New benchmark scores seven search APIs for AI agents

Artificial Analysis published its "Search Index" this month, rating seven search API providers across quality, cost, and speed for AI agent use cases. The Decoder reports that GPT-5.6 Luna, Parallel, Exa, and Firecrawl topped the rankings. The remaining three providers are not named in the summary, but the methodology is clear enough: this is the first systematic attempt to score the retrieval infrastructure that sits between an AI agent and the web.

That infrastructure matters more than most B2B communications teams have noticed.

The pipe nobody is watching

When a user asks an AI agent a question requiring live web information, the agent does not browse freely. It calls a search API, receives a set of results, and synthesises an answer from whatever that API returns. The search API is therefore a gatekeeper: if your content does not surface in a given provider's index at retrieval time, it does not exist for that agent's answer. Brand visibility in agentic AI is not merely a function of what models were trained on; it is a function of which search APIs the agent uses and how well those APIs index your domain.

This is the citation-pattern problem restated at the infrastructure level. Most attention to AI visibility has focused on training data: which sources did GPT-4 or Claude read, and how often do they cite them? The agentic shift moves the question downstream. Real-time retrieval APIs, not pre-training corpora, will increasingly determine whether a brand appears in an agent's response to a live commercial or policy query.

Artificial Analysis is applying to search APIs the same benchmarking logic it already uses for language models: measure on multiple axes, publish the numbers, let buyers decide. The quality dimension likely captures relevance and completeness of retrieved results; cost is straightforward (price per query); speed affects latency in agent pipelines where multiple API calls happen sequentially. A provider that wins on quality but loses badly on speed may be unsuitable for high-frequency agent use cases, regardless of how well it indexes a given domain.

Who wins and who loses

For brands in financial services, multilaterals, or major industrial groups, the practical consequence is uncomfortable. These organisations typically invest in owned content: white papers, research portals, microsite architectures with careful canonical structures. That investment was calibrated for Google crawlers and, more recently, for the training pipelines of large language models. Neither of those calibrations translates automatically to search API coverage.

The four top-ranked providers in the Artificial Analysis benchmark (GPT-5.6 Luna, Parallel, Exa, Firecrawl) each have distinct indexing philosophies. Exa, for instance, is built around semantic search rather than keyword matching; Firecrawl specialises in structured web scraping. A brand whose content is optimised for keyword retrieval may index well in one provider and poorly in another. The benchmark makes those differences legible for the first time.

For AI developers and platform teams building agents, the benchmark is a procurement tool. They will use it to select their retrieval stack. That decision, made largely out of view of marketing departments, will determine which sources their agents cite. A financial institution whose research portal indexes poorly in Exa may simply not appear in answers generated by agents that rely on Exa. The brand has no direct lever; the AI developer does.

The response is not panic but measurement. Communications and digital teams at institutions that care about AI visibility should treat search API coverage as an auditable variable, alongside training-data citation rates and direct LLM brand mentions. The Artificial Analysis benchmark provides a framework; the next step is testing how specific domains index across the top providers, and whether technical interventions (structured data, API-accessible content formats, faster page loads) shift the outcome.

The benchmark will presumably update as providers iterate. What it already establishes is that the retrieval layer in agentic AI is now competitive, differentiated, and consequential enough to rank. Brands that treat it as somebody else's infrastructure problem will find it is, in fact, their visibility problem.