What is a consideration set in AI search?

The consideration set is the pool of pages an engine pulls in as candidates when it builds an answer with live search. It is also called the candidate set, and the pages in it are called retrieval candidates.

Three things define it:

• It is per query. A different question, or the same question reworded, can produce a different set.
• It is per engine. ChatGPT search, Perplexity and Google AI Overviews each assemble their own.
• It only exists when search runs. If the engine answers from memory, there is no set at all. See retrieval-on vs retrieval-off answers for that split.

Think of four stages: what the model already knows, access, retrieval and generation. The consideration set is the output of retrieval.

1. Access. A crawler must be allowed and able to fetch your page. That means per-bot robots.txt permission, a firewall or CDN that does not return 403, and content present in the HTML before JavaScript runs. For Google, Google says a page must be indexed and eligible to show with a snippet to appear as a supporting link in AI Overviews or AI Mode.
2. Query rewriting. OpenAI's help centre says ChatGPT search rewrites your prompt into one or more targeted queries. So the query that builds the set is not the one the user typed.
3. Retrieval. The engine runs those queries and keeps a set of candidates.
4. Generation. The model reads the candidates and cites some of them. The cited list is usually smaller than the retrieved list.

You can see the retrieved list through some APIs, which is how you log retrieved vs cited sources. Bing Webmaster Tools is the one place that shows real grounding queries, which you can then compare with your prompt set.

Almost all GEO research starts after the set is built. A 2026 critical survey of 45 GEO studies (arXiv 2607.14035) concluded that already-retrieved content can change its own citation, but no technique shows a stable, cross-platform effect on discoverability.

C-SEO Bench (NeurIPS 2025) found most content-rewriting methods ineffective or harmful. It also found moving a document higher in the model's context mattered more. That position was set by the researchers, not earned through any index.

Ahrefs tracked 1,885 pages adding schema and found AI Overview citations fell 4.6%. Every page in that study already had 100+ AI Overview citations. So the study cannot say whether schema helps a page that is not yet in the set.

What that means: if you are absent from answers, the first question is access, not copy. Check crawler permission, firewall behaviour, rendering and index status. Where engines cite other sites, being named on those sites is associated with appearing in answers, though no effect is promised.

• "Rank top 10 and you are in." Ahrefs found 37.9% of AI Overview citations also appeared in the top 10 results (March 2026). Rank and the set overlap, but they are not the same thing.
• "My tracked prompts show the set." Prompts are a proxy. Engines rewrite them, so your list approximates an input you cannot see.
• "Not cited means not retrieved." A page can be retrieved and dropped at generation.
• "Schema or llms.txt gets you in." Untested for entry. Google says structured data is not required for its AI features and that Google Search ignores llms.txt.
• "Allowing GPTBot puts you in ChatGPT's set." GPTBot governs training. OAI-SearchBot governs ChatGPT search.

• Retrieval-on and retrieval-off: whether a set exists for an answer. Read the definition.
• Grounding: the engine basing its answer on retrieved pages.
• Citation: a candidate the model chose to credit.

For the full path from retrieval to citation, read how AI search picks sources. To check whether a given answer used live search, see how to tell if an AI answer used web search.