• AI visibility

Guide · AI search

Can optimizing for AI search backfire on your brand?

Yes, it can, and in controlled tests it often did. Popular rewriting tricks lowered pages in AI answers more often than many teams expect, and hype or manipulation was flagged or demoted when engines had even simple defenses. Careful work that adds real substance did not show the same downside.

By Underneath Team Guide Updated

The short version

  1. In a benchmark of AI answer engines, adding statistics to pages lowered their rank in 19 of 24 test settings (Puerto and colleagues (opens in a new tab)).
  2. In a shopping test, product descriptions rewritten with superlatives were flagged as questionable 100% of the time by every AI ranker that had a simple warning built in (Bagga and colleagues (opens in a new tab)).
  3. A research defense against manipulative rewriting cut its success rate from 50.32% to 6.20% (Li and colleagues (opens in a new tab)).
  4. Some defenses also catch honest brands: filters that caught at least 90% of planted fake brands also threw out roughly two thirds of genuine recommendations (Luo and colleagues (opens in a new tab)).
  5. Optimization does not have to make answers worse: a careful rewriting method raised visibility by 35.99% on average while keeping answer quality (Wu and colleagues (opens in a new tab)).

Can GEO tactics lower your visibility instead of raising it?

Yes. In the largest independent test of common rewriting tricks, many made pages less visible, not more.

The C-SEO Bench study, by Puerto and colleagues (opens in a new tab), rewrote pages using popular generative engine optimization (GEO) methods and measured where AI answers then cited them. Out of 54 combinations of method and topic, only 3 showed a reliable improvement. Adding statistics, a widely repeated tip, lowered a page’s rank in 19 of 24 test settings. For product recommendations on Anthropic’s Claude Haiku 3.5, 26 out of 30 cases moved pages down.

The effects were also unpredictable. For one rewriting method on retail products, pages moved up in 26.2% of cases and down in 12.8%. In 61.0% of cases, nothing changed. A 2026 survey of 45 studies by Martinez (opens in a new tab) describes a test that kept a simulated search step in place. There, rewriting only the page body cut how often pages reached the top 10 after re-sorting by 16%. Our related guide covers how GEO content tactics hold up in tests.

Do AI engines penalize content that looks manipulative?

Sometimes, and simple warnings were enough in tests. Hype and invented proof stood out to the stronger AI rankers.

Bagga and colleagues (opens in a new tab) built a shopping test with Amazon product listings and five AI rankers. They added one sentence to each ranker’s instructions, asking it to down-rank and flag misleading descriptions. Descriptions rewritten with superlatives were then flagged 100% of the time by every ranker. On GPT-5, those descriptions fell an average of 4.14 places. When the researchers let software rewrite copy to dodge the warning, it drifted to careful, fact-based prose. Their reading is that rank gains under the warning came from genuine content improvement, not manipulation.

Defenses built by researchers go further. Li and colleagues (opens in a new tab) tested a two-stage filter on five AI systems against seven manipulative rewriting attacks. It reduced the average attack success from 50.32% to 6.20%. These are lab defenses; no study we found confirms what ChatGPT or Google deploy today. We weigh the odds in whether AI search can filter manipulative GEO.

Live engines do not screen out every promotional format today. In our study of self-ranking “best of” lists, 24.2% of cited numbered lists with an identifiable publisher ranked that publisher first. Being cited now is no guarantee for later, since the lab defenses above target exactly this kind of self-interest.

Can honest content get caught by these defenses?

Yes. Defenses that block manipulation also remove some legitimate sources, so honest brands carry part of the cost.

The research defense above kept 94.12% of genuine evidence in use, which means a slice of honest material was dropped. Luo and colleagues (opens in a new tab) tested 12 AI systems on fake products planted in web pages. Two consensus filters, which keep only brands that other evidence backs up, caught the fake in at least 90% of cases. But they also discarded roughly two thirds of the real brands people would have been recommended.

The Martinez survey (opens in a new tab) warns that an overly strict anti-promotion filter “may penalize small publishers that legitimately describe their products.” If your content reads like an advertisement, it may sit closer to the line than you think.

Can aggressive optimization hurt how AI describes your brand?

Yes. Manipulative rewriting made AI answers worse in tests, and weakly sourced pages are common among optimized content.

Wu and colleagues (opens in a new tab) compared careful rewriting with “hijack” and “poisoning” attacks. The attacks raised visibility but always lowered answer quality and reliability. That is a poor trade for a brand whose name appears next to a misleading answer.

Chu and colleagues (opens in a new tab) scanned 10,095 pages returned by Google Search and Gemini for 1,000 real questions. On pages their detector judged to be optimized for AI, 69.34% of the sources those pages cited were rated low on how easily they could be checked. Separately, Khodayari (opens in a new tab) found hidden instructions to AI systems on live websites. AI systems obeyed them at most 8% of the time in tests; we cover the reputational risk in websites hiding instructions for AI search.

Does optimizing for AI search necessarily make answers less diverse?

No study shows that it does. The evidence points to gains canceling out, and to defenses as a separate risk to variety.

Visibility in an answer is shared, so one page’s gain is another’s loss. Puerto’s team found that the advantage of a rewriting method shrank as more sites adopted it, “eventually converging toward zero at full adoption.” That suggests a crowded race, not fewer sources.

There are two real pressures toward sameness. First, in Bagga’s shopping test, automated rewriting converged on one shared style; the authors list as an open question whether product descriptions would converge if everyone used it. Second, strict defenses can thin out legitimate choices, as Luo’s filters did. For now, Google’s AI Overviews draw on a wide range of sites. In a US study by Xu and colleagues (opens in a new tab), 56.2% of the sites they cited appeared only once in 40 days. Whether widespread optimization narrows that range has not been measured.

What should you do about it?

Treat GEO as a product-quality decision with brand risk, not as a bag of tricks.

  1. Ask any team or vendor for evidence that a tactic raised citations in a test with a comparison group, not a single before-and-after.
  2. Drop hype. Superlatives, invented awards and emotional pressure were the patterns AI rankers flagged most readily.
  3. Add substance that a reader can check: real figures with sources, clear product facts and plain answers. Whether polishing the prose alone helps is covered in readable writing and AI visibility.
  4. Never hide text or instructions for AI systems, and audit your site for any that agencies or plugins added.
  5. Keep investing in being findable in ordinary search. Puerto’s team found that a page’s position in the results handed to the AI mattered more than any rewrite.
  6. Measure before and after any rewrite, on the same questions, across more than one AI assistant.

For how to tell legitimate work from manipulation, see what separates legitimate GEO from manipulation. If you want help building that kind of program, see our generative engine optimization service.

What does the research not tell us yet?

Most of this evidence comes from lab tests, not from watching real brands over months.

  • No study we found tracked a real brand that was demoted by ChatGPT, Gemini or Google for optimized content.
  • The defenses tested are research prototypes; what engines deploy is not public.
  • Diversity under widespread optimization has been modeled and discussed, not measured on live engines.
  • The shopping and benchmark tests used fixed sets of pages, which can differ from the open web.
  • Results varied by AI model, so a finding on one engine may not hold on another.

Frequently asked questions

Can GEO hurt my Google rankings?

The research has not measured that directly. The Martinez survey reports a simulated search test where rewriting page bodies for AI made pages less likely to reach the top results.

Will ChatGPT penalize my content for being optimized?

There is no public evidence that it does today. In lab tests, a single warning in an AI ranker’s instructions was enough to flag superlative-heavy product copy 100% of the time.

Is adding statistics to content risky?

Adding invented or decorative statistics can be. In C-SEO Bench, adding statistics lowered rank in 19 of 24 settings, so add numbers only when they inform the reader.

If everyone optimizes for AI, does it stop working?

Gains from a shared trick shrink as more sites adopt it. Puerto’s team found the advantage of popular rewriting methods fell toward zero at full adoption.

Sources

Free strategy call

Some questions are easier to answer about your own business.

Bring the one that matters most. On a free 30-minute call we’ll take a first look at it and send you a short written read afterward.

Ask about my businessHow we work