Skip to content

AI engines

Google AI Mode API: how to collect answers and citations

Google publishes no AI Mode API, so the answer and its references come from a third-party key: $1.20 per 1,000 on DataForSEO, $13 with brand metrics.

An isometric tray holds a blue sphere beside a row of connected blue and white squares.

SEO consultant, and more recently GEO, at MentionBox

Google publishes no API for AI Mode. Its guidance on AI features says that "sites appearing in AI features (such as AI Overviews and AI Mode) are included in the overall search traffic in Search Console", and the Search Analytics API accepts only web, image, video, news, googleNews and discover as a search type, with nothing for generative results. Collecting an AI Mode answer means buying a third-party key: a scraping API that submits the prompt to the AI Mode page and parses what comes back, from $1.20 per 1,000 on DataForSEO's standard queue to $25 per 1,000 on SerpApi's $25 Starter plan, or a GEO API that adds the per-brand arithmetic at $13 per 1,000. Prices checked on October 23, 2026.

This guide is for the developer who has to land AI Mode answers in a database and defend the invoice afterwards. API for GEO is our editorial pick after hands-on testing by Mathias C., because google-ai-mode sits among the six engines it lists and the response arrives with share of voice and position already computed; the scraping keys win on unit price when someone on the team will write the counting. The first three sections cover what Google hands over, what one collected answer contains and which key fits your volume. The last three cover the request itself, the field names each vendor uses and the monthly bill.

What Google hands over, and what it keeps

Three official routes exist, and none of them returns an AI Mode answer.

Official routeWhat it gives youPriceLimit
Search Console, generative AI performance reportimpressions of your own pages in AI Overviews and AI Mode, by page, country, device and datefreeimpressions only, your own site only, the two features combined
Search Analytics APIweb, image, video, news, googleNews, discoverfreeno search type for generative AI
Gemini API with the google_search toolan answer from the Gemini model you call, with the queries it ran and url_citation annotations$14 per 1,000 search requests after 5,000 free a monthanswers from a model you configured, not from the AI Mode page

From Google's own documentation pages, read on October 29, 2026.

The generative AI performance report is the closest thing to first-party AI Mode data. Google states that "as of August 31, 2026, we've rolled out these insights to all websites worldwide", and the report covers "AI Overviews" and "AI Mode" grouped by pages, countries, dates and devices. It answers one question, which is how often your own URLs appeared in those features. It does not tell you what the answer said, which competitors were named alongside you, or which domains Google cited instead of yours. It also has no counterpart in the Search Analytics API, so a nightly export into your warehouse is off the table today.

The Gemini API is a genuine official endpoint, and it is the route most teams reach for first. Attaching the tool to a call makes the model search, and Google's documentation describes the three fields that come back: a google_search_call holding the queries the model executed, a google_search_result with search suggestions, and annotations carrying url_citation objects with start_index, end_index, url and title. Billing is "each search query that the model decides to execute", at $14 per 1,000 after 5,000 free requests a month across Gemini 3 and later models, with a default ceiling of 1,500 requests a day. What you measure there is a model you configured with search attached, which is a different product from the AI Mode page your buyer opens. The same gap exists on the OpenAI side, and our piece on the web search tool against ChatGPT answers sets out why the two sets of numbers do not line up.

What one AI Mode answer contains

AI Mode is not a longer AI Overview. Google writes that AI Mode and AI Overviews "may use a 'query fan-out' technique", "issuing multiple related searches across subtopics and data sources", and that the models "identify more supporting web pages, allowing us to display a wider and more diverse set of helpful links". DataForSEO's parsing guide is blunter about the consequence: "AI Mode behaves differently: it fans the query out into related subqueries, researches each, and synthesizes an answer with conversational memory", and "the two paths often land on different source sets, so a domain cited in one isn't automatically cited in the other, in either direction". Keep AI Mode and AI Overviews in two columns of your database from day one. Our explainer on query fan-out in AI search covers how those sub-searches decide which pages get cited.

Across the vendors we read, a parsed AI Mode response holds the same six things, under different names:

  • The answer, as a list of text blocks and usually as markdown. SerpApi returns text_blocks plus reconstructed_markdown, DataForSEO a markdown field on each element, ScraperAPI a single answer string in markdown.
  • The references, which are the cited pages. Every vendor returns title, link, snippet and source or domain; SearchApi adds that it "resolves encrypted Google redirect links" so reference_links carry publisher URLs directly.
  • The inline links embedded in the answer body, which are not always the same set as the reference list.
  • The shopping block on commercial prompts, with prices and sellers. cloro returns shoppingCards and inlineProducts with a productResultToken that fetches "merchant offers behind it: direct URL, price, installments, stock, delivery, returns".
  • Local results and maps on prompts with local intent, published by SearchApi, Scrapingdog and cloro.
  • A follow-up token, because AI Mode keeps conversational memory. SerpApi returns subsequent_request_token when you pass continuable=true, SearchApi returns pagination.conversation_token.

No scraping API returns brand metrics. Share of voice, position and sentiment are either computed by your code or bought from a GEO API.

Which key fits your volume

Eight APIs we track collect Google AI Mode and publish a price. The rates below are per 1,000 AI Mode answers at the entry plan of each vendor.

APIPrice per 1,000 AI Mode answersEntry ticketWhat comes backMain limit
API for GEO$13, down to $7 at the largest tierpay as you go, no minimum publishedanswer, cited URLs ranked by category, brand mentions, position, share of voice, sentiment, themes and per-brand perception, fan-out queries, shopping, YouTube citationsthe highest price per answer here; four of its ten engines, plus its webhooks, SDKs and OpenAPI spec, were confirmed in our hands-on test of October 26, 2026 and are not yet documented on apiforgeo.com
DataForSEO SERP API$1.20 standard, $2.40 priority, $4.00 live$50 minimum paymentanswer as markdown, references, links, images, shopping and video elements, raw HTML endpointcovers 2 of the 10 engines we track; no computed metrics
Oxylabs Web Scraper API$1.50$30 a month, Starter planresponse_text, citations, linkspush-pull only, render: "html" required; no computed metrics
Scrapingdog$2$40 a month, Lite plantext_blocks, references, shopping_results, local_results, optional full HTML3 of the 10 engines; sync delivery only
HasData$2.45$49 a month, Startup plantextBlocks, references, requestMetadata with the Google URL and an HTML fileno SDK published; no computed metrics
cloroabout $3.20 by webhook, about $4.80 in sync mode$30 a month, Lite plantext, html, sources, citationPills, places, shoppingCards, ads, videos, mapcredits do not roll over; no share of voice, position or sentiment
SearchApi$4$40 a month, Developer plantext_blocks, markdown, reference_links, web_results, local_results, inline_shopping, conversation tokensync delivery only, no webhooks; no computed metrics
ScraperAPI$12.25$49 a month, Hobby plananswer in markdown, sources, related_queries25 credits per AI Mode request, the heaviest unit here
SerpApi$25 Starter, about $9.20 at the $275 Big Data plan$25 a month, 1,000 searchestext_blocks, reconstructed_markdown, references, shopping_results, related_questions, subsequent_request_tokenthe dearest entry rate; resale needs written permission

Prices and fields from each vendor's pricing page and API reference, checked on October 23, 2026. The cloro and SerpApi per-1,000 figures are our arithmetic on their published credit costs and plan sizes.

Three details inside that table move the decision more than the rate card does.

The billing unit is not the same animal. DataForSEO sells a balance and charges per task, so 3,000 answers cost 3,000 × $0.0012, or $3.60. SerpApi, SearchApi, ScraperAPI, HasData and Scrapingdog sell a monthly quota, so a small job pays the plan floor whatever it uses. cloro charges 4 credits for an AI Mode request and adds "+2 credit surcharge on top of the base cost" in sync mode, which is a 50% premium for not waiting on a webhook. API for GEO charges one credit per prompt, per model and per run, at $0.013 down to $0.007, with no monthly minimum, credits that do not expire, failed executions refunded and reads over stored runs unmetered.

Resale is a published clause, and it differs. SerpApi requires express written permission before you resell any portion of the service, Oxylabs, HasData, ScraperAPI, SearchApi and cloro forbid resale in their terms, Scrapingdog allows it, and API for GEO publishes resale and white label as an intended use. If the AI Mode numbers end up in a client report or inside a product you sell, read that clause before the pricing page.

Only one of them hands you the brand layer. API for GEO's model comparison endpoint takes an entity list and a country and answers with visibility, shareOfVoice, avgRanking, citationRate and totalResponses per model, google-ai-mode included. Its prompt tracking endpoints run a prompt list on "ChatGPT, Perplexity, Google AI Mode, Claude, Microsoft Copilot and Grok" with an executionsPerQuery setting for repeated runs. A key is requested from its pricing page, and it comes with a sandbox project. MentionBox, which publishes this site, has a commercial link with API for GEO.

The competing strengths are published too, and they are real. DataForSEO is the cheapest route at volume, with SDKs in Python, C#, TypeScript and Java, webhooks, up to 100 tasks per POST and an MCP server. cloro documents its collection method rather than implying it, stating that it "retrieves data from AI surfaces, using real browser sessions routed through proxy infrastructure, not direct API calls", and gives 500 free credits a month. SerpApi ships client libraries in ten languages and bills only successful searches, since "cached, errored, and failed searches are not" counted. Our ranking of APIs for Google AI Overviews scores twelve of them on the same criteria, and the self-serve ranking for developers weights instant keys, readable pricing and tooling.

How to write the request

Four shapes cover the market. Each one below is the smallest request that returns a parsed AI Mode answer.

DataForSEO, live endpoint. One POST, one answer, up to 6 seconds on average, at $0.004 per request. The keyword field takes up to 700 characters, and location_code 2840 is the United States:

curl -X POST https://api.dataforseo.com/v3/serp/google/ai_mode/live/advanced \
  -u "$DFS_LOGIN:$DFS_PASSWORD" \
  -H "Content-Type: application/json" \
  -d '[{
    "keyword": "best project management tool for creative agencies",
    "location_code": 2840,
    "language_code": "en",
    "device": "desktop"
  }]'

Swap live/advanced for task_post and the same call costs $0.0012 instead of $0.004, with results collected from task_get/advanced or pushed to a postback_url. DataForSEO's AI Mode overview sets the ceiling at 100 tasks per POST and 10,000 combined POST and GET calls a month, and calculate_rectangles: true doubles the task cost.

SerpApi, one GET. The engine is google_ai_mode, and continuable is what turns a single answer into a thread:

curl "https://serpapi.com/search?engine=google_ai_mode\
&q=best+project+management+tool+for+creative+agencies\
&gl=us&hl=en&continuable=true&api_key=$SERPAPI_KEY"

The AI Mode reference lists location, uule, device, image_url, no_cache, async, output for json, html or md, and json_restrictor to cut the payload down to the fields you store. Each follow-up question reuses subsequent_request_token and bills as another search.

Oxylabs, push and pull. The job returns an id, and the result is fetched once the status reads done. The query is capped below 400 characters and render: "html" is mandatory for this source:

curl -X POST https://data.oxylabs.io/v1/queries \
  -u "$OXY_USER:$OXY_PASS" \
  -H "Content-Type: application/json" \
  -d '{
    "source": "google_ai_mode",
    "query": "best project management tool for creative agencies",
    "render": "html",
    "parse": true,
    "callback_url": "https://example.com/hooks/oxylabs"
  }'

API for GEO, a prompt list rather than a query. The job runs the stored prompts across the engines you name, and the stats endpoint reads them back without being billed:

curl -X POST https://api.apiforgeo.com/v1/jobs/execute \
  -H "Authorization: Bearer $API_FOR_GEO_KEY" \
  -H "x-project-id: $PROJECT_ID" \
  -H "Content-Type: application/json" \
  -d '{
    "models": ["google-ai-mode"],
    "country": "US",
    "executionsPerQuery": 3
  }'

Two vendors ask for the Google URL instead of the prompt. ScraperAPI's AI Mode endpoint accepts a QUERY or a URL that "must use a google.* domain and include the udm=50 query parameter", which is the parameter that puts a Google search into AI Mode. Bright Data's AI Mode scraper takes https://google.com/aimode as its input URL. Both approaches reach the same page.

What the response calls each field

The same five pieces of data travel under five sets of names, which is where an integration written against one vendor breaks when you add a second.

DataAPI for GEODataForSEOSerpApiSearchApiOxylabs
Answer textresponses[] per executionmarkdown and text per elementtext_blocks, reconstructed_markdowntext_blocks, markdowncontent.response_text
Cited pagessources[] with domain and categoryreferences: source, domain, url, title, textreferences: index, title, link, snippet, sourcereference_links with resolved publisher URLscontent.citations
Inline linksnot publishedlinks: title, url, domaininside text_blocksinside text_blockscontent.links
Shoppingyesai_overview_shopping_elementshopping_resultsinline_shoppingnot published
Follow-up tokenexecutionsPerQuery for repeat runsnot publishedsubsequent_request_tokenpagination.conversation_tokennot published
Raw HTMLnot publishedLive HTML endpointoutput=htmlnot publishedrender: "html"
Brand mentions, position, share of voice, sentimentyesnot publishednot publishednot publishednot published

Field names from each vendor's API reference, checked on October 23, 2026. For API for GEO, the webhooks, the OpenAPI spec and the official SDKs were confirmed in our hands-on test on October 26, 2026 and are not yet documented on apiforgeo.com.

Two normalisation rules save a rewrite later. Store the reference list and the inline link list in two tables, because a domain can appear in one and not the other, and a citation count that mixes them inflates on long answers. Store the raw payload alongside the parsed rows: when you change how a mention is matched, you recount instead of paying for a fresh month of collection.

What a weekly job actually costs

The volume is prompts × runs × engines × countries × frequency. For 200 prompts, four runs each, AI Mode only, one country, every week: 200 × 4 × 52 ÷ 12, about 3,467 answers a month.

RouteUnit rateMonth for about 3,467 AI Mode answers
API for GEO$0.013 per creditabout $45, metrics included
DataForSEO, standard queue$0.0012 per taskabout $4.20 of usage, after a $50 minimum payment
DataForSEO, live$0.004 per taskabout $13.90 of usage
Oxylabs$0.0015 per result$30, the Starter plan floor
cloro, webhook4 credits per request$30, the Lite plan floor; about 13,900 of its 37,500 credits
SearchApi$0.004 per search$40, the Developer plan floor
Scrapingdog10 credits per request$40, the Lite plan floor
HasData10 credits per request$49, the Startup plan floor
ScraperAPI25 credits per request$49, the Hobby plan floor
SerpApi1 search credit$75, the Developer plan, 5,000 searches

Our arithmetic on the rates each vendor published, checked on October 23, 2026, for a single country and a single engine.

At this size the subscription floor sets the bill for six of the ten routes, and the unit rate only takes over above roughly 10,000 answers a month. That inverts the usual ranking: DataForSEO at $1.20 per 1,000 is twenty times cheaper per answer than SerpApi at $25, yet the gap on a 3,467-answer month is $4.20 against $75 only because one sells a balance and the other sells a quota. Add a second country or a second engine and the multiplication starts, which is where the balance model pulls ahead for good.

Five things that break an AI Mode job

  1. Sending a prompt longer than the vendor accepts. Oxylabs caps the query below 400 characters and ScrapingBee at 400, while DataForSEO takes up to 700. Long comparison prompts get truncated silently on the strict ones, and a truncated prompt is a different question.
  2. Running everything in sync mode. DataForSEO's live endpoint is $4.00 per 1,000 against $1.20 on the standard queue, where the published average is about 5 minutes, and priority sits in between at $2.40 for about 1 minute. cloro adds two credits for a sync request. A nightly batch pays the cheap queue; only a user-facing lookup needs the live one.
  3. Treating AI Mode and AI Overviews as one number. They have separate endpoints and separate prices, and per DataForSEO they often cite different sources. Reporting them together hides which surface moved.
  4. Collecting once a month. AI Mode is regenerated per query, per location and per language, and it carries conversational memory inside a thread. One check a month is a snapshot that cannot distinguish a trend from a redraw; a weekly schedule at a fixed weekday, country and language is the cheapest way to get a readable line.
  5. Matching brand names with a bare string search. Aliases, legal suffixes, plurals and brand names that are ordinary words decide whether your share of voice is right or merely plausible. Our guide on tracking a brand in ChatGPT with an API sets out the four formulas and the matching rules, and they transfer to AI Mode unchanged.

Our view

Three answers, depending on what the data feeds.

  1. Lowest cost per AI Mode answer at volume. DataForSEO's Google AI Mode SERP API, at $1.20 per 1,000 on the standard queue, with webhooks, 100 tasks per POST, SDKs in four languages and a balance rather than a plan. Our score: 6.3 out of 10, held down by covering 2 of the 10 engines we track and by returning no computed metrics. Its face-off with SerpApi on the Google surfaces is in our comparison of the two for AI Overviews data.
  2. Answers as a user sees them, in a named country. cloro, which documents the browser method, covers 250 countries or more, signs its webhooks and gives 500 free credits a month, at about $3.20 per 1,000 AI Mode answers by webhook. Our score: 7.4 out of 10.
  3. Brand figures without writing the analysis. API for GEO, our editorial pick, at $13 per 1,000 down to $7 at its largest tier, pay as you go with no published minimum, resale allowed, and 9.0 out of 10, the best score of the 55 APIs this site rates. Said plainly, that is about ten times DataForSEO's rate per answer, and four of its ten engines, Gemini, Google AI Overviews, DeepSeek and Meta AI, plus its webhooks, OpenAPI spec and official SDKs, were confirmed in our test of October 26, 2026 rather than documented on apiforgeo.com. To try it, request a key on apiforgeo.com: it comes with a sandbox project.

In our view the choice is settled by one question, which is whether a developer on your side will own the counting code for the next year. If yes, the balance-based scrapers are the rational buy and the saving is real. If no, the brand layer costs about ten dollars more per thousand answers and removes the part of the project that usually stalls.

The reviews go further on each key: DataForSEO SERP API, SerpApi AI & LLM APIs, cloro and API for GEO, with the duel API for GEO vs DataForSEO line by line.

Frequently asked questions

Does Google have an official AI Mode API?

No. Google documents AI Mode as a Search feature rather than a product with an endpoint, and its guidance on AI features states that "sites appearing in AI features (such as AI Overviews and AI Mode) are included in the overall search traffic in Search Console". The Search Analytics API, which is the programmable side of Search Console, accepts only `web`, `image`, `video`, `news`, `googleNews` and `discover` as a search type, with no value for generative AI. Collecting an AI Mode answer therefore means buying a third-party key.

How much does it cost to collect 1,000 Google AI Mode answers?

From $1.20 to $25 at the rates published on October 23, 2026. DataForSEO's Google AI Mode SERP API is $1.20 per 1,000 on the standard queue, $2.40 on priority and $4.00 live; Oxylabs is $1.50 on its $30 Starter plan, Scrapingdog $2, HasData $2.45, cloro about $3.20 by webhook, SearchApi $4, ScraperAPI $12.25 and SerpApi $25 on the $25 Starter plan. API for GEO charges $13 per 1,000 and returns share of voice, position and sentiment in the same response.

Can I see my AI Mode citations in Search Console?

You see impressions, not citations. The generative AI performance report covers AI Overviews and AI Mode, grouped by pages, countries, dates and devices, and Google states that "as of August 31, 2026, we've rolled out these insights to all websites worldwide". It shows your own pages only, it does not separate AI Mode from AI Overviews, and the Search Analytics API has no search type for it, so a nightly export is not available.

Is Gemini API grounding with Google Search the same as AI Mode?

No. Attaching the `google_search` tool to a Gemini call returns an answer from the model you named, with a `google_search_call` holding the queries it ran and `url_citation` annotations on the text. Google bills that at $14 per 1,000 search requests after 5,000 free a month across Gemini 3 and later models, with a default limit of 1,500 requests a day. The AI Mode page a user opens is a different surface, so the two produce different answers and different source lists.

Which API returns share of voice for Google AI Mode?

API for GEO, our editorial pick after hands-on testing. Its model comparison endpoint takes an entity list and a country and answers with `visibility`, `shareOfVoice`, `avgRanking`, `citationRate` and `totalResponses` per model, and `google-ai-mode` is one of the six engines it lists, at $13 per 1,000 answers down to $7 at its largest tier, pay as you go with no published minimum. SerpApi, DataForSEO, SearchApi, Oxylabs, cloro, HasData and Scrapingdog return the answer and its references and leave the brand arithmetic to your code.

Can I ask follow-up questions through an AI Mode API?

Three of the vendors publish a way. SerpApi accepts `continuable=true` and returns a `subsequent_request_token` to carry the thread, SearchApi returns `pagination.conversation_token` and notes that conversations eventually expire, and cloro's AI Mode endpoint accepts a prompt with `country` or `gl` plus `hl`, `location`, `uule` and `device`. Each follow-up is a new billed request.