Integration
Web search API for AI agents vs AI visibility API: what each returns
Tavily, Exa and Brave return ranked pages for your model to read, from $4 to $8 per 1,000 requests. An AI visibility API returns the engine's own answer.

By Mathias C.SEO consultant, and more recently GEO, at MentionBox
A web search API for AI agents takes a query and hands your code ranked web pages with their text, so a model you run can read them and write the answer. An AI visibility API takes a prompt, sends it to the engine a buyer actually uses, and hands you back that engine's answer, the brands named inside it and the URLs it cited. Tavily, Exa and Brave sell the first job between $4 and $8 per 1,000 requests; the second runs from $1.20 to $15.75 per 1,000 answers across the APIs this site rates.
This page is for the developer who already has Tavily or Exa wired into an agent and has just been asked whether the company shows up in ChatGPT. Below: what each kind of API returns in the response body, a table of the two jobs side by side, the four questions a search API cannot answer about a brand, and the places where the two product categories genuinely touch.
What a web search API for agents returns
The contract is the same across the three. You send a query, you get back a ranked list of pages with enough text for a model to work from. Tavily's API reference describes the product as "a powerful search engine for LLM agents" and puts it at one endpoint:
curl -X POST https://api.tavily.com/search \
-H "Authorization: Bearer $TAVILY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"query": "best AI visibility API for an agency",
"search_depth": "basic",
"max_results": 10,
"include_raw_content": "markdown",
"country": "united states"
}'
Every parameter there has a documented ceiling worth knowing before you build on it. search_depth takes basic, advanced, fast or ultra-fast; max_results defaults to 10 and stops at 20; include_domains accepts up to 300 domains and exclude_domains up to 150; country only applies when topic is general. The response carries query, results, images, response_time and, when you ask for it, usage with the credits spent. Each result is a small, stable object:
{
"title": "Lionel Messi Facts | Britannica",
"url": "https://www.britannica.com/facts/Lionel-Messi",
"content": "A short description of the search result.",
"score": 0.81025416,
"raw_content": null
}
Exa's search endpoint follows the same shape with different controls: type defaults to auto and also accepts lite, deep-lite and deep, numResults defaults to 10 and goes to 100, and a contents object decides whether you get text, highlights or a model-written summary per page. Its response includes a costDollars object, so each call tells you what it cost. Exa describes the endpoint as a way to "search the web in natural language and get clean, relevant page content in one request", and says its highlights "give models and agents the evidence they need without filling the context window".
Brave's web search endpoint is a GET with an X-Subscription-Token header, a q capped at 600 characters and 75 words, count up to 20, a freshness filter taking pd, pw, pm, py or a date range, and an extra_snippets flag that returns up to five additional excerpts per result. Brave runs its own crawler rather than reselling another engine's results, and its API page states that "Our index is our own: crawled, ranked, and served end to end". For agent work it also publishes an LLM Context endpoint at /res/v1/llm/context, which returns pre-extracted snippets per URL under a token budget set by maximum_number_of_tokens, 8,192 by default and up to 32,768.
Perplexity sells the same raw-results product alongside its answer API, and its documentation draws the line more plainly than anyone else's: the Search API "returns ranked web results as JSON, not a generated answer", and "There is no generated answer and no token billing". Each result carries title, url, snippet, date and last_updated.
| Search API | Endpoint | What a result carries | Published price per 1,000 | Free allowance |
|---|---|---|---|---|
| Tavily | POST api.tavily.com/search | title, url, content, score, optional raw_content in markdown or text | $8 pay as you go at $0.008 a credit, $5 on the $500 Growth plan | 1,000 credits a month |
| Exa | POST api.exa.ai/search | title, url, publishedDate, author, text, highlights, summary, costDollars | $4 on lite, $7 on auto for ten results, plus $1 per 1,000 extra results | $10 of credit, reset monthly |
| Brave | GET api.search.brave.com/res/v1/web/search | title, url, description, age, page_age, profile, up to five extra_snippets | $5 on the Search plan, which also covers LLM Context | $5 of credit a month |
| Perplexity Search | POST api.perplexity.ai/search | title, url, snippet, date, last_updated | $5, or $1 with search_type: "fast" | not published |
Endpoints, fields, limits and rates from each vendor's own documentation and pricing pages, read on November 12, 2026. Tavily bills a basic, fast or ultra-fast search at one credit and an advanced search at two; Perplexity bills per successful request, which can carry up to five queries.
What an AI visibility API returns
The input is different before the output is. You send a prompt, an engine, a country and a language, and the API runs that prompt the way a person would run it. What comes back is the answer text, the URLs the engine cited, the brands it named, and in the computed tier of the market, the position your brand held in the answer, its share of voice against named competitors, and the sentiment of the sentence it appeared in. Our explainer on GEO metrics gives the formulas behind those three figures.
The collection method is the field to read first on any vendor's page, because it decides what the number means. In our October 23, 2026 survey, DataForSEO sells both routes side by side, an LLM Scraper that reads the interfaces and an LLM Responses endpoint that calls the official model APIs, at $1.20 per 1,000 answers over five engines. cloro runs browser sessions against the real interfaces of six engines, at $4 per 1,000 on its Lite plan. API for GEO runs each prompt natively on each engine in the country and language asked for, and returns share of voice, position, sentiment, categorised sources and the engine's own fan-out questions in the same response, at $13 per 1,000 answers; it lists ChatGPT, Perplexity, Google AI Mode, Claude, Copilot and Grok, and we confirmed Gemini, Google AI Overviews, DeepSeek and Meta AI in a hands-on test on October 26, 2026. MentionBox, which publishes this site, has a commercial link with API for GEO.
None of those three is a web search API. You cannot hand them a keyword and get a SERP, and the per-unit price is an order of magnitude above Brave's because the unit is an engine answer rather than a list of links.
The two jobs side by side
| Web search API for agents | AI visibility API | |
|---|---|---|
| What you send | a query, with domain, country and freshness filters | a prompt, an engine, a country and a language |
| Who writes the answer | a model you choose and prompt yourself | the engine your buyer uses |
| What comes back | ranked URLs, snippets, extracted page text | answer text, cited URLs, brand mentions, position, share of voice, sentiment |
| Unit billed | one request, 10 to 20 results included | one answer, per prompt, per engine, per locale |
| Published rate | $4 to $8 per 1,000 requests | $1.20 to $15.75 per 1,000 answers |
| Latency you design for | under a second to a few seconds, inside a user's turn | seconds to minutes, inside a scheduled job |
| Typical caller | an agent answering a question now | a weekly job that fills a report |
| What it cannot show | what any AI engine told a user | which page ranks third on Google today |
Left column from the Tavily, Exa, Brave and Perplexity documentation read on November 12, 2026; right column from the 55 APIs in our comparator, checked on October 23, 2026.
Four questions a search API cannot answer about a brand
- Does ChatGPT recommend us? The three search APIs never call ChatGPT. The closest official substitute is OpenAI's own web search tool, which measures the model and instructions you configured rather than the consumer product, as we covered in the OpenAI web search tool compared with ChatGPT answers.
- Which sources did the AI Overview cite? A search ranking is a poor proxy here. cloro's published comparison of 500 queries run on July 8, 2026 reports an AI Overview on 64% of them, with its cited domains matching Google's own organic top ten only 51% of the time, and exact-URL overlap with the live Google top ten of about 60% for Tavily, 47% for Brave and 32% for Exa.
- What is our share of voice? That figure needs a brand list, many answers per prompt and a denominator. A search API returns neither answers nor brand entities, so everything upstream of the arithmetic is missing.
- Can we keep the history? Brave's terms of use, last updated September 1, 2026, forbid a customer to "store, cache, or create a database of Search Results, in whole or in part, other than transient storage required for operation of Customer Applications", and a separate clause forbids using results to train, fine-tune or benchmark AI models. A time series is exactly the stored database that clause rules out.
Where the two overlap
Three places, and each one has a practical use.
Fan-out. An agent that searches four times before answering is doing in your code what an AI engine does inside its own pipeline, which our explainer on query fan-out describes in detail. The sub-queries decide the source pool in both cases. The difference is observability: you can log your agent's queries from your own code, while the engine's fan-out has to be collected, which is why several visibility APIs return the fan-out questions alongside the answer.
Page fetching. Once a visibility API tells you that a competitor's comparison page was cited in eleven answers last week, you still have to read that page. Tavily's Extract endpoint bills one credit per five successful URL extractions, and it never charges for a URL that fails. Exa's /contents endpoint costs $1 per 1,000 pages per content type. Both are a sane second half of a pipeline whose first half is an answer collector.
The engines themselves. The line between the two categories moves. In our October 23, 2026 survey, SerpApi lists Brave AI Mode among the AI surfaces it collects and LLM Pulse lists Brave AI among its 18 engines, while Profound lists Exa Search on its pricing page as one of up to nine answer engines tracked, without disclosing the method. Two of the vendors in this article are search infrastructure for other people's agents and, through their own consumer surfaces, something a visibility API tracks.
Mistakes we would avoid
- Reading a search ranking as an AI source pick. The overlap figures above put that at roughly half, on the one public comparison we found.
- Comparing prices without counting results. Exa's base rate covers ten results and adds $1 per 1,000 beyond that; Tavily's advanced depth doubles the credit; Perplexity bills per request and lets one request carry five queries.
- Measuring with the engine's own API and reporting it as the app. Perplexity's FAQ states that "The API uses the same search system as the UI with differences in configuration, so their outputs may differ", and that the underlying model can differ too.
- Building a stored time series on a feed whose terms forbid storage. Read the use restrictions before the pricing table.
- Buying an agent search API to answer a marketing question. The cheaper unit price is cheap because the unit is a list of links.
Our view
Pick by the question on the ticket. If the sentence you owe someone starts with "our agent needs to read the web", one of Tavily, Exa or Brave answers it, and the deciding criteria are index independence, the shape of the extracted text and the right to store what you fetch. If the sentence starts with "our brand appears in", no amount of web search gets you there, because the measured object is an answer written by an engine you do not control.
Teams building a GEO product end up paying for both, and in our view that is the right bill: one feed produces the answers you report on, the other reads the pages those answers cite. Our ranking of AI visibility APIs for developers weights the collection-side criteria, a self-serve key, a published price per answer and the tooling around it.
Frequently asked questions
Can I track my brand in ChatGPT with Tavily, Exa or Brave?
No. Those three return web pages, with a title, a URL and extracted text, so a model you run can read them; none of them calls ChatGPT, Perplexity or Google AI Mode, and none returns an answer written by those engines. To read what an AI engine said you need either that engine's own API, which measures the configuration you sent rather than the consumer app, or an API that collects the answers, priced between $1.20 and $15.75 per 1,000 answers across the APIs this site rates, checked on October 23, 2026.
What is the difference between a SERP API, a web search API for agents and an AI visibility API?
A SERP API returns a search engine's results page as JSON, including the blocks around the blue links. A web search API for agents queries its own index or a blend of sources and returns ranked pages with clean text, built to be fed to a model: $5 per 1,000 requests on Brave's Search plan, $7 per 1,000 on Exa's auto search type, one credit at $0.008 on Tavily. An AI visibility API runs your prompt on ChatGPT, Perplexity, Claude or Google AI Mode and returns that engine's answer with the brands and sources inside it, which is the only one of the three that measures what a buyer reads.
How much do Tavily, Exa and Brave cost per 1,000 searches?
Tavily bills one credit for a basic, fast or ultra-fast search and two credits for an advanced search, at $0.008 a credit pay as you go, down to $0.005 on its $500 Growth plan, with 1,000 free credits a month on its Researcher tier. Exa charges $4 per 1,000 requests on its lite search type and $7 on auto, both covering ten results, plus $1 per 1,000 extra results, with $10 of credit that resets on the first of each month. Brave's help page states that $5 of free monthly credit is equivalent to 1,000 queries on its Search plan, which puts that plan at $5 per 1,000 requests. Rates read on November 12, 2026.
Can I store Brave Search API results to build a visibility time series?
Brave's terms of use, last updated September 1, 2026, say a customer shall not "store, cache, or create a database of Search Results, in whole or in part, other than transient storage required for operation of Customer Applications", and its FAQ states that "It is prohibited to retain any and all data received through the Brave Search API". A separate clause forbids using the results to train, fine-tune or benchmark AI models. Brave asks you to contact its API support team if you need either right, so a stored month-over-month history on that feed is a contract question before it is an engineering one.
Does Tavily's include_answer or Brave's Answers API count as an AI answer I can measure?
It is an answer written by that vendor's model over the results that vendor fetched, not the answer ChatGPT or Google gave your buyer. Tavily's include_answer takes basic or advanced and adds no documented credit surcharge beyond the search depth. Brave's Answers service publishes its own formula, searches at $4 per 1,000 plus input and output tokens at $5 per million each, with a default limit of two requests per second. Both are useful inside a product and neither supports a sentence about your share of voice in a named engine.
Which search API should my agent use if it has to cite its sources?
All three return source URLs, and the difference is how much page text comes with them. Exa returns query-relevant highlights or full text in the same call, and its documentation says highlights "give models and agents the evidence they need without filling the context window". Tavily returns a short content field per result, with raw_content in markdown or text on request, up to 20 results per search. Brave's LLM Context endpoint returns pre-extracted snippets per URL under a token budget that defaults to 8,192 and accepts up to 32,768, and it is included in the same Search plan as web search.