GEO is barely 18 months old: how to tell good AI-search advice from hype

If you are asking “is GEO worth it,” start with a date. The term Generative Engine Optimization first appeared in a research paper submitted to arXiv on November 16, 2023 by a Princeton-led team. The job most people mean by GEO, getting your brand named inside a live ChatGPT or Perplexity answer, is younger than that. So when someone sells you a “proven, multi-year GEO system,” check the math. The surfaces it claims to have mastered mostly shipped in the last 18 months.

That is not a reason to dismiss GEO. The demand is real. It is a reason to read AI-search advice the way you would read advice about any young field, with interest and a filter. This post hands you the filter. You get the actual timeline with sources, a clear line between what holds up and what is hype, and an honest look at how I run GEO at BlueJar. I build this stuff for a living, and I will tell you which parts are still guesses.

How old is GEO, really

There are two clocks here. The academic one and the practitioner one.

The academic clock started in late 2023. The paper “GEO: Generative Engine Optimization” was submitted to arXiv in November 2023 and later published at the ACM SIGKDD conference in 2024. It coined the term, introduced a benchmark called GEO-bench, and reported that certain optimization methods raised a source’s visibility in generated answers by up to 40% in their tests. That is where the word comes from.

The practitioner clock, the one that matters if you are about to spend money, started later. Getting cited inside a live AI answer was barely a job until the engines could read the live web and put answers in front of millions of people. Those catalysts landed through 2024 and 2025. Count from there and the working practice is roughly 18 to 24 months old.

The timeline, with sources

Here is the sourced version, so you can date any claim you hear against it.

Look at that last row. A core surface of AI search rolled out about a year ago. Any “proven system” for it has, at most, a year of evidence behind it, and the engines have changed their behavior several times since.

Why nobody has a multi-year track record

The honest version is simple. The engines that decide whether your brand gets named are moving targets. ChatGPT, Perplexity, Gemini, and Copilot all change how they retrieve and rank sources on a regular cadence, sometimes every few weeks. A tactic that earned citations in one quarter can stop working the next, not because you did anything wrong, but because the engine changed which sources it pulls.

I have watched this happen on our own pages. A format that was getting BlueJar named in answers for a cluster of queries quietly stopped, and the content had not changed at all. The engine had. That is the field you are buying into.

So when you read confident GEO advice, including mine, know what it actually is. A set of educated guesses, refined over months of testing, that are probably directionally right. That is genuinely useful. It is not the same as a law of physics, because the field is too young to have laws yet.

I will say this plainly, because it is the point of the whole post. Our playbooks at BlueJar are guesses too. Good guesses, tested against hundreds of prompts per audit and updated when the data moves. But the moment any vendor, us included, tells you their system is “proven” with the confidence of a decade-old discipline, your skeptic filter should switch on.

Tells of hype vs what is real

Skepticism is not cynicism. The goal is not to reject GEO. It is to catch the overclaims and keep the parts that hold up. Here is the side-by-side I use.

The SEO overlap, and the part that is new

The most common skeptic line is “GEO is just rebranded SEO.” It is half right, and the honest answer is more useful than either the hype or the dismissal.

Concede the overlap first, because it is true. What you optimize for AI answers is close to what good SEO already asks for: clear answers high on the page, structured content, consistent entity information, citation-ready statistics, real backlinks, schema. If you have done citation-readiness work, you have already done most of the on-page part of GEO. Do your SEO first. Most “GEO courses” that are really repackaged SEO are not wrong about the fundamentals. They are just overpriced for what they teach.

Now the part that is actually new, and it sits on the measurement side. In SEO you check a search results page and get a clean rank: position 4, done. In GEO, the same query asked twice can return different answers and different sources, because the engines are non-deterministic. A single check is a vanity metric. To know whether anything worked, you need a panel of prompts run across variations and across engines, then you report consistency or a range, not one number. That measurement discipline, plus the off-site work of getting placed in the third-party sources AI keeps citing, is the genuinely new workstream. The deeper GEO-versus-SEO breakdown walks through why the engines diverge.

One more distinction SEO cannot make, and it is the one I find most useful. You can rank number one organically and still be invisible the moment someone asks the query inside an AI tool. Rank is not citation. The click your ranking was built for never happens if the reader gets an answer and never scrolls to a blue link.

What is worth doing right now

None of this means wait. It means act like a scientist, not a shopper. Here is the short list of things that are genuinely real and worth doing today, whichever way the engines drift next.

• Do the basics. Clear, direct answers near the top of each page. Structured, scannable content. Consistent entity information across the web. These help in AI answers and in classic search, so the downside is close to zero.
• Get mentioned where AI pulls from. Engines lean on a small, repeatable set of third-party sources per topic: comparison posts, roundups, community threads, review sites. Being present in those sources is a parallel job to your own pages. See how to get cited by ChatGPT for the off-site side.
• Measure before and after, with discipline. Run a panel of real prompts across engines, record what names you and what does not, then re-run after you change something. One spot check tells you nothing.
• Test on your own brand first. Do not trust a generic playbook on faith. Run it on your site, measure what actually moves, keep that, drop the rest. The cost of testing is low and the learning is yours.

That is the whole method. Test, measure, keep what works. It is boring on purpose, and boring beats confident-but-unproven in a field this young.

How we approach this at BlueJar

We built BlueJar around the experiment-first idea, because the alternative, selling certainty in an 18-month-old field, is the exact thing this post is warning you about. An audit runs a structured set of 400-plus prompts across ChatGPT, Perplexity, Gemini, and Copilot and reports where you are named, cited, or invisible, per zone and per intent. It is a point-in-time measurement on purpose. We do not pretend a single number stays true forever, and we tell you to re-run it on a cadence instead of treating it as a one-time verdict.

It will not guarantee you a citation, because nobody can. What it does is swap guessing for measurement on your own brand, which is the one move that survives every engine update. If you want to compare what is on the market before deciding, our roundup of GEO audit tools is a reasonable starting point, and the state of AI search piece has the adoption context.

Want to separate real from hype on your own site? Run a free GEO audit at app.bluejar.ai. It is a point-in-time measurement across ChatGPT, Perplexity, Gemini, and Copilot, so you can see what is actually true for your brand today before you trust anyone’s playbook, mine included.