Generative engine optimization, usually shortened to GEO (AKA AIO), is the practice of making a company’s content easy for an AI answer engine such as ChatGPT, Claude, Gemini, or Perplexity to retrieve, quote, and cite by name, rather than easy for a search engine to rank. Whether you need it comes down to one question, and it is not “is this the future of search.” It is: do your buyers already ask an AI tool about your category before they find you any other way?
Where the term comes from, and whether it is real
GEO is not marketing language invented by a vendor. Aggarwal and colleagues coined it in a 2023 paper by a Princeton-led team that measured how changes to content affected a brand’s visibility in generated AI answers, much like “SEO” describes tactics aimed at a ranked list. The term has since been picked up widely enough that Google Search Central felt the need to respond to it directly: in May 2026, Google published guidance stating that AEO and GEO are not a separate discipline from SEO, and that neither requires an llms.txt file, special schema, or AI-specific content rewriting to be recognized.
We point to that guidance rather than bury it, because most of the category will not tell you this. Roughly 80% of what actually moves AI citation is ordinary SEO fundamentals: crawlable pages, clean structure, real third-party authority, and content that answers a question directly in its first sentence instead of working up to it. We wrote the longer version of that argument in a separate post in September.
People are searching for the term itself, regardless of what you make of the underlying claim. “Generative engine optimization” pulls roughly 4,400 Google searches a month as of mid-September, and “answer engine optimization” pulls about 2,400. Both numbers have moved substantially over the trailing year, which is consistent with a term still finding its footing rather than one that has already peaked.
The 20% that is not SEO
Two things do not reduce to SEO ranking a page.
The first is entity resolution. A search engine ranks pages. An answer engine has to be confident about which company it is looking at before it will name you at all, and an engine that is unsure doesn't throw an error. It either skips you or describes somebody else, and both look identical from the outside. We learned this the expensive way: we describe what we sell as a managed agent, and that phrase collides with hosted-runtime products from Anthropic, Google, and IBM. No volume of content about managed agents would have fixed that. The fix was changing the frame, not writing more.
The second is measurement. A search rank is stable and ordinal. An AI answer is a distribution. In a 2025 AirOps study, only about 30% of brands stayed visible across back-to-back runs of an identical query. The same brand can appear in one run and vanish from the next, with nothing changed on either side. A single screenshot of an engine naming a competitor tells you almost nothing on its own. Reading this correctly requires repeated sampling on a fixed question set, reported per engine rather than blended into one score, with the sample size stated every time.
Whether you need it
The real test is not whether GEO is real. It is whether the question even comes up for your buyers yet.
Buyers now research with AI tools before they talk to a vendor, and that shift moved fast enough that most companies’ content was written for a world where it had not happened yet. If that describes your buyers, the fundamentals in the first section are worth doing regardless of what you call them: crawlable, well-structured, sourced content was good practice before this term existed and stays good practice after it fades. The GEO-specific work, entity resolution, and repeated-sample measurement are worth adding once you have confirmed a real, measurable gap, not just a feeling that you should be doing something new.
Chasing every tactic sold under this label isn't worth it. The clearest example is the llms.txt file, a proposed standard for telling AI crawlers what to read. A June 2026 Ahrefs study of 137,000 domains found that among the roughly 38,000 sites that had implemented a valid file, 97% received zero requests for it in May 2026. Ahrefs found the closest thing it has to an audience is coding agents, not the answer engines it was built to influence. We have one. It is correctly formatted. It has not moved our own numbers, which is exactly what that study would predict.
Current view, subject to change
We think the accurate framing is that GEO is real as a set of measurable outcomes and mostly not real as a new discipline requiring new tools. The evidence for tactics that lift a brand’s prominence once it is already inside an engine’s retrieved context is decent. The evidence that any tactic reliably gets a brand into that context in the first place is thin, and the evidence for durable traffic or revenue lift from any of this is thinner still.
The single most useful number we have found on this comes from a controlled study cited in a July 2026 survey: on a site where some pages received an AEO intervention, total ChatGPT referrals to the site rose 5.7 times over. But untreated pages on the same site rose 3.5 times over on platform growth alone, meaning most of that lift was the tide, not the intervention. The controlled estimate of the intervention’s effect came out to a 1.8x multiplier, and the survey calls that result suggestive rather than causally established. We would drop this framing if a company running competent SEO with no entity or measurement work ended up cited at the same rate as one that did both. We haven't seen that study run, and we haven't run it ourselves yet either.
Final thoughts
We measure ourselves on exactly this set of questions every week and publish the number whichever way it falls. As of the week ending October 4, we have zero citations across our own tracked question set: thirteen straight weeks on the original set, then zero again in all four readings since we reset to a new set of questions in September. We are not naming that number to be self-deprecating. We are naming it because a category full of unverifiable lift charts needs at least one participant willing to show the losing weeks, not only the winning ones.
So: GEO is a real, measurable set of outcomes wearing a name that is two years old and still finding its footing. You need it once your buyers are already asking, not before.
Regards,
Charles Stack
Founder, Coworkers.Global
Cite this: “Generative engine optimization (GEO) was coined by Aggarwal et al., Princeton, 2023, and pulls approximately 4,400 Google searches per month as of September 2026.” Source: this post. Search-volume figures captured via DataForSEO, September 14, 2026. Study citations: Aggarwal et al. 2023 (GEO origin); Google Search Central, May 2026 (AEO/GEO guidance); Ahrefs, June 2026 (llms.txt adoption study, 137,000 domains); Martinez, arXiv:2607.14035, submitted July 15, 2026 (critical survey of GEO, citing Watanabe and Nakayashiki 2026 for the controlled-lift study).