2026-09-17, GEO · Fundamentals

The AI search engines in 2026: how each one grounds answers and shows sources

An AI search engine is a product that answers a question in generated prose after retrieving web content, and links the pages it used. In 2026 the ones that matter for most brands are ChatGPT search, Perplexity, the Gemini app, Microsoft Copilot, Grok, Google’s AI Overviews and AI Mode, and Claude with web search. They differ in where results come from, how sources are shown and which crawler decides whether your pages are eligible. Everything below is from each vendor’s own pages, checked 2026-09-17, with gaps in the documentation marked.

Summary table

Each row is expanded below. For the mechanics behind the retrieval column, and how to test them, see AI grounding by engine. For the crawler column in depth, see AI crawlers.

ChatGPT search

ChatGPT search is the web search capability inside ChatGPT. OpenAI’s ChatGPT Search help article (checked 2026-09-17) says it is available to Free, Plus, Team, Edu and Enterprise users, including logged-out free users, and that ChatGPT searches automatically when a question might benefit from web information. Users can also start a search manually.

Grounding: The same article says ChatGPT search “sometimes partners with other search providers”, typically rewrites the prompt into one or more targeted queries sent to those providers, and may send more specific follow-up queries after reviewing the first results. It lists Bing and Shopify as third-party search providers whose privacy policies apply. It also says ChatGPT uses general location derived from the IP address to improve local results, without sharing the IP address itself with providers.

Sources: Responses that use search “may include inline citations”; when they don’t, a Sources button under the response opens a panel with cited sources and other relevant links.

Eligibility: OpenAI’s crawler overview (checked 2026-09-17) says sites opted out of OAI-SearchBot “will not be shown in ChatGPT search answers, though can still appear as navigational links”, and that robots.txt changes take about 24 hours to take effect for search. The Publishers and Developers FAQ adds that ChatGPT appends utm_source=chatgpt.com to referral URLs.

Perplexity

Perplexity is an answer engine built around search. How does Perplexity work? (checked 2026-09-17) says it searches the internet in real time, lets Pro users pick the underlying model, and offers a Research mode that “performs dozens of searches automatically”.

Grounding: Perplexity runs its own index. Its post Introducing the Perplexity Search API (checked 2026-09-17) describes the API as giving access to “the same global-scale infrastructure that powers Perplexity’s public answer engine”, with an index covering “hundreds of billions of webpages”.

Sources: “Each answer includes numbered citations linking to the original sources.”

Eligibility: Perplexity’s crawler page (checked 2026-09-17) says PerplexityBot surfaces and links websites in search results and is not used to crawl content for AI foundation models, and recommends allowing it in robots.txt and permitting its published IP ranges.

Gemini (the app)

Gemini is Google’s assistant app, with mobile and web experiences according to Google’s overview.

Grounding: Google’s overview of the Gemini app (checked 2026-09-17; the page says it was last updated July 25, 2024) says Gemini “relies on external sources such as Google Search” and calls the process retrieval augmentation. The Gemini Apps help page on Connected Apps says Gemini automatically uses public information from Google Search, Google Flights, Google Hotels, Google Maps and YouTube.

Sources: View related sources & double-check responses (checked 2026-09-17) says that when sources are available, a Sources button appears at the bottom of the response or in-line, opening a side panel, and that “not all responses include related links or sources”.

Eligibility: Google’s common crawlers list (checked 2026-09-17) says the Google-Extended token manages whether crawled content may be used for training future Gemini models and “for grounding … in Gemini Apps and Grounding with Google Search on Vertex AI”, and that it does not affect inclusion or ranking in Google Search. That makes Gemini different from AI Overviews: the two are governed by different controls.

Microsoft Copilot

Microsoft Copilot is Microsoft’s consumer assistant, available free at copilot.microsoft.com according to Microsoft Support’s comparison of Copilot experiences (checked 2026-09-17), separate from the work versions tied to Microsoft 365.

Grounding: The Transparency Note for Microsoft Copilot (for individuals) (dated August 18, 2026, checked 2026-09-17) says that for conversations where users seek information, Copilot “is grounded in web search results”, centering its response on “high-ranking content from the web”. The note does not name the search engine. For the work product, Microsoft’s Copilot Chat privacy page says web grounding sends short generated queries to the Bing search service.

Sources: Grounded text responses carry “hyperlinked citations listed below the text”.

Eligibility: Bing’s crawler overview says robots.txt controls how Bing crawlers interact with a site, and Bing Webmaster Tools’ AI Performance report (February 10, 2026) shows citations of a site’s pages across Microsoft Copilot, AI-generated summaries in Bing and select partners.

Grok

Grok is xAI’s assistant on grok.com, in the mobile apps and on X.

Grounding: The Grok consumer FAQ (checked 2026-09-17) says Grok “has a unique feature that allows it to search public X posts and perform real-time web searches”. xAI’s developer docs describe separate Web Search and X Search tools for the API, with X Search supporting keyword, semantic, user and thread lookups.

Sources: The consumer FAQ does not describe the citation UI. The API’s citations documentation describes a full list of sources encountered plus optional inline [[N]](url) citations, and notes that not every URL in the full list is referenced in the answer. That describes the API; the app may render citations differently.

Eligibility: we found no first-party xAI page documenting crawler user agents as of 2026-09-17.

Google AI Overviews and AI Mode

AI Overviews and AI Mode are two AI features inside Google Search. Google’s AI features and your website (checked 2026-09-17) describes AI Overviews as a summary shown “only … when our systems determine that it is additive to classic Search”, and AI Mode as suited to queries that need “further exploration, reasoning, or complex comparisons”.

Grounding: Both “may use a ‘query fan-out’ technique, issuing multiple related searches across subtopics and data sources”, and they “may use different models and techniques, so the set of responses and links they show will vary”.

Sources: supporting links alongside the response. To be eligible, “a page must be indexed and eligible to be shown in Google Search with a snippet”, and “there are no additional technical requirements”.

Eligibility: Googlebot’s robots.txt rules, plus nosnippet, data-nosnippet, max-snippet and noindex. Google’s guide to generative AI features adds that site owners don’t need special machine-readable files such as llms.txt for Google Search, because Google Search doesn’t use them. More on that in llms.txt.

Claude with web search

Web search is a feature of Anthropic’s Claude apps. Anthropic’s help article Enable and use web search (checked 2026-09-17) says users switch it on per chat, and on Team and Enterprise plans an owner must first enable it for the organization.

Grounding: “Claude invokes a search tool to inform and ground its generated responses with content from the live web.” The consumer article names Bing only for image results (“Image search is powered by Bing”). It does not name the provider for text results.

Sources: “Every response includes citations”, with source links and relevant quotes where appropriate.

Eligibility: Anthropic’s crawler article (April 7, 2026) says disabling Claude-SearchBot “may reduce your site’s visibility and accuracy in user search results”, and disabling Claude-User may reduce visibility for user-directed web search.

What the documentation does and doesn’t tell you

Three patterns hold across all seven vendors:

1. Retrieval comes first, then writing. Every engine documents some version of “search, then answer with links”. None publishes how it picks which retrieved pages to cite. How AI engines choose citations treats selection as a hypothesis you test.
2. Several engines rewrite the question. OpenAI and Google document query rewriting and fan-out explicitly. Microsoft documents short generated Bing queries for Copilot Chat. The query that retrieves your page is often not the prompt a user typed (ChatGPT query fan-out).
3. Eligibility runs through a named crawler or index. For Google’s AI features it’s Googlebot and indexing in Search; for Gemini’s grounding, Google-Extended; for ChatGPT, OAI-SearchBot; for Perplexity, PerplexityBot; for Claude, Claude-SearchBot. Blocking the wrong token removes you from one engine while leaving the others untouched.

The documentation doesn’t tell you outcomes: whether your brand is named, which of your pages are cited, which competitors appear instead, and how that changes week to week. For that you need the answers themselves, collected on a schedule.

Monitoring the engines with one request format

This API returns each engine’s answer as JSON: the text, sources[], inline citationPills[] where the engine shows them, and, depending on the engine, the searches it ran. Claude is not covered.

Credits per request are explained in credits; plan prices are on /pricing.

A minimal cross-engine check with Python, run as async tasks:

import os
import requests

API = "https://api.answerline.dev"
HEADERS = {"Authorization": f"Bearer {os.environ['ANSWERLINE_API_KEY']}"}
PROMPT = "best project management software for a 20-person agency"

tasks = [
    {"taskType": "CHATGPT", "payload": {"prompt": PROMPT, "country": "US"}},
    {"taskType": "GEMINI", "payload": {"prompt": PROMPT, "country": "US"}},
    {"taskType": "COPILOT", "payload": {"prompt": PROMPT, "country": "US"}},
    {"taskType": "GROK", "payload": {"prompt": PROMPT, "country": "US"}},
]

ids = []
for task in tasks:
    resp = requests.post(f"{API}/v1/async/task", json=task, headers=HEADERS, timeout=60)
    resp.raise_for_status()
    ids.append(resp.json()["task"]["id"])
print(ids)

When a task is COMPLETED, GET /v1/async/task/{id} returns the answer under response.result. Store it with the engine, prompt, market and date. For scheduled prompt sets, submit batches with a webhook instead of polling (async tasks, webhooks).

Once results arrive, compute the same three numbers for every engine: whether your brand is named in text, whether your domain is in sources, and which competitors take the remaining citations. AI share of voice turns those into a report.

Pitfalls when comparing engines

• Don’t treat AI search as one channel. The engines use different indexes and controls, so a page can be eligible for AI Overviews and blocked from ChatGPT search at the same time.
• Field names differ by engine. Perplexity returns its searches as search_model_queries, ChatGPT as searchQueries (with include.searchQueries), Copilot and Grok as searchQueries. Check each engine’s schema in the API reference.
• One run is a sample. The same prompt can return different sources on repeated runs. See answer volatility before drawing conclusions.
• Keep the market fixed. ChatGPT documents location-aware rewriting, so pass country, and state for US prompts on the chat engines, and compare like with like.
• Vendor pages change. Several cited here were updated in 2026; re-check before acting and record the check date.

For per-engine request and response details, start with the engine pages, or make a first call with the quickstart.

Keep reading