GEO vs SEO: what the research actually says
Every agency now sells generative engine optimization on a 40% figure. The survey that collected the evidence rejects that number as a general claim, and found only three of 54 tactics that transfer across domains.
Type "geo vs seo" into Google and it offers you "geo vs seo vs aeo", "geo vs seo vs aio", and "geo vs seo reddit". The industry selling you this has not settled on what to call it. That is worth holding onto while you read the pitches, because the confident number in most of them comes from a single 2023 paper, measured under conditions that do not describe your website.
What is GEO, and how is it different from SEO?
SEO tries to place your page in a list. GEO tries to get your page quoted inside an answer that may never show a list at all.
That difference has a cost attached. Where an AI Overview appears, organic click-through fell 61%, and Pew found people clicked a traditional result in 8% of visits with a summary against 15% without. Informational searches now end without a click 74% of the time, while transactional ones sit at 31%. Question-shaped searches trigger an AI Overview about 86% of the time.
So the honest framing is that a large part of what used to be traffic has become mention — and a mention is worth something different to your business than a visit. It builds recognition with someone who may arrive weeks later by typing your name, which is harder to measure and slower to arrive than a click.
One piece of good news sits underneath that. Roughly 53% of the domains cited in AI Overviews never appear in the organic top ten for the same query. You do not have to outrank an established publisher to be quoted alongside it — which is the first thing in a decade that has favoured a small site over a large one.
Does GEO actually work?
Sometimes, in ways that depend heavily on the page, the question and the engine. Anyone telling you otherwise is selling.
The foundational GEO paper measured real gains. Adding quotations moved a source's citation share from 19.3% to 27.2%, which is the 41% relative lift the industry rounds to "40%". Adding statistics and citing sources landed in a similar range. Keyword stuffing scored worse than changing nothing.
The catch is the setup. Those numbers come from documents already placed in the model's context — five of them, fixed — so the paper measures which of five already-retrieved sources gets quoted. It does not measure whether your page gets retrieved in the first place, which is the part you actually need. An agency quoting that figure at you is describing a race you have not yet qualified for.
When researchers tested the same tactics across different domains and task types, the results mostly evaporated. C-SEO Bench found three of 54 method-domain combinations significantly positive — which is close to what you would expect from chance alone. The survey's conclusion about the field as a whole is blunter than anything you will read in an agency deck: no reviewed technique shows a stable, longitudinal, cross-platform causal effect.
That does not make the work pointless. It makes the work specific. What transfers between domains is thin, so the useful question stops being "what is the GEO checklist" and becomes "what does this engine quote for the questions my customers ask" — which you can only answer by asking it.
Which tactics have evidence behind them?
| Tactic | What was measured | How much to trust it |
|---|---|---|
| Relevance to the question | Dominates across 252,000 trials | Strongest finding in the field |
| Position in the model's context | Moving a source higher beats most rewrites | Strong, and mostly outside your control |
| Adding quotations | 19.3% → 27.2% citation share | Real, but inside a fixed context |
| Adding statistics | Similar range | Real, same caveat |
| Citing your sources | Similar range | Real, same caveat |
| Structure and headings | "Moderate and heterogeneous" | Test it, do not assume it |
| Schema markup | Not required for AI features | Useful for rich results, not for this |
| Keyword stuffing | Below baseline | Actively harmful |
| llms.txt | Ignored | Does nothing |
Read down that table and a pattern shows up. The tactics that survive are the ones that make a page genuinely more useful to quote — a real number, a real source, a direct answer. The ones that fail are the ones that try to signal quality rather than have it.
There is a wrinkle worth knowing if you are already ranking well. In the original study, citing sources gave a fifth-ranked page a large visibility gain and cost the first-ranked page visibility — the same edit, opposite outcomes, decided by where the page already stood. Any advice that does not ask where you currently rank is advice given without the one variable that changed the sign of the result.
Why does a page win on Google and vanish in Perplexity?
Because these systems are not reading the same shortlist. Bing Chat and Perplexity overlap on only about 26% of cited domains. Google's AI Overviews overlap with Google's own organic results at a Jaccard similarity of 0.11 to 0.18.
There is no single ranking to win, so "we will get you into AI search" is not a coherent promise. Visibility is specific to an engine and a surface, and a vendor who cannot tell you which one they measured has not measured anything.
The other structural fact is concentration. Reddit alone accounts for roughly 40% of citations among the most-cited domains, rising to 46.7% on Perplexity, and the top fifteen domains take about 68% of consolidated citation share. What these systems appear to reward there is discussion depth rather than link authority — a thread where five people argue about a product reads as evidence in a way a product page does not.
The practical reading is uncomfortable for anyone selling on-page work. A genuine, non-promotional presence where your customers already discuss their problems probably does more for AI visibility than most of what happens on your own site. Spamming those places gets your domain banned — which is worse than never having posted, because the ban follows the domain rather than the account.
What does Google itself say to do?
Google published guidance for its AI features, and it is unusually direct about what not to bother with. You do not need llms.txt; Search ignores it. There is no AI-specific schema. There is no requirement to break your content into tiny pieces. Seeking inauthentic mentions "isn't as helpful as it might seem".
What it asks for is "non-commodity content" with "a unique point of view" — first-hand expertise rather than a summary of what other people already wrote. Measurement lives in the Generative AI performance report in Search Console, which is where you should be looking instead of at a vendor's dashboard.
That advice is inconvenient for the GEO industry, because it says the lever is having something to say.
What should you actually do?
Four things, in the order they pay off.
Check you are not blocked. GPTBot, OAI-SearchBot, PerplexityBot and ClaudeBot each need to reach your pages. Plenty of sites quietly block them through a CDN setting nobody remembers enabling — and no amount of writing fixes that. It takes about a minute to test with a command-line request per crawler, and it is the only item on this list that can silently cancel all the others.
Answer one real question per page, in its first sentence. These systems quote passages rather than pages. A section that only makes sense after the three above it is a section that cannot be lifted — so "as we saw above" is a phrase with a measurable cost.
Put one number in that is yours. A measurement from your own work, with a date and a method, is the single most quotable thing a small business can publish — and it is the one thing a competitor cannot copy off you. Ten sites can write the same guide. None of them can write your measurement.
Stop paying for the 40%. If a proposal quotes that figure without naming the fixed-context condition it was measured under, the people writing it have not read the paper. That is a reasonable thing to ask them about before you sign.
None of this is fast. Search brings first movement at three to six months. What you are building is the chance of being the source an answer cites, and that compounds in a way a rented click never did.
Frequently asked questions
Is GEO replacing SEO?
No. The technical work overlaps almost entirely — crawlability, speed, clear structure. What changes is the outcome you are optimising for: being quoted inside an answer rather than listed beneath one.
Do I need an llms.txt file?
No. Google states that Search ignores it. Some smaller tools read it, but no major engine has published evidence that it changes citation.
Does schema markup help with AI search?
Not directly. Google says structured data is not required for its generative AI features. It still earns rich results in classic search, which is reason enough to keep it.
Is the 40% improvement figure real?
The measurement is real; the claim built on it is not. It describes which of five already-retrieved documents gets quoted, and the 2026 survey rejects it as a general statement about visibility.
Should my business be posting on Reddit for this?
Participating honestly, yes — it is the most-cited domain across every major engine. Posting promotional content, no. Reddit bans domains for it, and a banned domain is worse than an absent one.
How do I measure whether any of this worked?
Search Console's Generative AI performance report for Google surfaces, and manual prompt checks on ChatGPT and Perplexity for the rest. Because the engines cite different sources, one number across all of them would be a number about nothing.
Published 25 September 2026. Sources: the 2026 arXiv survey of generative engine optimization, the original GEO paper, C-SEO Bench, Google Search Central's AI features guidance, and Pew Research.