SEO for AI Agents: Can Search Bias Influence Tech Choices?

Overview

A growing number of consumers and developers are skipping traditional search engines in favor of AI agents for product and technology research. Recent data indicates that nearly half of AI users now begin their shopping journey on generative platforms, a significant shift from just two years ago. This migration raises critical questions about visibility and influence: if an agent decides the answer, who decides what the agent sees?

New research suggests that by controlling the search results an agent retrieves, vendors may be able to steer its recommendations. In a controlled experiment using Google Cloud infrastructure, a researcher found that when an AI agent relied on search-grounded information rather than its training data, it was far more likely to recommend specific commercial tools—particularly those hosted on the platform where the test was conducted.

The Shift from Search Engines to AI Agents

The behavior of both retail consumers and software developers is changing rapidly. A survey by L.E.K. Consulting, summarized by MarTech, reveals that 46% of AI users now start product research on standalone platforms such as ChatGPT, Gemini, Perplexity, or Claude. This is a sharp increase from 25% in 2024. Conversely, the percentage of people beginning their research on traditional search engines dropped from 43% to 24% over the same period.

This trend is not limited to consumer retail. Developers are increasingly relying on coding agents to make architectural decisions, from choosing databases to selecting hosting providers. A September 2026 report from Emplifi and Semrush found that 66% of consumers have used AI tools to research products, with 11% making it their initial stop.

However, reliance on agent advice does not guarantee a specific outcome. A study by Armature published in September 2026 analyzed 5,292 coding-agent sessions across Claude Code, Codex, and Cursor. The researchers found that when asked to add a voice agent feature, the agents often chose different tools: Claude Code selected Twilio, Codex opted for the OpenAI Realtime API, and Cursor chose Vapi. The study also noted that mere mention did not equate to selection; PayPal was referenced 139 times but chosen zero times.

Experiment Design: Controlling the Information Stream

To understand whether search bias could influence these technical recommendations, a researcher based at Google Cloud designed an experiment to test if they could push an agent toward Google Cloud products. The study utilized Claude Code with Claude Sonnet 5, running through the Gemini Enterprise Agent Platform.

The experimental setup involved 310 prompts covering eight categories, including AI/ML, infrastructure, databases, and security. None of the prompts named a specific vendor. To ensure complete control over the data the agent accessed, the researcher blocked the agent’s native WebSearch and WebFetch tools. Instead, they deployed a custom MCP server called gemini-search-mcp.

This server acted as a intermediary between the agent and Google Search. It logged every query generated by the agent and allowed the researcher to alter the query string before it reached the search engine. The tool returned written summaries, source lists, and the specific queries executed. This setup enabled a controlled environment where the agent’s knowledge source could be manipulated without altering the model’s underlying training data.

Results: Training Data vs. Grounded Search

The experiment consisted of a baseline and five distinct tests to measure the impact of search availability and encouragement on the agent’s recommendations.

In the baseline control, prompts were sent directly through the Anthropic API with native search enabled. Sonnet 5 searched only 3 out of 930 trials, resulting in a search rate of 0.3%. For high-level advisory architecture questions, the model predominantly answered from its training data, which has a cutoff of January 2026. This was significantly lower than the 30% search rate observed in the broader Armature study, likely because Armature’s agents were writing code in real repositories rather than answering conceptual questions.

When the custom search tool was introduced, the dynamics changed. In tests where the agent was encouraged to search for current documentation, capabilities, and pricing, the agent began to retrieve and rely on real-time web data. The critical finding was that when the search results were manipulated to favor Google Cloud products, the agent’s recommendations shifted accordingly.

Key Finding: When the agent was forced to rely on search-grounded information rather than its static training weights, it became significantly more susceptible to the content it retrieved. By shaping the search results, the researcher could steer the agent toward recommending specific Google Cloud services such as Document AI and Speech-to-Text.

Implications for Generative Engine Optimization

The results highlight the emerging field of Generative Engine Optimization (GEO). Just as traditional SEO aims to rank websites higher in search engine results to attract human readers, GEO seeks to ensure that AI agents cite and select specific tools or products when generating recommendations.

The experiment demonstrated that different agents have vastly different habits regarding web search. Codex searched in 94% of sessions in the Armature study, while Claude Code searched only 30% of the time. This variance means that a strategy effective for one agent may fail for another. Furthermore, the ease with which an agent’s recommendations can be influenced by controlling its search inputs suggests that the "truth" an agent presents is often a reflection of the information ecosystem it is configured to access.

As AI agents become the primary interface for technical and consumer research, the ability to influence the data sources behind those agents will become a critical competitive factor. Vendors may need to optimize not just for keywords, but for the specific retrieval behaviors of the coding and research agents their customers use.