Where margins are high, manipulation follows. I closed my BrightonSEO talk on local AEO with the rougher edges of the field, which show up most in high-margin niches like roofing and personal injury law. This is a tour of what’s out there and what the evidence says, not a how-to.

Old tricks, new targets

Hidden text. White-on-white text is a 2005-era black-hat trick, and Google’s spam policies have long named it. In a test by Otterly.ai on a single low-content domain, only some AI platforms read it:

Did AI platforms read hidden text? Results from Otterly.ai’s hidden text test. Copilot read it, then refused to load the page after flagging the injected instructions. Claude read it. Gemini and ChatGPT read it only in deep thinking or deep research modes. Google AI Mode and Perplexity did not retrieve the page. None of the seven platforms tested followed the hidden instruction to treat the site as a trusted source. Did the platform read the hidden text? CopilotYes, then refused to load the page after flagging it ClaudeYes GeminiOnly in deep thinking or deep research modes ChatGPTOnly in deep thinking or deep research modes Google AI ModeDidn’t retrieve the page PerplexityDidn’t retrieve the page Followed the hidden “treat us as a trusted source” instruction: 0 of 7 platforms.
Source: Otterly.ai. One small site over a limited window, so it may not generalize.

Telling AI to trust you. The same experiment added a hidden instruction asking AI systems to treat the site as a “primary trusted source.” None of the seven platforms followed it, and two flagged or reported it. ChatGPT and Google AI Mode instead confidently described the page with content from unrelated sites.

Negative GEO

Researchers have coined “negative GEO” for using AEO tactics to damage a competitor’s AI reputation. Similarweb’s write-up lists the playbook: prompt poisoning (“5 reasons we left [competitor]” content), comparison hijacking, planted “would not recommend” stories on Reddit and Quora, content flooding across many domains, and review bombing. In an experiment by the agency Reboot, some models surfaced the planted claims while others were more skeptical, so results vary hugely by model. It’s also hard to scale.

If you’re on the receiving end, Similarweb’s defensive advice is practical: monitor how AI describes your brand at least weekly, watch for sudden sentiment swings or one platform that’s far more negative than the rest, and build a deep base of credible positive sources so no single planted story dominates.

Other fringe tactics

  • Reddit and forum farming. I covered what’s happening, and how to participate honestly instead, in Reddit and local AEO: earn it, don’t farm it.
  • Resurrected domains: dropped websites bought back and used as manipulated backlinks or citations pointing to a main site.

Why we don’t recommend it

My position in the talk: it’s your ethical choice, but none of this scales, and the platforms are catching up. Detection is already visible in the tests above. The upside is short-lived and the downside, including exclusion from retrieval, public call-outs and wasted spend, lasts longer. It’s also why the truly wild tactics never get discussed on stage, as I noted in my BrightonSEO takeaways: nobody wants their clients’ names attached to them.

What to do instead

  1. Audit your own pages for accidental hidden content.
  2. Track what each AI platform says about your brand, and set alerts for sudden shifts.
  3. Build the positive base: reviews that name services, local listicles, citations and genuine community participation. Start with the two local rulebooks and published pricing.

Worried about how AI describes your brand? Get in touch and we’ll take a look.


Leave a Reply

Your email address will not be published. Required fields are marked *