Does GEO actually work, or is it hype?

The evidence supports GEO for narrow questions; results for generic questions remain unproven. Three kinds of evidence are available: a research paper using a synthetic benchmark, large studies of what assistants cite published by SEO tool vendors alongside the platforms’ own documentation, and small experiments on sites such as ours. All three support work on pages that answer specific prompts. None establishes what buyers often want most: a recommendation for “the best” provider in a category full of established names.

Two terms before the evidence. Citation rate is the share of a fixed set of buyer prompts where an assistant links to one of your pages. Share of voice is your share of brand mentions on that same set, against the competitors named beside you. Generative engine optimization (GEO) is the work of moving those two numbers.

The evidence supports changes to content and crawler access for specific prompts. It does not yet establish broader gains in brand visibility. The table assesses how useful each source is when deciding whether to invest.

Three things, and a buyer should hold on to all of them. First, Google’s position is documented: its AI Overviews and AI Mode draw on the same index as classic search, require nothing extra, and no AI-specific file or schema changes eligibility. If a vendor’s GEO package for Google surfaces is llms.txt and “AI schema”, the skeptic is right and the package is housekeeping. Our own guide on llms.txt says the same.

Second, most large-sample studies are published by companies that sell SEO or GEO services, ours included. That does not make them wrong, and Ahrefs in particular publishes its caveats, but it means there is no independent index of GEO results yet. Third, generated answers vary between runs. In our August 2026 runs of one prompt family in ChatGPT, twelve different competitors appeared beside us across five clean sessions. A single screenshot proves nothing, in either direction.

Where the skeptic overreaches is in concluding that nothing can be measured. A fixed prompt set, run several times in clean sessions every few weeks, produces citation rate and share of voice with enough repetition to see movement. The variance is a reason to sample more, and the sampling is what a measurement protocol is for.

get-geo.ai went live in June 2026 with the technical layer mostly in place: server-rendered text, language tags (hreflang) across eight languages at launch, structured data, llms.txt. On August 4 an audit of our own site found a redirect loop between the bare domain and www, the kind of thing Google’s crawler tolerates and AI fetchers, which often give up after one or two requests, frequently do not. We fixed it that day.

On August 13 we asked ChatGPT in five clean sessions to recommend a GEO agency for a company working in several languages, varying the phrasing. It named get-geo.ai in five of five. On August 17 a sixth run, in English and Spanish, did the same. From launch to the first citations we could reproduce on those prompts: roughly two months.

The limits, in our own words. Six runs across two days is evidence of a first result, and the prompts fit our positioning unusually well, so this is the narrow-prompt case working and nothing more; it says nothing about the generic “best GEO agency” prompt, where established names still dominate. The rest of the limitations are in the case study and in our guide on how long GEO takes.

These four situations account for most of the cases where prospective clients tell us GEO produced no results.

You do not need to buy anything to find out. Pick ten prompts your buyers would actually type, in one language, and run each three times in a clean session (logged out, or a temporary chat with memory off) in one assistant. Record who is named, in what position, and which pages are cited. That is your baseline, and it takes an afternoon. Our guide on how we measure GEO results describes the full version.

Then fix access: fetch three important pages with an AI crawler’s user-agent, or simply view the page source, and make sure the text is there without JavaScript, on one host, with the crawlers allowed. Write one page for the prompt where you are furthest from being cited: the prompt as the headline, a 40–60 word direct answer, then the detail a model would verify. Wait three weeks, run the same ten prompts the same way, and compare.

If the new page is cited on its prompt, GEO works for you at the level the evidence supports, and the question becomes how many prompts are worth a page. That is also the only honest way to answer “is GEO worth it”: GEO ROI is the referral traffic and share of voice on prompts you chose, against the cost of the pages, rather than a percentage from a vendor deck. If it is not, the first suspect is still access, and the second is that the page answers a question nobody asked. Either way you now hold a baseline, which is more than most proposals will offer you. Our free AI visibility audit is this test, done for you, on ten prompts.