5 things not to do when you start GEO

⏰ The short version

• Keyword stuffing — an SEO-era habit. AI doesn't count repetitions, and the writing gets worse.
• Mass-generating 30 posts with AI — they all look alike, so none of them wins the citation, and you take on policy risk.
• Knocking competitors — if AI quotes it, that's your brand saying it. Asymmetric downside.
• Unverified numbers — they're easy to cite, which is exactly what makes them dangerous.
• Opening every crawler — search crawlers and training crawlers are different decisions.

1. Keyword stuffing

It used to work. Put "moisturizer for oily skin" in ten times and you climbed.

AI doesn't count repetitions. It looks for a sentence that answers the question. Ten mentions with no answer gets you nothing; one mention with a clear answer gets cited.

And stuffing makes the prose worse. Harder for humans to read — and hard-to-read writing is also less likely to be quoted. You lose twice.

Instead: organize around questions, not keywords. If you can't say in one sentence which question a page answers, the structure needs work.

2. Mass-generating 30 posts with AI

"We have 30 customer questions, let's generate 30 articles." Tempting, and common.

Two problems.

They converge. Same model, similar prompts, similar output. And your competitors are doing the same thing. When near-identical articles are everywhere, AI has no reason to pick yours.

There's policy risk. Search engines treat "scaled content produced primarily to manipulate rankings" as a violation. Drop out of the index and your AI citations go with it.

Instead: use AI for drafting, then add what only you have. Real customer questions, your own measurements, the edge cases you hit in practice. That's the part that differentiates — and the reason to cite you.

3. Knocking competitors

"Brand A's product has these drawbacks." Comparison content gets cited well, so the temptation is real.

But the downside is asymmetric. If AI quotes that line, it circulates as "[your brand] says this." Thin evidence turns into a defamation or false-advertising problem, and it looks bad regardless.

Instead: compare at the category level. Not "Brand A vs us" but "chemical vs mineral sunscreen," "kibble vs fresh." That's how customers ask anyway, and it gets cited without naming anyone.

If you must name a competitor, stick to verifiable facts with sources attached.

4. Unverified numbers

"98% satisfaction." "3x faster absorption." AI loves these — specific, easy to drop into an answer.

Which is precisely the risk. An unsupported number that gets cited propagates across many answers. There's no recall process.

Instead: ship the conditions and source with the number.

• ❌ "3x faster absorption"
• ⭕ "In our in-house test (n=20, Nov 2025), absorption time was one-third of the control"

Qualified numbers read as more credible, and AI prefers quoting them. Not a cost — a gain.

5. Opening every crawler without thinking

"To get cited we should allow everything." Half right.

AI crawlers do different jobs:

Crawler	Purpose
OAI-SearchBot, PerplexityBot	Retrieval for answers — allowing these helps citation
GPTBot, ClaudeBot	Training data collection — little direct effect on citation
Google-Extended	Google AI training — separate from search indexing

If citation is the goal, allow the retrieval crawlers. Whether to allow training crawlers is a judgment call about your content as an asset — there isn't one right answer.

What is clearly wrong is never looking. In our diagnostics a meaningful number of sites failed to crawl at all. Of 370 Korean brands attempted, only 284 completed.

Instead: open your robots.txt once and confirm the current state is intentional rather than inherited.

The common thread

All five come down to the same thing.

Trying to game AI mostly backfires. Repetition, mass production, and exaggeration are search-era tactics. AI is checking something else: is this sentence usable as the answer to this question?

Which makes the safest strategy also the most effective one: answer real questions honestly. Unglamorous, but it's what works.

Honest footnotes

• Search engine and AI policies change often. Current as of July 2026.
• Verify crawler names and behavior against each provider's official documentation — they change.
• #5 depends on how you view content licensing. The point is to decide deliberately, not to prescribe an answer.