What can GEO not do?

GEO cannot make a model invent a recommendation that contradicts its training, cannot guarantee a mention next month, cannot replace a bad offer, and cannot ethically place a health or drug claim into an AI answer. It is a structure and clarity practice. If the underlying business, price, or claim does not hold up, better structure only makes the weakness easier to read.

Hard no's

• — Fake reviews or testimonials, including "example" quotes presented as real customers.
• — Impersonating a competitor, buying their brand name as a page title, or writing comparison pages that misstate their offer.
• — Medical or drug efficacy pages without a named, credentialed reviewer — and never a page designed to get a treatment recommended by a model.
• — "As seen in ChatGPT" or "AI recommended" badges. No engine issues those.
• — Networks of near-duplicate pages built to flood retrieval.
• — Markup that disagrees with the visible page.

Limits that are nobody's fault

1. Model memory. Broad category questions often resolve from training data, which no page can edit.
2. Volatility. Engines ship new versions and re-rank sources without notice; a mention is a snapshot, not a position.
3. Attribution gaps. An engine may use your facts and cite nobody, or cite a directory that copied you.
4. Query coverage. Google shows AI Overviews on some queries and not others, and that set changes.
5. Personalization. Two users asking the same question can get different answers.

Who this is for

• — Owners deciding whether to buy GEO work and wanting the downside stated plainly.
• — Marketers who have been pitched a guaranteed-placement package.

Who this is not for

• — Anyone looking for a workaround to the limits above.

What to expect instead of a guarantee

A defensible outcome looks like this: a scored baseline across a fixed prompt set, a set of pages that answer those questions accurately, and a rerun that shows what moved. Some questions will move. Some will not. Where a national brand is the entrenched default, expect no change and put the effort into narrower questions where you can be the best available source.

We put this in writing on every engagement: we deliver a visibility snapshot and published pages. We do not guarantee placement inside any model, Google ranking, leads, or revenue.

Sources

Claims about spam and structured-data policy should be checked against Google Search Central's spam policies. Claims about model behavior should be checked against each vendor's own documentation.

What a good engagement promises instead

A defensible contract says three things. First, what gets measured: 20–30 named customer questions across multiple AI/search systems, with the observations recorded consistently. Second, what gets delivered: a defined set of improvements, published with sources and appropriate structure. Third, what is explicitly not promised: placement, ranking, leads, or revenue. If a proposal skips the third item, ask for it in writing. Ask also how corrections are handled when an engine states something false about you, because the answer is process, not leverage: publish the correct fact, keep it crawlable, and file feedback with the engine.