AEO and GEO
Which GEO and AEO tactics backfire or waste your money?
Five shortcuts in our corpus of 320 publishers are mostly called useless or risky: mass-produced AI pages (12 publishers against, 0 for), hidden prompt-injection text, fake reviews or planted Reddit posts, expired domains and paid link schemes, and keyword stuffing. Each tries to game the index or the model instead of earning a real record. Some worked in lab tests on some models; the sources that watched them in the wild report them fading or getting removed.
Updated 19 November 2026
In our corpus of 320 publishers, Mass-produced AI pages and four other shortcuts are all labelled mostly called useless or risky. For Mass-produced AI pages alone, 12 publishers are against it and 0 for. The table below shows the same lopsided result for hidden prompt-injection text, fake reviews and planted Reddit posts, expired domains and paid link schemes, and keyword stuffing. University researchers did show hidden text can push a product up in some models' answers, but a Microsoft Research study found some models resisted every manipulation they tried, and OpenAI's and Anthropic's own rules for apps and tools forbid hidden instructions. In the paper that named GEO, keyword stuffing did worse than no edit at all.
What our source count shows
Each row below is one tactic, and each cell counts how many of the 320 publishers in our 435 saved sources took that stance. A publisher can take more than one stance, so the columns do not add up to a total.
| Step | Recommend it | Measured it working | Call it useless or warn | Measured no effect | Measured harm | Our label |
|---|---|---|---|---|---|---|
| Mass-produced AI pages | 0 | 0 | 12 | 0 | 0 | Mostly called useless or risky |
| Hidden text or prompt injection aimed at the model | 1 | 2 | 9 | 1 | 0 | Mostly called useless or risky |
| Fake reviews or planted Reddit posts | 2 | 1 | 7 | 1 | 0 | Mostly called useless or risky |
| Expired domains, parasite pages, paid link schemes | 0 | 1 | 8 | 0 | 0 | Mostly called useless or risky |
| Keyword stuffing | 0 | 0 | 0 | 1 | 2 | Mostly called useless or risky |
What sources that measured it reported
- Microsoft Research / Arizona State University, 27 October 2025: "Sonnet-4.5 exhibited the strongest resistance, showing virtually no susceptibility to any manipulation strategy tested." (source)
- UC Berkeley, 25 September 2024: "We then present a tree-of-attacks-based jailbreaking technique which reliably promotes low-ranked products." (source)
- Harvard University, 2 September 2024: "We observe that the strategic text sequence significantly enhances the visibility of both products by increasing their chances of appearing as the top recommendation." (source)
- Microsoft Research: "Some frontier models resisted everything, but others were completely compromised. All payments were redirected to the attackers." (source)
- Kalicube / Search Engine Land, 17 February 2026: "The fake personas had volume, they had mentions, but what they lacked was cascading confidence" (source)
- Ahrefs, 4 December 2025: "Many lack any human presence and look almost entirely like they were built for link building and / or citation purposes." (source)
- SE Ranking, 24 November 2025: "Broad titles and URLs that simply describe the main topic receive over 2x citations on average, compared to highly keyword-optimized ones." (source)
- Princeton, Georgia Tech, Allen Institute for AI, IIT Delhi (arXiv 2311.09735v3, KDD 2024): "While widely used for Search Engine Optimization, we find such methods offer little to no improvement on generative engine’s responses." (source)
What this means for your business
- Walk away from anyone offering hundreds of unedited AI-written city or comparison pages; 12 publishers warn against mass-produced pages.
- Never hide instructions to the model in your pages or product text; the platforms' own rules forbid it where they address it, and it can cost you your listing.
- Skip fake Reddit personas, bought reviews and keyword-stuffed titles; put that money into real service and a few well-researched pages.
We saved 435 guides, studies, forum threads and platform docs from 320 publishers in full, recorded 1592 word-for-word method statements across 99 tactics, and logged 216 reported results with numbers. Every count is by publisher, not by article, so one prolific blog only gets one vote per stance, and every quote is checked word for word against our saved copy. A high count is not proof that a tactic works: many GEO writers sell GEO services or software, and a measured claim usually means one site's before-and-after numbers rather than a controlled test.
Common questions
Does hidden prompt injection work to get recommended by ChatGPT or Claude?
Researchers have tricked some models in lab tests, but a Microsoft Research study found some models resisted every manipulation tried, and OpenAI's and Anthropic's rules for apps and tools forbid hidden or manipulative instructions. Hidden text puts your listing or integration at risk for a trick that may already have stopped working.