What GEO Optimization Actually Costs You

By Muntazir · Published 14 September 2026

A company selling course-hosting software wrote its own “best learning management systems for selling courses” listicle and put its own product at the top. Google’s AI Overview picked up that page as a source anyway. Then it recommended four other companies — Kajabi, Thinkific, LearnWorlds and Teachable — every one of them named somewhere inside the very list the first company had written. Search Engine Land reported the pattern on 18 June 2026, drawn from an analysis of 100 B2B software queries run by SEO researcher Lily Ray: self-promotional “best of” pages got cited as a source 323 times, and in 224 of those cases — about 69 percent — the AI Overview named a competitor instead of the brand that had written the page. Optimizing the content got it cited. It didn’t get the company chosen.

That gap is the honest starting point for anything selling “GEO” — generative engine optimization, the practice of shaping a site so an AI answer engine can find, parse and quote it. The practice does something real. It also has a cost that rarely makes it into the pitch.

What does optimizing for AI answers actually buy a site?

The genuine upside is narrow and mechanical: it makes a page easier for a model to lift text from. A model composing an answer needs short, complete statements it can quote without losing meaning, not a paragraph that only makes sense after three sentences of preamble. One AI visibility audit engine scores exactly this — how many paragraphs on a page read as self-contained passages of roughly fifteen to eighty words, against how many run past a hundred and fifty words with no internal structure. A paragraph in the second group tends to get summarized by a model rather than quoted directly, and the summary loses both the site’s own wording and the attribution that came with it. A free page-level scorer checks the same thing on a single page without a full audit.

Clear facts, valid structured data and short paragraphs are the actual mechanics behind most GEO advice, and none of it is snake oil. It genuinely raises the odds of a page turning up as raw material behind an answer.

Does getting cited mean getting recommended?

No, and the LMS example above is the plain version of why. An AI Overview treats a page as a source of facts, not as an endorsement of the business that published it. It can read a company’s own comparison table, pull the competitor names sitting inside that table, and recommend one of them — while still linking back to the original page as where it got the information. Citation proves the content was useful. It proves nothing about which name the model reaches for once it actually answers.

What does chasing that citation cost the site itself?

Two costs show up directly in how sites respond to GEO advice, and both are things a technical audit can see — the full methodology covers what else gets checked alongside them.

The first is content added for a model and hidden from a person. A common move is an FAQ section built specifically to answer questions an AI model might ask, then collapsed into an accordion so it doesn’t clutter the page for a human visitor. A check built to find exactly this counts blocks of real prose sitting behind a click — inside a collapsed element, a hidden tab panel, anything marked to be invisible by default — and flags a page once more than a handful turn up. The content exists in the markup. Whether a crawler gives it any weight at all isn’t something the site controls, and a collapsed FAQ is one of the most common ways a page ends up burying the exact section it built to be found.

The second is sameness. Told to answer more and narrower questions, a site can end up with a run of pages that restate the same paragraph under a different heading — a “best X for freelancers” page next to “best X for small teams” next to “best X for agencies,” each one roughly the same text with the audience swapped out. A duplicate-content check that compares pages by overlapping five-word sequences catches this directly: page pairs sharing more than 80 percent of their main content get flagged as competing with each other rather than adding anything distinct. A model choosing what to cite has no reason to prefer one over the other, and often ends up citing neither.

What does it cost beyond the site?

There’s a harder-to-measure cost sitting one level up, and a German trade-press piece from meedia.de makes the sharper version of the argument: the more a brand shapes its own communication around what a machine wants to read, the greater the risk it becomes interchangeable. The piece’s warning is specific — that competition shifts away from earning a click on a company’s own site and toward earning a mention inside someone else’s answer, a surface the company doesn’t own, can’t format, and can’t fully see the workings of. A comparison page written to sound quotable is, by construction, written to sound like every other comparison page a model might quote instead — the same problem a business runs into when its identity isn’t distinct enough for a model to tell it apart from a competitor in the first place.

That’s the complication the LMS example already showed in miniature. The company’s page was written well enough to be cited. The citation didn’t stop the model from choosing Kajabi.

Is it still worth doing?

The groundwork — clean structure, real facts, pages that don’t reread as a template — is worth doing regardless, because a badly structured page loses to a well-structured one on citation with nothing else in play. What the chain above shows is that the payoff stops there. Getting quoted isn’t the same as getting chosen, and scaling up the tactics that earn a citation — more FAQ blocks, more comparison pages, more of the same paragraph with a new heading on top — tends to produce exactly the duplication and clutter a model has the least reason to prefer.