The article examines whether vendors can use SEO or manipulated search results to influence an AI agent’s product recommendations. The author created 310 vendor-neutral architecture and technology prompts spanning eight Google Cloud competitor categories, ran each three times, and tested Claude Sonnet 5 under five search configurations, for 5,580 trials in total. With native search available but no instruction to use it, the agent searched in only 3 of 930 trials, or 0.3%, and selected a Google Cloud product in 9.8% of all baseline trials. Blocking search produced a similar 9.2% selection rate, while connecting a search MCP server without encouraging its use led to 14.2% of trials searching and a 10.0% selection rate. Explicitly telling the agent to check current documentation caused searches in 99.5% of trials and increased Google Cloud selections to 13.0%. The strongest test rewrote every search query to favor Google Cloud while still returning legitimate Google Search grounding; this raised selection to 17.1%, although competitors were still chosen in 706 of the 925 searched trials. The manipulation increased product mentions much more than primary recommendations: from 56.0% to 89.1% in the comparison between no-search and encouraged, biased search, while selections rose from 9.2% to 17.1%. The author presents this as a theoretical ceiling far beyond ordinary SEO. The main conclusion is that recommendations in this setting were shaped mostly by what the model learned during training, while prompt instructions and whoever controls the search tool determined whether current information could influence the answer. The author notes that results may differ for coding tasks or models that search more often, and advises developers to request searches when current documentation or pricing matters and to scrutinize third-party search tools that can silently rewrite queries.
