Skip to content

Articles / How engines answer

How AI answer engines choose which brands to name: searches, sources and what the engines publish

AI engines name brands found in the pages their searches return. What Google, OpenAI, Anthropic and Perplexity publish about it, plus our observed data.

Published
Reading time
10 min read
By
AmpliRank
Tags
AI Search · Citations · Query Fan-out · AI Crawlers

AI answer engines with web search on name the brands they find in the pages their searches return. Each engine decides whether to search, turns the question into one or more searches, reads the results and cites some of them; the brands in those cited pages are the brands the answer can name. None of the engines publishes how it weighs one brand against another, so the useful evidence is what they do publish and what their answers cite.

In short

  • Google, OpenAI, Anthropic and Perplexity all document the same broad steps: decide whether to search, search (often more than once), then answer with cited sources.
  • Google says its AI features may run several related searches across subtopics, a step it calls query fan-out. Anthropic says comparative research can take 10 or more searches.
  • No engine publishes a ranking of brands or a list of factors for naming one. Google says there are no special optimizations needed for its AI features beyond being indexed and eligible for a snippet.
  • What you can act on: whether the engines’ crawlers can reach you, whether your pages answer the searches they run, and whether you appear in the third-party pages they cite.
  • In our own data, 130 of 346 cited sources across 45 answers were “best tools” roundups. The pages the engines read are often not the brands’ own.

This guide covers:

What decides which brands an AI answer names?

The pages an engine reads decide which brands it can name. An answer written from search results draws its facts, and its brand names, from the pages those searches returned and the engine chose to use.

That makes brand naming a chain of four steps, each documented to some degree by the engines themselves:

  1. The engine decides whether to search. Without a search, it answers from what the model learned in training.
  2. It turns the question into searches. Often several, each narrower than the buyer’s words.
  3. It reads the results and picks sources. Some of what it reads is cited; some isn’t.
  4. It writes the answer. The brands named are the ones the chosen sources support.

A brand can drop out at any step: if no search is run, if no search finds a page about it, or if the pages found don’t mention it. The rest of this guide takes the steps in order, using only what each engine publishes.

When does an engine search before answering?

Engines search when the question needs current or specific information, and the model makes that call itself. All four publishers describe the decision the same way.

Engine What its documentation says about the decision to search
Claude (Anthropic) Claude “determines when to search based on the prompt”, and searches for information that is current, changing or about “specific organizations, people, or products that might have changed”.
Gemini (Google) “The model analyzes the prompt and determines if a Google Search can improve the answer.”
OpenAI models “The model can choose to search the web or not based on the content of the input prompt.”
Perplexity Perplexity searches “the internet in real-time” to answer a question.
Google AI Overviews and AI Mode AI Overviews are shown “when our systems determine that it is additive to classic Search”.

A question about which software to buy names specific products whose details change, the kind of question Anthropic lists as a reason to search. In our own data, 40 of 45 answers to 15 buyer questions about AI-visibility tools drew on web search (88.9%, range 76.5% to 95.2%), on 13 November 2026 (data note).

How does one question become several searches?

Engines often run several searches for one question, each aimed at part of it. Google calls this query fan-out: its guidance says AI Overviews and AI Mode may issue “multiple related searches across subtopics and data sources” to build a response.

The other engines document the same behaviour in their own terms:

  • Gemini: “the model automatically generates one or multiple search queries and executes them.”
  • Claude: searches “can repeat multiple times throughout a single request”. Anthropic’s guide adds that single-fact questions typically use 1 to 3 searches, while “comparative or multientity research can use 10 or more.”
  • OpenAI: a reasoning model “can refine its searches as it reads results.”

A buyer’s “best tool for X” question is comparative and multi-entity by nature, so expect many searches behind it. Those searches, not the buyer’s exact words, decide which pages get read. What is query fan-out? shows what a fanned-out question looks like and how to use the searches.

In our data the engines searched for things like “AI search visibility tracking tools pricing comparison 2026”, “AI search visibility tracking tools for SEO agencies” and “LLM brand visibility tracking platform weekly executive report”. Each is narrower than the question asked, and each returns its own set of pages.

Which pages do the engines cite?

The engines cite a selection of the pages they read, and they say so. The cited sources are the best public evidence of which pages shaped an answer.

Engine What its documentation says about citations
Claude “Citations are always enabled for web search.” Each citation carries the page’s URL, title and up to 150 characters of the cited text.
OpenAI models Responses “include inline citations for URLs found in the web search results”. A separate list of every URL consulted is often longer: “The number of sources is often greater than the number of citations.”
Gemini Each citation annotation “links a text segment … to a source URL.”
Perplexity “Each answer includes numbered citations linking to the original sources.”
Google AI features AI Overviews and AI Mode “display a wider and more diverse set of helpful links” than classic results.

OpenAI’s note matters for measurement: the cited list is shorter than the list read. A page can inform an answer without being cited, so citations undercount influence. They are still the only part of the reading the engines show.

What gets cited, in practice? In 45 answers from Gemini, Claude and Google AI Mode on 13 November 2026, the 346 cited sources were mostly third-party pages:

Cited page type (by title) Citations Share of 346 95% range
Roundup (“best”, “top N”, “tools”) 130 37.6% 32.6% to 42.8%
How-to or guide 35 10.1% 7.4% to 13.7%
Comparison (“vs”, “alternatives”) 0 0.0% 0.0% to 1.1%

Reddit was the domain cited in the most answers, 13 of 45. The data note on roundups has the method, the full tables and the limits: one category, one day, three engines, and citations that aren’t independent of each other.

What do the engines say about ranking or optimization?

Little, and Google says the least is needed. Its guidance for site owners states: “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.” A page must be “indexed and eligible to be shown in Google Search with a snippet”, and “There are no additional technical requirements.”

Google adds that “You don’t need to create new machine readable files, AI text files, or markup to appear in these features.” Its recommended practices are ordinary search hygiene:

  • allow crawling in robots.txt and in any CDN or hosting layer;
  • make content findable through internal links;
  • keep important content in text form;
  • make structured data match the visible text;
  • provide a good page experience.

None of OpenAI, Anthropic or Perplexity publishes ranking factors for which brands an answer names. Anything more specific you read about “how ChatGPT ranks brands” is inference from observed answers, not documentation. Treat it the way you’d treat any observed pattern: as a hypothesis to measure.

Can a site be left out of AI answers by accident?

Yes: by blocking the crawler an engine uses for search. Each engine documents a separate agent for search, and blocking it can remove your pages from that engine’s answers while leaving others untouched.

Engine Crawler for search What the documentation says blocking it does
ChatGPT search OAI-SearchBot “Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers”, though they can still appear as navigational links.
Claude Claude-SearchBot, Claude-User Blocking either may reduce your site’s visibility in user search results or user-directed web search.
Perplexity PerplexityBot Perplexity recommends allowing PerplexityBot “to ensure your site appears in search results”.
Google AI Overviews and AI Mode Googlebot Googlebot’s robots.txt rules are the control; nosnippet, data-nosnippet, max-snippet and noindex limit what is shown.

These search crawlers are distinct from the crawlers some publishers use for model training, so a site can refuse training and still be readable for search. Check your robots.txt, your CDN’s bot rules and any firewall rules: a rule written to block training can catch a search agent too.

Check what the engines can reach. A free AmpliRank account includes the AI readiness check of your site, which shows whether the engines’ crawlers can read it. Run the readiness check.

What does this mean for a content team?

It means working on three things the engines document and one you can observe. The documented steps say where a brand drops out; the cited sources say which pages to work on.

If the engines… Then the gap is… What to do
can’t crawl your site for search access Allow the search crawlers; check CDN and firewall rules.
run searches no page of yours answers your pages Write the page for the sub-question they search: pricing, security, comparisons, use cases.
cite third-party pages that leave you out other people’s pages Earn an accurate place in the roundups, review sites and threads they cite.
cite pages that describe you wrongly accuracy Correct the facts at the source, and state them plainly on your own pages.

The order of that work, with checklists, is in how to get your brand named in AI answers. The product tour shows the searches and cited sources for a sample run.

What can’t the documentation tell you?

The documentation can’t tell you why one brand was named and another wasn’t. The engines publish the steps, not the weights.

Three more limits apply:

  • Documentation changes. The quotes here were checked on 14 November 2026. Engines update their search behaviour often, and their pages with it.
  • Developer documentation describes the engine’s search tools. It is the most specific public description each publisher gives, but a consumer product can add its own behaviour on top.
  • Observed answers vary. The same question gets different answers, searches and citations each time. Any pattern you see in a handful of answers needs a sample and a range before it becomes a decision; how to measure AI visibility explains how.

Where to start

Pick three buyer questions that matter to you. Ask each in the engines your buyers use, more than once, and write down the sources each answer cites. Mark which sources mention you, which don’t, and which are yours. Then check your robots.txt against the crawler table above. Those two lists, access and sources, are the start of your plan.

Sources

■ Ask · Measure · Compare · Trace · Fix · Share

Measure what AI says about you.

Cookies