now carry at least one of them, usually priced as a separate line from the SEO retainer that was already running.
They describe the same shift. People increasingly receive an answer rather than a list of links, and that answer names a handful of sources. The vocabulary has not settled. The work underneath the three names has, and a good share of it is work you were already doing.
So the useful question is not which acronym to buy. It is which parts of your existing SEO carry straight over, and which four things actually broke.
Key Takeaways
GEO, AEO and LLM SEO are three names for one shift, not three disciplines to staff separately.
Four things did change: the retrieval unit, crawler access, how you measure, and what counts as an outcome.
Ranking well no longer implies getting cited. Ahrefs put the overlap at 12% across 15,000 prompts.
The most common technical failure in this whole category costs nothing to fix and lives in a single file.
Why This Stopped Being a Niche Concern
Two numbers explain the urgency better than any vendor deck.
SparkToro's analysis of the first four months of 2026 found that 68% of Google searches ended without a click. Ahrefs measured that when an AI Overview appears, click-through rate for the top organic result drops by 58%.
The second number matters more than the first. Zero-click search is not new, and featured snippets have been eating clicks for years. What is new is that the answer now names sources, and being one of them is a different contest from ranking first.
What Actually Changed
The unit is the passage, not the page
An AI answer is not written from your page. With search enabled, these systems run a query against a live index, pull back a handful of documents, cut them into chunks, and hand a few of those chunks to the model as context. The model writes out of that context and links to whichever chunk it leaned on.
The consequence is awkward for anyone who has spent a decade optimising pages. A page ranking first with its answer buried in paragraph nine loses the citation to the page ranking seventh that put a self-contained answer directly under a heading. Position is an input to retrieval. It is not the decision.
This is also why heading structure stopped being a formatting preference. Headings are the most reliable chunk boundary available to a retriever. Two thousand words under two headings produces a few enormous passages, each covering several topics and each diluted against any specific query. The same two thousand words under eight question-shaped headings produces eight tight ones, each able to win a different question.
Crawler access became a second, separate gate
This is the one new failure mode, and it is binary.
Googlebot access does not imply OAI-SearchBot access. A site can be indexed cleanly by Google, ranking respectably, and be completely absent from ChatGPT search because of one line somebody added to robots.txt in 2023.
The confusion that causes most of the damage is that GPTBot and OAI-SearchBot are different crawlers doing different jobs. GPTBot collects text for model training. OAI-SearchBot builds the index that ChatGPT searches. Plenty of sites disallowed GPTBot on principle, which is a defensible policy choice, then copied the same block onto every OpenAI-looking user agent and quietly removed themselves from ChatGPT search. The same pair exists at Google, where Google-Extended is the training opt-out and blocking it does not remove you from AI Overviews.
There is a second trap in the file format itself. Under RFC 9309, a crawler obeys exactly one group: the one whose user-agent token is the longest match on its name. It ignores every other group in the file, including User-agent: *. So the moment you add a named group for PerplexityBot, your wildcard rules stop applying to it at all.
Two more failures show up constantly. A Disallow: /*? written years ago to keep faceted URLs out of Google now blocks every AI crawler from every URL carrying a query string. And bot-manager rules at your CDN are invisible in the file entirely, so robots.txt can say Allow while your edge returns 403.
There is no single result page to check
Ask the same question twice in the same product and the citations can differ. Ask it in ChatGPT, Perplexity and Google's AI Overviews and you get three different sets.
There is no rank number here. What you can measure is presence across repeated runs, which means the instrument is a sample rather than a screenshot. A single check tells you nothing, and a fair number of the dashboards sold on this premise are selling you a sample size.
Clicks stopped being the only outcome
A cited answer frequently ends the session. Someone reads that your product does the thing, forms a view, and never visits.
That is hard to sit with, and it is also why brand-level mentions inside answers matter now in a way impressions never did. The mention is the outcome. Judging it against session data will make a win look like a failure.
What Did Not Change
Most of it, which is the part the acronyms obscure.
ChatGPT search and Perplexity pull their candidates from web indexes. The pages they cite are usually pages that already rank somewhere. A page that cannot rank rarely gets quoted either. Being in an index, having pages that load, sensible internal links, saying something nobody else says: all of that carries over unchanged.
Ahrefs' finding that only 12% of AI-cited URLs also rank in Google's top ten for the same prompt gets quoted as proof that SEO stopped mattering. It shows something narrower. The top ten is not the retrieval pool, but those cited URLs still came out of an index, and they still had to be findable to get there.
How to Check Where You Stand Without Buying Anything
Work in one order: access, then structure, then citability. Each one makes the next measurable.
Access
Read your robots.txt line by line against each crawler's user agent, then check server logs for arrivals by name. Arrivals are ground truth; the file only records intent. Tracetify, an SEO tool that traces how competitors got their first users, publishes a free checker that fetches your robots.txt and evaluates it per bot rather than per wildcard, which settles the intent half in about a minute and needs no account.
Structure
Fetch the URL without a browser and confirm the text is present in the HTML before any JavaScript runs. Then read your heading outline on its own and ask whether each section would still make sense pulled out and read alone.
Citability
Write down fifteen to twenty questions your buyers actually ask. Run them monthly across ChatGPT, Perplexity and Google, and log which domains get cited. Run each prompt more than once, because output varies between runs.
When It Is Worth Paying Someone
Published GEO retainers start around $1,000, and most programmes that move anything sit between $4,000 and $15,000 a month.
That is money well spent when your gap is authority, meaning almost nobody outside your own site mentions you and the fix is earning third-party coverage you cannot manufacture internally. It is poor value when your gap is technical, because a blocked crawler, a JavaScript-only page and a two-heading article are all things your own team can fix in an afternoon.
Run the three checks above before taking any calls. They cost nothing, and they tell you which conversation you are having.
Frequently Asked Questions
Is GEO different from AEO and LLM SEO?
Not in practice. All three describe optimising for AI-generated answers instead of ranked links. Some practitioners draw a line between winning one specific answer and being chosen as a source at all, but the work underneath is the same: let AI systems fetch your pages, and give them something worth quoting.
Does blocking GPTBot hurt my Google rankings?
No. GPTBot is OpenAI's training crawler and has nothing to do with Googlebot or with ChatGPT search. What does cost you visibility is blocking search-time crawlers by accident, which is exactly what a firewall rule aimed broadly at "AI bots" tends to do.
Do I need an llms.txt file?
Not urgently. It is a proposed convention for publishing a curated map of your site for language models, and it costs about an hour to add. No major AI search product has documented using it for retrieval. Fix crawler access and page structure first, because those are read today.
Is traditional SEO dead now that AI answers questions?
No, and the two are difficult to separate. The pages AI systems cite are overwhelmingly pages that already rank somewhere. This work extends what you were doing rather than replacing it.
Conclusion
The three acronyms are marketing. The four changes underneath them are not.
If you take one thing from the comparison, make it the order of operations. Crawler access is binary, costs nothing to check, and renders every other improvement worthless when it fails. Structure comes second, because it decides what a retriever is even able to hand the model. Only after both of those does the content question become worth real money: why would a model name you rather than someone else.
Most of what is currently sold as a new discipline is that third item, priced as though the first two had already been handled.
