Callers of Google's Custom Search JSON API must pick a new web search provider before January 1, 2027

A host swap is safe only if the new endpoint reproduces the old one's edge cases. Callers sent key, cx and q and got JSON with an items array back ^[3]. According to the developer who built cse-compat and described it on dev.to, code that has run against that contract for years quietly depends on details well beyond the items array ^[10].

When a query has no results, the items key is missing entirely, and plenty of code checks `if "items" in res` ^[11]. Matched words in htmlTitle and htmlSnippet arrive wrapped in `<b>` tags, and some interfaces render them directly ^[12]. Most pagination loops stop on queries.nextPage ^[13]. A request cannot reach past result 100 through start plus num, num must fall between 1 and 10, and Google returns a specific 400 error when a caller breaks either rule ^[14]. Even totalResults is a string ^[16]. "Getting these right is most of the work," the author wrote ^[17].

Errors are where a plain provider swap breaks first ^[15]. Retry logic often checks for a 429 carrying RESOURCE_EXHAUSTED ^[15]. A new provider that returns 402 Payment Required sends the client into a code path it has never seen ^[15]. The cse-compat Worker returns Google's error format along with Google's parameters and response fields ^[18].

"So if you were using it for general web search, there's no official drop-in path," the author wrote ^[8]. Vertex AI Search is a different API, so even teams that only ever searched their own sites face a rewrite ^[6]. In my view the shim is the right tradeoff when the search call is spread across several services or buried in a library the team does not own. That is the case the author calls painful for a rewrite against Brave, Serper, Exa or Tavily, because each returns its own JSON shape ^[9]. For one small script, the author calls a rewrite fine ^[9], and it leaves one fewer service to operate.

Running the shim puts a Cloudflare Worker on the request path. It is licensed Apache-2.0 and runs on Cloudflare's free plan ^[18]. Inside the Worker, the Serper key is a Wrangler secret; apps send a proxy key the operator makes up in place of the old Google key, so the calling services never hold the provider credential ^[20]. Results come from Serper today, with a Brave adapter included, even though the response shape is Google's ^[19]. For real use you deploy your own copy, since the public demo has a daily cap ^[25].

The client guidance is the most careful part of the post. Google's official client has Google's host built in but accepts an override: in Python through `client_options={"api_endpoint": "https://cse.yourdomain.workers.dev"}` ^[22], and in Node through `rootUrl` ^[24]. The Python example also sets `static_discovery=True`. That flag makes the client use its built-in API description and skip fetching one from Google, so nothing depends on Google's servers after the shutdown ^[23].

Budgeting starts from the old rate: 100 free queries a day, then $5 per 1,000 ^[4], or half a cent a query ^[2]. The post does not include prices for Serper, Brave or the other providers it names. The unit to price is the call. A loop that pages to the end of the 100-result window at 10 results per page makes 10 calls for one search ^[1].