Three Vendors, Three Kinds of GEO

Say you run a small or mid-sized company. You’ve noticed customers showing up already having asked ChatGPT about things, so one evening you open your laptop to work out who you’d even hire for GEO. One evening of looking, and you’ll meet three vendors — in three different ways.

The first one you find by searching. A long-established communications group has formally published its GEO solution on its own website, the methodology laid out in six steps: first define the key questions your stakeholders will ask, then calibrate the consistency of your brand’s own statements, reinforce third-party external signals, circle back to verify the technical foundation, and finally set up governance and ownership. You nod along the whole way down — until step two. That step is about working out where your brand currently stands inside AI answers, and the word they reach for to name that measurement is GPS.

The second one you already know. The forum platform you scroll every day has, at some point you can’t quite place, started selling a “GEO package” too. Click in and look at what’s inside: they’ll line up a few influencers plus a batch of throwaway accounts, post five glowing reviews on-platform, seed twenty positive comments underneath, and “get AI to pick up your reputation.”

The third one you don’t have to find — it finds you. A marketing firm’s ads are everywhere: in the elevator, before the video plays, all over your feed. The founder has written a book, launched a course, and runs seminar after seminar teaching people how to get found in the AI era; of the three, this one looks the most authoritative. The pitch is all about AI visibility — getting seen by more engines, landing inside the answer. But click through to what it actually teaches and sells: using AI to auto-generate a hundred articles, auto-fill Schema, auto-build internal links. And the name it hangs on the whole thing isn’t GEO — it’s “AI SEO.”

All three hang the GEO label. Not one word of what they sell overlaps. That fact alone is the answer: whatever this market is selling right now has stopped having much to do with the name on the label.

And the most dangerous of the three is usually not the most ridiculous one. It’s the one that looks the most credible.

Let’s Start With Something That Cuts Against Us

There is a real emperor’s-new-clothes problem in this industry, and it isn’t small.

An overseas search professional with twenty-five years in the field recently published a takedown of the entire GEO services business. His point wasn’t that AI search doesn’t matter — it was that the visibility report you’re paying for can’t be verified by anyone. AI platforms don’t expose an official data interface; there’s no equivalent of Search Console. Ask the same question five times and get five different answers. Point two different tracking tools at the same brand and the number of prompts they catch can differ by several multiples. In an environment like that, when a consultancy emails you “your AI visibility is up 23% this month,” where does that number come from, can it be rerun, does it still hold on a different machine — most of the time, the client can’t ask, and the vendor can’t answer.

He’s half right, and it’s the important half.

The black-box monthly report isn’t a problem with GEO as a discipline — it’s a problem with this particular business model. AI engines really do select which sources to cite; your site really can go unread and never get cited. Both of those are true. But “there is a real problem” doesn’t mean everyone selling something under that problem’s name is solving it. The old line about gold rushes — the real money is in selling shovels — is already playing out in the GEO market. The only wrinkle is that a lot of what’s being sold as a shovel isn’t one.

One more thing worth noticing: that takedown was written overseas. You will not find an article with that kind of force written in Taiwan — not because Taiwan doesn’t have these problems; read on and you’ll find all four fakes alive and well here. It’s because here, the people who see this industry clearly enough to write it are mostly the ones making a living off these very problems.

What follows are the four things most commonly packed into the GEO box and sold right now.

Fake GEO #1: Reading the G as GPS

To be clear up front: GEO stands for Generative Engine Optimization. The G is Generative, as in generative AI.

But in the marketing world, the geo- root has been welded to geography for far too long. Geo-targeting, geo-fencing, LBS, GPS — marketers have been saying these words for over a decade. So when a group that has spent decades in brand and media work starts talking about GEO, the gravity of that root pulls them straight back toward familiar ground.

Not long ago, a long-established communications group formally announced its GEO solution. The methodology runs six steps: define the key questions stakeholders will ask, calibrate the consistency of the brand’s own statements, reinforce third-party external signals, verify the technical foundation, set up governance and ownership. It reads smoothly.

And — this is the part of the whole piece most worth stopping on — a lot of what it lists is correct. When it gets to how you should read a brand’s current situation inside AI, it names things that genuinely do matter: how stable the mentions are, which sources are holding those mentions up, which kinds of questions the brand performs steadily on and which ones it drifts on. This is insider vocabulary. Someone who does this work every day would name those same things. It even states outright that the measurement has to be repeated across many rounds — and multi-round repetition is the single most important step in this entire discipline, and the one most often skipped.

And then, in the same breath, the word they reach for to name the act of measuring is GPS.

GPS is a satellite positioning system orbiting the Earth. It has nothing whatsoever to do with whether a language model will cite your website in an answer.

You could argue it’s just a metaphor — like GPS, we’ll locate your brand’s coordinates in the AI world. But the metaphor is exactly what gives the game away, because the defining property of GPS is the one property AI visibility measurement can never have: GPS gives you a precise, unique, repeatable coordinate. Stand in the same spot and measure a hundred times, and it returns the same set of numbers.

AI answers don’t work that way. Ask the same question ten times and it might mention you in six and skip you in four; change the model, change the day, change the account, and the distribution shifts again. Anyone who genuinely does AI visibility measurement every day doesn’t reach for satellite positioning as their mental image — they reach for opinion polling. You have to ask many times, log every round, and then talk about distributions and confidence intervals, not coordinates.

And that is exactly where the sentence detonates itself: “multi-round measurement” and “GPS positioning” cannot logically coexist. The reason you need many rounds is precisely that there is no such thing as a coordinate here. If there really were a GPS-style coordinate — precise, unique, repeatable — you would measure once and be done. Why would you need multiple rounds at all? Whoever put those two phrases in the same sentence is simultaneously asserting that this thing has an exact answer and that it doesn’t. He doesn’t know what he’s writing.

And this is the embarrassing part, far more embarrassing than plain ignorance would be. Someone ignorant gets all of it wrong. That isn’t what happened here. Here, every noun is right, and only the verb is wrong. They know what to look at — stability, sources, which question types hold and which drift — and they don’t know what looking at it actually is.

Nouns can be collected. Verbs can’t. You can assemble a full page of correct GEO terminology in half an hour, arrange it neatly, and slice it into six steps. But how you measure is something that only grows in the hands of a person who actually sat down, fixed a set of prompts, asked them for ten straight days, and pasted every answer into a spreadsheet. Someone who has done that will watch it mention them six times today and three times tomorrow — and will never again reach for satellite navigation to describe it.

So the correct nouns and the wrong verb came from different places. The nouns were collected. And whoever collected them has no hands.

Verdict: this isn’t a competence-boundary problem — it’s a problem with how the thing was produced. A methodology nobody has actually run, one that merely arranges the right nouns in a tidy row, can be produced very fast — fast enough that an entire “solution” appears in a single afternoon, all six steps present and accounted for.

Long-established communications groups do have real expertise, and it’s called media relationships and credibility building. But the other half of GEO is engineering: whether AI crawlers can reach you, whether a model can parse your page structure, whether your entity information is consistent across pages, whether your content contains a passage that can be lifted straight into an answer. That half does not complete itself just because you put it on slide five under the heading “Technical Foundation Verification.”

One more thing worth noting: “I can give you a precise coordinate” is exactly what every black-box monthly report in this market is selling. We’ll come back to that one shortly.

There’s a thirty-second test, and I mean this seriously: ask them directly, “what does the G in GEO stand for?” If they have to think for three seconds, don’t bother turning to the next slide.

Fake GEO #2: Sponsored Content Wearing a New Label

The second version is darker.

A number of content platforms and forum-style communities have had a rough few years — organic traffic eaten by search engines and AI summaries (though the more fundamental reason is simply that they couldn’t keep up with the newer platforms), ad revenue drying up, on-platform commerce never taking off. What do they still have? A stable of writer accounts, a sponsored-content production line, and a brand reputation as “the place everyone checks for reviews.”

How broke are they? Broke enough that when they entered the GEO business, the page design was traced straight from our own site. The thing they want to sell you is “make AI recognize who you are” — and their own storefront is traced from someone else’s.

Out of that, the “GEO package” was born. The pitch is polished: AI now references community discussion, so you need a presence on our platform; we’ll arrange a few genuine-sounding user reviews, seed some natural-looking discussion threads underneath, and once AI picks it up, it’ll start mentioning you in recommendations.

Strip it down and it’s the same writers, the same sponsored-post pipeline, the same seeded comments — only the invoice line item changed from “word-of-mouth marketing” to “GEO optimization.”

This stopped working reliably back in the SEO era, and it’s even less useful now, not for moral reasons but for mechanical ones: when AI engines pick citation sources, they weight the credibility of the source itself, and user-generated-content platforms already carry a lower baseline weight. Worse, when ten accounts praise the same brand in similar phrasing within three days, that isn’t buzz — that’s an anomaly signal. What you’re paying for may be turning your own brand entity into a contradictory dataset in the AI’s eyes: it reads a batch of enthusiastic recommendations that don’t line up with your own website, your independent third-party data, or your actual scale. When AI hits that kind of contradiction, its usual response is to skip you and cite a competitor whose data is internally consistent instead.

This isn’t speculation — there’s published research you can check. A September 2025 study from the University of Toronto, Generative Engine Optimization: How to Dominate AI Search (arXiv:2509.08919), measured and compared the source composition of AI search results against Google’s. In Google’s results, social content still holds ten to twenty percent; in AI search, the measured share of social sources across multiple industry categories is 0% to 0.3% — the paper’s own phrase for it is “near-total absence.” The thing you’re being sold — “getting AI to pick up your reputation on our platform” — is precisely the slot AI barely looks at.

Why fake reviews and seeded comments backfire on a brand is covered in more depth in the full anatomy of fake testimonials and social-media astroturfing; we won’t repeat it here.

Verdict: This kind of package has one identifying feature — it doesn’t touch your own asset. The money goes in, and the output stays on someone else’s platform, sitting in a handful of accounts that can be deleted at any time. The day the contract ends, you walk away with nothing.

Fake GEO #3: Using AI to Do SEO Is the Opposite of Optimizing for AI

The third version is the one most likely to get your guard down, because it’s the one packaged to look the most like an expert: there’s a book, there’s a course, there’s a stage.

Its actual product is an “AI SEO” automation suite: connect your site, AI analyzes your keyword gaps, auto-generates a hundred blog posts, auto-fills Schema markup, auto-builds internal links, auto-submits for indexing — “no humans, fully scalable.” None of this is new. The only new thing is the noun. When GEO became the hot word, the product didn’t change by a single line — but the ads and the seminars changed what they call it: AI SEO is GEO, and once you’ve run this package, your GEO is done.

There’s a directional error buried in here worth naming precisely.

“AI SEO” treats AI as a production tool, and the goal is still traditional ranking: more pages, more keywords, faster output. “Optimizing for AI” treats AI as the reader, and the goal is getting a language model to reach your page, understand what it says, work out who you are, and be willing to cite you in an answer. The first opens the tap wider. The second checks who’s actually on the other end of the pipe.

Auto-generating a hundred articles does almost nothing for the second goal, and it can actively hurt it. When an AI engine decides whether a domain is worth citing, it’s reading overall health: how consistent the content is, how checkable the facts are, how uniform your entity information is across the whole site. A hundred template-spun articles that repeat each other and contain no first-hand information will drag every one of those signals down. You don’t become easier to cite — you become a domain with a lot of content and not one sentence worth quoting.

Auto-injected Schema has the same problem. Schema is a structured statement of fact aimed at machines, and it only works if those facts are correct and consistent across pages. Tool-generated Schema routinely shows a product price that doesn’t match the checkout page, a company name spelled three different ways on three different pages, a rating with no source behind it. That kind of markup doesn’t earn points — it tells the AI this domain’s data can’t be trusted.

And here’s something that doesn’t occur to you while you’re sitting in the audience: a person who genuinely knows how to get you cited by AI doesn’t need to buy ad slots, publish a book, and run seminar after seminar to get onto your radar — they get cited naturally, because the substance holds up. The higher the authority stack and the wider the ad spend, the more it looks like scaffolding propping up a core that can’t stand on its own. They teach you to bury AI under mountains of content, and their own visibility is nothing more than that same trick of content marketing — just aimed at you instead of at the AI.

Verdict: Automation has a legitimate place in GEO, but that place is execution, not strategy — checking, not producing; the actual division of labor between tools and engines is mapped out in the honest map of AI visibility tools. But with this kind of vendor, the real problem was never the tool. It’s honesty: to win your business, they took two things that point in opposite directions and sold them as one noun. A vendor willing to lie to you at the level of the noun, before you’ve paid a cent — how honest do you expect their monthly report to be after you have?

Fake GEO #4: The Report You Can’t Verify

The fourth version is the most expensive, because it doesn’t even bother wearing a costume — what it sells is something you cannot check.

The monthly report arrives: “AI visibility up 23% this month,” “brand-mention rate up 1.8x,” “cited 7 times across 12 target prompts.” The numbers look great, the charts look professional. You want to double-check, so you open ChatGPT and ask one of the listed prompts yourself — you’re not in the answer.

You go back to the vendor. There are three standard responses: “AI answers are random,” “we tested a different model version,” “this month’s figure is a weighted average.” None of the three is technically a lie. None of the three can be falsified, either. A number that can’t be falsified is worth zero in a business context.

Here’s where it’s worth being precise: the randomness in AI answers is real, but that randomness is not an excuse for a black box — it’s exactly why measurement has to be rigorous. Asking a question once and drawing a conclusion was always going to misfire; the correct approach is fixed prompts, repeated across multiple rounds, every raw answer logged, citation frequency and position tallied, and the entire raw record handed to the client. A number like “cited in 3 out of 10 runs,” produced that way, is something you can rerun yourself, verify yourself, and use to measure next month’s progress yourself. We laid out this methodology in full in citation drift and multi-round measurement, and the four layers of metrics a verifiable results report should contain are in how to actually measure GEO ROI.

Verdict: The value of a GEO report isn’t how good the numbers look — it’s whether you can rerun it yourself. A monthly report that won’t hand over the raw prompt-and-answer log is the same thing as a traffic report that won’t hand over the GA login.

Sometimes All Four Fakes Are on the Same Price List

I’ve laid out the four kinds separately, as though each came from a different kind of company. It doesn’t work out that way.

There’s an outfit selling GEO in Taiwan right now whose public services page carries all of the following at once: a guarantee that your keywords will be recommended by the major generative AI engines; a claim that its methods can establish a “monopoly” over what AI recommends; a service menu where AI optimization sits alongside word-of-mouth marketing, a writer corps, managed five-star review campaigns, bloggers and influencers, and press-release media placement, pick whatever you like; and, finally, a reseller program whose commission is half of what you pay.

Those four things aren’t four separate flaws. They’re a machine whose gears mesh.

The guarantee is the fuel. Nobody can guarantee AI will cite you — the model companies themselves won’t even guarantee that asking the same question twice returns the same answer. So why would a promise that is structurally impossible to honor be printed on a company’s website? Because it wasn’t written for delivery. It was written for closing.

The reseller commission explains why the guarantee is mandatory. You pay a hundred; half of it goes straight into the pocket of the person who sold it to you, and what remains still has to cover the vendor’s own margin and overhead. How much is left to fund the actual engineering of making AI able to read your website? Under that cost structure, the package has to sell fast, sell in volume, and sell without needing to be explained — and the most effective way to make something sell fast is to attach a guarantee that sounds airtight. The guarantee isn’t confidence. It’s a product of the cost structure.

And the writer corps and the five-star reviews sitting on the same menu is not a coincidence. It leaks the seller’s mental model: they believe that getting AI to speak well of you and getting people to think other people speak well of you are the same job, which is why the same writers can take orders from both sides. We’ve already covered what happens next — AI reads a pile of enthusiastic recommendations that contradict each other and don’t line up with your entity data, and the cheapest way for it to resolve that is to skip you. The thing you paid for is making you harder to cite.

Verdict: when the word “guaranteed” shows up in a GEO proposal, don’t start by looking at what’s being guaranteed. Start by asking how the money is split. Guarantees and fat commissions almost always travel together, because they’re two ends of the same business — and that business is selling the thing, not making it work.

Three Questions That Knock Down All Four Fakes

You don’t need to understand GEO to protect yourself. You need three questions for the proposal meeting.

Question one: what does the G in GEO stand for?

It sounds like a cheap shot. It’s the single most efficient question you can ask. Anyone who can answer “Generative” at least knows what they’re selling. Anyone who answers “geography,” or “positioning,” or waves it off with “you know, it’s about making you visible on AI” — you already know exactly how deep their technical bench goes. This question eliminates fake #1.

Question two: are you changing my assets, or someone else’s platform?

Anyone actually doing GEO work is touching your domain, your content, your structured data, your entity-information consistency — things that stay with you after the contract ends. Anyone working on someone else’s platform (sponsored posts, planted comments, farmed accounts) is producing something that was never yours, and it may actively contaminate your brand data. This question eliminates fake #2.

Question three: how do I re-run this number? Show me the raw prompt-and-answer logs.

Ask them what prompts they tested, on what date, on which engine, with the full text or screenshot of every round’s answer. The ones who can do it will hand you a shared folder on the spot. The ones who can’t will start explaining to you that AI is random. This question eliminates fake #4, and it takes out most of the automation tools along with it — their dashboards usually hand you a score, not a process.

And note that a single screenshot is not a record. When a vendor hands you a “look, AI recommends us” image, it proves exactly one thing: that on that one draw, they showed up. Screenshots can be cherry-picked — the best-looking of twenty runs, and you’ll never see what the other nineteen looked like. If you’re going to look, you look at all of them; that’s the whole meaning of the word “re-run.”

Three questions, four fakes down. That’s a far more efficient use of your time than trying to learn twelve dimensions of technical detail, and it doesn’t require any technical background at all.

One More Test — And You Don’t Even Have to Ask

The three questions above require you to open your mouth. This last one doesn’t — you just go look at the vendor’s own website.

There’s a platform in Taiwan doing exactly this right now. Before you judge how well it’s built, look at what it sells: an online storefront system, POS checkout, payments, e-invoicing, international logistics, convenience-store logistics, product sourcing and listing, marketplace inventory sync, one-click cross-border listing, member VIP tiers, a community ecosystem, ad performance tracking, business analytics dashboards, AI agent integration — and then, after all that, an AI version of SEO and GEO, plus “use AI to predict and auto-generate content.” The largest line of type on the homepage introduces the whole thing as a place that will get you recommended more often by Google and by AI.

Start by counting how long that list is. More than a dozen product lines, spanning storefronts, payments, logistics, community, advertising, and AI. No team on Earth can be excellent at all of that simultaneously — this isn’t an unfair demand, it’s arithmetic. A service list that long isn’t an inventory of capabilities. It’s a net: cast it wide enough and it will catch someone.

And two items on that list tell you how the net gets hauled in. One shows you an org chart of the members you referred in. The other is cross-channel referral commission splitting. In plain terms: its growth engine isn’t doing the work well, it’s getting you to bring in the next person. This and the outfit from the previous section paying out half the fee are two models of the same machine.

Then we spent thirty seconds looking at the homepage source. Every item below is something you can re-verify yourself:

I’m not here to mock anyone’s aesthetics — whether a website looks good is a matter of taste, and I don’t care. But whether AI can read you isn’t a matter of taste. It happens to be the exact thing they claim to be selling.

Verdict: this doesn’t even rise to bait-and-switch — with bait-and-switch, there’s at least something in the pot. This is printing the entire menu first, hanging every dish on the wall, and learning to cook whatever you order after you’ve ordered it. And the fact that their own doorstep has never been swept for AI has already graded that menu for you: a company that can’t make AI understand itself wants to take your money to make AI understand you.

The good news is that this is the cheapest test on the market. You don’t need to know what robots.txt is. You only need to ask one thing: “Have you ever run a health check on your own website? Show me the score.“

(Incidentally, every check in this section — robots.txt, structured data, heading hierarchy, text-to-image ratio, missing alt text — is something the free health check at geoweb.tw scans automatically. Paste the vendor’s URL in, and you’ll have the answer in thirty seconds.)

What Real GEO Actually Does

Said in the positive, briefly.

When an AI engine answers a question, it looks for sources it can cite. It needs to be able to reach your page (technical accessibility), understand what you’re saying (structure and semantics), confirm who you are (entity consistency), judge whether you’re trustworthy (authority and evidence), and, when it needs one, find a ready-to-quote passage (citability). All five of these happen on your own domain, and none of them reset to zero just because you didn’t spend on ads this month.

It isn’t fast. It doesn’t work like advertising, where money in today means volume tomorrow.

And there’s something that has to be said plainly, because almost nobody in this market is willing to say it: AI itself is a black box, and it’s the kind that’s shut all the way. Nobody knows why a model picked A over B on a given question, which corpora it was fed, or which internal weight rules were pulling against each other. This isn’t a case of outsiders being in the dark while insiders know — the researchers who built these models can’t explain it either. Interpretability remains an unsolved research problem to this day. So when someone tells you they’ve mastered the AI’s algorithmic rules and can build a monopoly over what it recommends, you don’t need to be polite about it: that person is either bluffing you, or doesn’t know that they’re bluffing you.

But “can’t see inside” is not the same as “can’t measure,” and the conflation of those two is precisely what lets the black-box monthly report exist. Go back to the earlier comparison: an opinion poll can’t read anyone’s mind either, and it still manages to ask a thousand people, log every answer, compute a distribution, and then run the same method again next month to see whether anything moved. GEO works the same way. What you operate on is the health of the input side — whether AI can reach you, whether it can parse you, whether who you are is stated clearly, whether there’s a passage sitting there ready to be lifted. What you observe is the distribution on the output side — how many times out of ten you get cited, how you get described, who you appear alongside. The black box in the middle is one you don’t need to open, and couldn’t open anyway.

So what can be verified was never “why AI cites you” — nobody can answer that one. What can be verified is whether AI cites you right now, whether that’s more or less than last month, and whether you get the same result when you re-run it yourself. That is the only standard this entire article asks you to hold to, and it’s the only thing an honest person can give you and a dishonest one cannot.

When something is slow, re-runnable, and stays in your own hands all at the same time, it’s usually real. The emperor’s new clothes were never the slow option — they were the one that promised results tomorrow and claimed to understand what the AI is thinking.

So, This Week

  1. Run that proposal through the three questions — and scan the firm’s own website while you’re at it. Pass all three and come back with a clean structural profile, and it’s worth a conversation; fail any one of them and send it back for an explanation.
  2. Spend twenty minutes backtesting your own brand. Open an incognito window, write four questions in the words your customers would actually type (not your own brand name — customers don’t ask that way), ask each one five times, and log all twenty results. When you’re done you’re holding a number that doesn’t depend on anyone and that nobody can fake — that’s your baseline, and without it, whatever anyone tells you afterward, all you can do is believe it. The full procedure and a copy-paste log table are here: Screenshots Lie, Backtests Don’t. Most people’s reaction after this step is silence — because they aren’t in the answers.
  3. A backtest tells you whether you’re being cited right now. It doesn’t tell you why you aren’t. That part is structural — whether AI can reach you, whether it can parse you, whether who you are is stated clearly, whether there’s a passage sitting there ready to be lifted. To see that layer, run a free site-wide analysis on your site at geoweb.tw; it takes thirty seconds. And if the score is ugly, don’t rush to sign with anyone: the gaps are usually cross-page, cross-system engineering, not something one package or one tool can finish for you. To find out how the gaps get closed, which systems it touches, and how long it takes — send the report to [email protected] and we’ll look it over and come back with a specific read, including the parts you can handle yourself without hiring us.

Further Reading