Skip to main content

AI Search Overview

Required add-on: Ai Search

The AI Search add-on lets listeners search your episodes by meaning rather than exact words. Someone who types what do I do when I’m burned out? finds your episode called Managing Creative Fatigue, even though the two share no words. It also gives you a search panel for your own back catalog while you prep, and a related-episodes list that suggests episodes close in subject.

It runs on your own account with an AI provider: OpenAI, Google Gemini, Cohere or Anthropic. You bring the API key and pay your provider directly at their published prices. Benecaster adds nothing on top.

AI Search works on every Benecaster plan, Launch included. Buy it from its shop page, install it like any other add-on (see Installing Add-ons), and switch it on for the shows you want under Benecaster → Settings → Add-ons → the AI Search card → Active on.

Before You Start

Connect your provider first: Benecaster → Settings → AI Search, choose a provider, add its API key and click Test connection. One provider and key serve all your shows. Until a key is accepted, the dashboard’s Needs Attention list shows Connect your AI Search provider. The settings are on Setting Up AI Search, and getting a key from each provider is on Setting Up Your AI Search API Key.

Which provider you choose decides how AI Search finds episodes, and what it costs.

OpenAI, Google Gemini, Cohere Anthropic
How it searches Builds a search index of your episodes ahead of time, then compares each search against it Reads your show’s episode list on every search and ranks the episodes that answer it
Cost Indexing costs a little once per episode; each search after that costs very little Nothing to index, but each search costs more
Transcripts Searched, with Transcription Service Not sent, so never matched
Related episodes by meaning Yes No: the list shows your latest episodes

With Anthropic you also choose the model on Settings → AI Search: Claude Haiku 5.5 (the default, and the lowest cost per search), Claude Sonnet 5.5 or Claude Opus 5.5. The episode list Anthropic reads is the same on every search, so a search within five minutes of the last one on the same show costs much less. Setting Up AI Search compares the models’ costs.

What Gets Searched

AI Search reads each published episode of every show it is switched on for: its title, description and show notes, as the words a reader sees. Drafts, scheduled and private episodes, and shows with AI Search off, are never sent to your provider.

Before anything is sent, the text is cleaned. HTML, images and their alt text, embeds, shortcode markup and web addresses are left out (a link’s text is kept). So is text that adds nothing to a search:

  • a description that only repeats the episode’s title or the opening of its show notes;
  • any paragraph a show repeats word for word in more than half of its published episodes, and in at least three, such as a fixed footer or a standard sponsor read. A footer you reword from episode to episode is caught only in the wording most episodes share.

Long show notes are split into pieces at paragraph and sentence breaks, so they are searchable all the way through.

Transcripts. With the Transcription Service add-on active and OpenAI, Google Gemini or Cohere as your provider, each episode’s transcript is indexed too: the spoken words and speaker labels, cleaned the same way. A transcript set to Hidden, or deleted, is never sent, and leaves the search index as soon as it becomes so. A search matches what was said in an episode only for a listener allowed to see that episode’s transcript under its visibility. For anyone else the episode is placed by its title, description and notes alone, and an episode found only through its transcript is not listed. Results never show transcript text. If you switch Transcription off, transcripts already indexed stay searchable until it is back.

Keeping the Search Index Current

With OpenAI, Google Gemini or Cohere, indexing starts by itself once your provider has accepted the key. Every published episode is indexed in the background, 20 at a time, and Settings → AI Search shows how far it has got.

After that, AI Search follows your changes:

  • A published or edited episode is indexed again about 15 seconds after you save it, without slowing the save.
  • Only new or changed text is sent. AI Search remembers a fingerprint of every piece it has indexed, so an episode whose text has not changed costs nothing, and editing one paragraph re-sends only the pieces around it.
  • An episode taken off Published (back to draft, scheduled, private, or into the trash) or deleted leaves the search index at once.
  • Switching provider or model re-indexes everything in the background, with no prompt. Until an episode is done, it is missing from results, so results fill back in as the job runs.

When your provider is busy or can’t be reached, indexing waits and tries again: after 5 minutes, then 10, 20 and 40, and at most every hour after that. Nothing is lost while it waits. If your provider refuses one episode’s text, that episode is set aside so the rest carry on, and Settings → AI Search says how many could not be indexed. If your provider stops accepting your key, indexing stops and the dashboard’s Needs Attention notice comes back.

Put the [benecaster_ai_search] shortcode on any page or post. Settings → AI Search shows it in its Add search to your site box, ready to copy. With the free Blocks & Widgets Library, the search box is also a block and an Elementor widget, with the same choices and no shortcode to type.

The box is labeled Search episodes by topic, with the placeholder Try “episodes about remote work” and a Search button. It searches only when the visitor presses Search or Enter, not as they type, because each search is a call to your provider.

Which show it searches. On a show’s or an episode’s page, that show. Anywhere else on a single-show site, your only show. On a site with several shows, add show_id outside a show’s pages: [benecaster_ai_search show_id="12"]. The box draws nothing when no show can be worked out, or when AI Search is off for that show.

Results list up to 10 episodes, best match first: each episode’s title as a link, its date and a short excerpt of its description. A visitor sees only episodes they can open, unless the box shows the others locked (see below).

What a visitor may see instead:

  • No episodes matched your search. Try describing the topic in other words.
  • Too many searches. Please wait a moment and try again. Each visitor can search 10 times a minute.
  • Search isn’t available right now. Please try again later. when your provider doesn’t answer, or the show has reached its daily limit on paid searches. A visitor never sees an error.

What a Guest Said

With Guest Manager and Transcription Service also on for the show, the box answers what has [guest] said about [topic]? A search that names a guest shows a section headed In their own words above the episode list: up to five passages of that guest’s own words from the show’s transcripts, closest to the search first. Each has the line {guest name}, in {episode title}, at {m:ss}: the title links to the episode, and the time (h:mm:ss from an hour) is where the passage starts in the audio. The episode list below is unchanged.

  • Naming a guest. Their full name (any case, ‘s allowed), or one part of their name, three letters or more, that no other quoted guest shares: what has Okafor said about rest. At most three guests per search, matched against their name in Guest Manager now.
  • Only on the topic. A passage is quoted only when it is at least as close to the search as the fifth-best episode result. When none is, the section doesn’t appear: what has Okafor said about quantum physics shows no quotes rather than Okafor’s words on something else.
  • What it needs. OpenAI, Google Gemini or Cohere as your provider (with Anthropic nothing is indexed, so the section never appears), and a speaker in the episode’s transcript matched to the guest in the Speaker names box on its Transcript tab. Only turns of four words or more are quoted.
  • Who sees a passage. Only a visitor who may open its episode and see its transcript. See Result Visibility.

It costs nothing extra: it reuses the search’s own call to your provider.

Placements

Each box chooses its own layout with placement:

placement What it looks like
full-page (the default) A wide box. Results show title, date and excerpt
sidebar A narrow box with Search under the input. Results show title and date
modal A Search episodes button that opens the full-page box in a pop-up. Esc, the × (Close search) or a click outside closes it

A misspelled value gives the full-page box. Themes can restyle each placement by its class; see Style the AI Search box’s placements.

Episodes a Visitor Can’t Open

By default, a box leaves out episodes the visitor can’t open. With result_visibility="locked" it shows them in their place instead, as a card with a padlock and a link to subscribe. Each box has its own setting, so a public search bar can show locked episodes while a members’ area hides them. See Locked Episodes in AI Search Results.

The built-in [benecaster_search] searches by keyword as the visitor types, at no cost. Both shortcodes can sit on the same site, or the same page, with no conflict. To offer AI Search only to some tiers, wrap it in [benecaster_tier_gate] as Standard Search’s page shows.

If AI Search is switched off for every show, WordPress prints [benecaster_ai_search] as plain text. A tier gate around it avoids that.

Pinned Results

Pinned results put a chosen episode, post or page first whenever a listener’s search contains your words: your getting-started episode for getting started, or a sponsor’s page for its name. They are set per show, under Benecaster → Shows → [your show] → Settings → AI Search (the tab appears on shows with AI Search on).

To add one, under Add a pinned result:

  1. In When a search contains, type the words.
  2. Under Show first, type a title and click Find, then choose the result. It can be one of the show’s own episodes, or any published post or page.
  3. Set Priority: lower numbers come first when several pins match one search. The default is 0.
  4. Click Add pinned result.

Each pin has Active and Priority, which you can change in place, and Remove. To change a pin’s words or result, remove it and add another.

Matching ignores capital letters but not wording. A pin for getting started fires for which episode explains Getting Started? but not for how do I get started, so add a row for each wording you want caught.

Pinned results come first, are not repeated further down, and still show when your provider is down or the show has reached its daily limit. A pin never shows a listener an episode they can’t access: a box that hides those leaves it out, and a box that shows them locked shows it locked. A pin whose post or page is later unpublished shows a warning on the tab and stops showing. Pins apply to the search box only, not to your own search panel.

AI Search adds a choice to the Related Episodes list on your episode pages: Related by meaning (AI Search). Pick it under Shows → Settings → Templates → Episode Page Blocks → the gear on the Related Episodes row → Which episodes to list (see Episode Page Renderer), or use it anywhere with [benecaster_related_episodes query="by_semantic"]. Each episode page then suggests the published episodes of the same show closest in subject. Episodes a visitor can’t open show with a lock, as in every related list.

It costs nothing per page view. Lists are worked out in the background from the search index, without calling your provider, and kept for each episode. Once the index is built, AI Search works out every episode’s list ahead of time, so a page has its list on its first view. An episode published since then shows your latest episodes until its list is ready, about 10 seconds after its first view.

A list is worked out again when the show’s search index changes (an episode published, edited or unpublished; a transcript updated, hidden or deleted), when it reaches the age you set under Keep each list for, after a provider or model switch, and when you click Clear related lists. An out-of-date list keeps showing until its replacement is ready, so pages don’t flip to your latest episodes after every edit. An unpublished episode drops out of every list at once. A list holds at most 50 episodes.

Three choices on Settings → AI Search settle close calls: Prefer episodes close in date moves an episode published near the one being viewed up among episodes similar in subject, Prefer episodes from the same season does the same for episodes with the same Season number, and, with Guest Manager, Prefer episodes with the same guest does the same for episodes that share a guest. Neither ever puts a weaker subject match ahead of a clearly closer one. See Setting Up AI Search.

Related episodes by meaning need OpenAI, Google Gemini or Cohere. With Anthropic or no provider, with AI Search off for the show, or if anything goes wrong, the list shows your latest episodes instead. Listeners never see an error.

What Searching Costs

Every search your provider answers is billed to your account, at your provider’s own prices.

With Anthropic, a rough guide. For a show with 300 episodes, our best estimate as of 9 November 2026 is about $0.005 a search on Claude Haiku 5.5, about $0.09 on Claude Sonnet 5.5, and about $0.18 on Claude Opus 5.5. A new search on the same show that starts within five minutes of the last one costs far less, because Anthropic reuses the episode list it was just sent: about $0.0004 on Haiku, $0.008 on Sonnet and $0.009 on Opus. Each search restarts the five minutes, and publishing or editing an episode changes the list, so the next search pays the full price again. These are estimates, not quotes: a show with more or longer episodes costs more, and Anthropic sets and changes its prices. Check Anthropic’s pricing for the current figures.

Several things keep that bill small and predictable:

  • Repeated searches are free. The same words searched again soon after are answered from a saved copy: for a day with OpenAI, Google Gemini or Cohere, and for an hour with Anthropic.
  • 10 searches a minute per visitor. After that, the visitor is asked to wait.
  • AI Search for signed-in visitors only, if you choose. Signed-out visitors then get Benecaster’s free keyword search box instead, and their searches never reach your provider or count toward the daily limit. It is off by default; see Who Can Use the Search Box.
  • A daily limit on paid searches per show, 500 by default. Once a show reaches it, its search box shows Search isn’t available right now. Please try again later. to new searches until midnight UTC. Searches already made that day, and pinned results, still show. You see AI Search reached today’s limit on paid searches for {show}. Its search box shows no results until midnight UTC. in the dashboard’s Needs Attention list, with Change the daily limit →. It clears when the day turns over or you raise the limit. Set the limit on Settings → AI Search (see Daily Limit on Paid Searches).

The limit, the per-visitor cap and signed-in-only search are what stop a busy or abused search box from running up your bill. Your own searches in the admin panel are not limited.

Your Own Search Panel

AI Search also adds a search panel beside every Benecaster admin screen, for finding the episode where you covered a topic while you prep. Open it from the AI Search card on the dashboard or with Alt+Shift+S. See Searching Your Own Episodes.

Privacy

Your provider receives your episodes’ text, and the words listeners type into the search box. Nothing identifying a listener is sent. The details, by provider, are on Privacy and Data Handling.

The Tour

AI Search Tour is listed under Tours in the Benecaster admin while AI Search is on for at least one show. It starts on Settings → AI Search, shows the Add search to your site box, and ends on the dashboard’s AI Search card.

For Developers

Code examples are in the recipe library.

See Also

Need this built rather than just documented? See our services →

Last updated