Most GEO is a scam. Here is the part that is not.

Type "is GEO a scam" into a search box and the suggestions finish the sentence for you. Just SEO rebranded. Snake oil. Nobody measuring it. The honest answer is that most of what is sold under the name deserves those suggestions. It is ordinary technical SEO under a new invoice, or a tactic with no evidence behind it, or a dashboard charting noise. A narrow set of work survives scrutiny. We sell that set and nothing else. The rest of this article is the line between the two.

One caveat. We sell GEO, so read this with more suspicion than usual. An agency explaining why its category is mostly fraudulent but its own corner is fine has an obvious interest. The only defence is to show the evidence and let you check it.

Is GEO a scam? For most of what is sold, yes

Each criticism is at least partly right.

It is mostly SEO with new column headers

Every AI answer surface retrieves from a crawled index. Google's AI Overviews and AI Mode use the same index and Googlebot as ordinary search. Copilot grounds on Bing. Perplexity and ChatGPT run their own crawlers over the same web. A page that cannot be crawled cannot be cited, and making one crawlable already has a name. We wrote the long version in is GEO replacing SEO. The short version is that a "GEO audit" which finds broken redirects and blocked crawlers is a technical SEO audit with the title changed.

None of that is fraud. The problem starts when the same work is priced as if it were new.

The signature tactics have been tested and failed

This is where the category earns the word scam. Schema markup is the loudest example. Ahrefs ran a matched difference-in-differences study on 1,885 pages that added JSON-LD against 4,000 control pages. AI Overview citations fell 4.6 percent, a statistically significant decline. AI Mode moved 2.4 percent and ChatGPT 2.2 percent, both indistinguishable from zero. Schema still earns conventional rich results. It does not cause AI citations, and anyone who says otherwise is quoting a correlation the controlled study already explained.

llms.txt is the second. Google's own guidance says no special files, no llms.txt and no special markup are needed, and that Google Search ignores the file. FAQPage rich results ended in Google Search on 7 May 2026. speakable never left a news-only beta. All three fit neatly on an invoice. None changes whether an assistant names you.

The dashboards measure noise

LLM outputs are non-deterministic. Sampling maths puts the margin of error on a five-repetition run at around plus or minus 27 points, and a defensible measurement needs 12 or more repetitions per prompt. Most vendor dashboards run each prompt once and chart it as a trend. A line that climbs on that basis is a chart of dice rolls.

The traffic is small and badly attributed

Conductor's benchmark puts AI referrals at roughly 1 percent of traffic. Ahrefs published its own numbers: 0.5 percent of visits, though 12.1 percent of signups. Much of that traffic arrives with no referrer and lands in analytics as Direct, so nobody has a clean read on it yet. A vendor who reports precise AI revenue is reporting a guess with a currency symbol in front of it.

The part that is not a scam

Four things hold up. They are unglamorous, which is why they are undersold.

Crawler access

There are two kinds of AI crawler and most robots.txt files confuse them. GPTBot and ClaudeBot are training crawlers. OAI-SearchBot, Claude-SearchBot, PerplexityBot and Bingbot are answering crawlers, the ones that fetch your page so it can be cited live. Block an answering crawler and you are removed from that surface entirely. Google-Extended is the common trap: blocking it does nothing to AI Overviews or AI Mode. The full breakdown is in the AI crawlers and robots.txt guide. Checking takes ten minutes and it is the highest-confidence item on the list.

Rendering

No major AI crawler executes JavaScript. Vercel and MERJ analysed more than 500 million fetches and found GPTBot, ClaudeBot and PerplexityBot fetched JavaScript files and never ran them. Only Google's AI surfaces and Copilot render. If your menu, hours or services are assembled in the browser, ChatGPT sees an empty shell. The same analysis found 404 rates near 34.8 percent for ChatGPT's crawlers and 34.2 percent for Claude's, against 8.2 percent for Googlebot. Stale URLs and redirect chains eat a third of what those crawlers try to read.

Answer-first pages

Search Engine Land instrumented ChatGPT's retrieval and found the stored snippet for a page is about 200 characters, taken from the text near the H1, frozen at index time and identical whatever the query. Pages over 4 MB are rejected outright. So the paragraph after your H1 carries the page. It should say what you are, where you are and who you serve, with no warm-up sentence. This is copywriting, and it is the only content work we have seen evidence for.

Mentions on sites you do not own

Ahrefs studied 75,000 brands and found branded web mentions correlated with AI Overview visibility at 0.664, against 0.218 for backlinks. Muck Rack analysed more than 25 million citations and found earned media accounted for 84 percent, paid and advertorial for 0.3 percent. Both are correlational and neither proves cause. But they point the same way, towards the largest input to being recommended sitting outside your website.

One caution from Semrush's ghost citations study: 61.7 percent of citations produce no brand mention. ChatGPT cites a source 87 percent of the time and names a brand 20.7 percent of the time. Being linked and being recommended are different outcomes, and only the second pays.

What we refuse to sell

The last row deserves a sentence. The story that Foursquare powers ChatGPT local results circulated widely and no test supports it. A vendor still selling it has told you how they check claims.

Questions that separate the two

Ask any GEO agency, ours included, these five.

1. How many repetitions per prompt does your measurement use? Under 12 is a guess.
2. Which first-party report do you read? Bing Webmaster Tools has shown Copilot citations and grounding queries since February 2026. Google Search Console's Generative AI report, global since August 2026, shows impressions only. A vendor who names neither is working from their own tool. How to track AI search visibility covers what each can tell you.
3. What is the first change you would make to my site? If the answer is a file or a markup type, end the call.
4. What is the difference between a citation and a mention? A pause means they have not read the studies.
5. How long until I appear? Nobody knows, because there is no index status to check.

What this costs if you do it yourself

The honest part is a robots.txt review, a JavaScript check, a rewritten opening paragraph on each service page, a crawl for 404s, and a list of the publications in your city that write "best of" round-ups. That is an afternoon of technical work and an ongoing PR job. Neither needs a monthly retainer for markup.

We charge for doing that work properly and for measuring it with enough repetitions to mean something. We do not charge for the rest, because we have read the same studies you now have. Disable JavaScript on your homepage and reload it. What is left on the screen is everything ChatGPT can see.