Pimp My IDE / garage dispatch
Back to garage
October 5, 2026 | agents / retrieval / source custody

A provider switch is not a search policy.

Cloudflare's new Web Search API turns three search services into one request shape. That makes switching easy. It also makes the differences in retention, price, result text, logging, and crawler behavior easier to overlook.

Normalize the response. Keep the provider decision visible.

One parameter now moves the search engine.

Cloudflare released Web Search API in open beta on October 2. A request names Ceramic.ai, Exa, or Linkup. Cloudflare routes the query through AI Gateway and returns normalized items with URLs, titles, and descriptions.[1][2]

The clean interface is useful. It is not proof that the providers are interchangeable. The provider page lists different prices, retention support, search modes, and description behavior. Ceramic is the default. Exa uses its auto mode and returns highlights. Linkup uses its fast depth with raw search results.[3]

Shared JSON removes adapter work. It does not remove procurement work.

Zero retention and gateway logging are separate controls.

Cloudflare's provider table marks Ceramic and Linkup as supporting Zero Data Retention. It marks Exa as not supporting it. That statement concerns provider retention for requests made through this product. It does not mean the gateway keeps no record.

AI Gateway logging is enabled by default. Cloudflare says a log can include the prompt, response, provider, time, status, usage, cost, duration, and user agent. The gateway setting can disable collection, and a request can override that setting with a header.[4]

For search, the query may carry unreleased product names, incident details, customer identifiers, or a private error string. Decide whether the query belongs in a gateway log before the tool call leaves the agent.

A source link is an address, not a supported claim.

The API returns up to ten items. Each item can include a URL, title, and description. The response metadata can include the original query, a request ID, and latency. That is enough to discover pages. It is not enough to prove that a drafted sentence follows from the page.[2]

Fetch the selected page. Read the relevant passage. Save the exact URL beside the claim. If the page cannot be retrieved, mark the claim unsupported instead of treating a search description as evidence.

Crawler rules matter upstream.

Cloudflare requires the participating search providers to use crawlers that meet its verified-bot rules and to return a source link with each result. Its verified-bot policy requires deterministic identification, respect for robots.txt and crawl directives, reasonable request rates, and non-abusive behavior.[5]

That is a useful upstream commitment. It does not tell you whether a result is current, complete, independent, or suitable for the claim your agent wants to make. Provider conduct, result quality, and claim support need separate checks.

Interactive makeover / search customs manifold

Plug in a provider. Keep four inspections in series.

Traditional purpose replaced: a provider dropdown. Better version: a native provider selector drives a visible cartridge manifold while four independent checks build a review sheet. No search is sent.

Choose the provider route

Prices and retention labels below come from Cloudflare's provider page, read October 5. They are vendor documentation, not measurements from this page.

Search cartridge
Preflight inspections
0 of 4 inspections selectedCeramic.ai
Provider cartridge / normalized bus

Search customs manifold

Provider tokenceramic
Listed ZDRYes
List price / 1K$0.25

The provider is selected. No inspection is selected.

Next review: query class recorded.

All four selected inspections means the route-sheet structure is ready. It does not prove the provider honored retention, the gateway saved or omitted the expected fields, the result supports a claim, or a search ran.

One request / four decisions

Keep the easy switch. Expose the hard differences.

01 / QUERY

Classify before sending

Record what names, incidents, customer details, or unreleased work the search phrase can reveal.

02 / PROVIDER

Name the search service

Record the provider, search mode, contract, retention statement, and price source.

03 / LOG

Choose the gateway record

Decide whether to keep request metadata, payloads, both, or neither for this query class.

04 / REPLAY

Read the nominated page

Fetch the source, match the claim to a passage, and keep uncertainty visible.

Sources read

Source log and evidence boundary
  1. Cloudflare changelog, "Introducing Web Search API", published October 2 and read October 5, 2026. It announces the beta, names the three providers, states the Zero Data Retention and verified-crawler commitments, and describes AI Gateway billing and logging.
  2. Cloudflare Web Search API, "How to use", read October 5, 2026. It documents the REST and Workers routes, request limits, normalized response fields, request ID, latency field, provider default, and bring-your-own-key behavior.
  3. Cloudflare Web Search API, "Providers", read October 5, 2026. It lists provider modes, prices, Zero Data Retention status, defaults, and description behavior. These are Cloudflare's product statements, not our benchmark results.
  4. Cloudflare AI Gateway, "Logging", updated September 24 and read October 5, 2026. It says logging is enabled by default, names fields a log can contain, and documents gateway-level and per-request controls.
  5. Cloudflare Bots, "Verified bots", updated July 1 and read October 5, 2026. It defines deterministic identity, non-abusive behavior, crawl-rule compliance, and direct versus intermediary operation.
  6. Hacker News discussion, "Web Search API", opened October 5, 2026. We resolved the exact discussion ID through the HN search API. The discussion was discovery context, not evidence for product claims.

Evidence boundary: Cloudflare documents its beta, provider contracts, prices, retention labels, crawler requirements, and logging controls. Pimp My IDE designed the manifold and route sheet. We did not create an AI Gateway account, load credits, store a provider key, send a search request, inspect live logs, compare relevance, or verify provider-side deletion.