How AI Engines Decide What to Recommend

Two mechanisms, not one

When a user asks an AI engine for "the best CRM for a ten-person sales team", the answer comes from two places:

1. Training data. The model recalls patterns from the text it was trained on. Brands that appear often, in authoritative sources, in the context of a category, are the ones it names first.
2. Retrieval. Engines with browsing (Perplexity, Copilot, Google AI Overviews and AI Mode, ChatGPT with search) fetch pages at answer time and cite them. Here the question is whether your page is indexed, ranks for the query and contains a passage that answers it directly.

Citadex's Citation Sources page shows you the second mechanism at work: the exact domains and pages each engine cited for your prompts.

Six factors you can influence

• Entity clarity. Say plainly what you are, for whom, and what you do. Vague positioning is the most common reason an engine describes a brand as the wrong kind of company. Citadex flags this as a context mismatch in Prompt Tracking.
• Direct answers. A sentence that answers the question in one line is quotable. Marketing prose is not.
• Structure. Headings that match the questions, FAQ sections, comparison tables and schema markup make passages easy to extract.
• Citation network. Review sites, community threads, editorial coverage and comparison articles that mention you are what retrieval engines pull from. The Citation Sources gap map ranks them by how often engines cite them without you.
• Recency. Retrieval engines prefer recently updated pages. Republishing an improved page beats leaving it untouched.
• Sentiment. Engines aggregate how sources talk about you. The Sentiment view under the Workspace shows which sources shape negative framing.

Why a competitor appears instead of you

Usually one of these: they are covered by more third-party sources the engine trusts, they have a comparison or list page that answers the prompt directly, their brand entity is clearer, or their content is newer. The Competitors page diagnoses this per prompt (missing comparison, missing guide, missing citable asset) and generates the brief to fix it.

Cited is not recommended

AI can cite your page and still recommend someone else. The Workspace card From mention to recommendation separates the steps: mention you, cite your site, recommend you, first pick, and how many of the answers that cite you also recommend you. Cited but rarely recommended usually means the page gives facts but no reason to choose you (use cases, comparisons, pricing, real examples). See From mention to recommendation.

Each answer gets one of five recommendation levels: first pick, recommended, conditional (recommended only under a condition such as “if your budget allows” or “if you need X”; counted on its own, not in Recommend you), listed option, or mentioned only. Being first in a list is not a first pick; criticism or “I’d avoid it” counts as mentioned only. Levels are judged by AI, and the evidence panel shows the words the level rests on plus “AI judged” or “Rule-based”. A language switches to AI judging only after it reaches 90% accuracy and an F1 of 0.85 for every level on a labelled set; until then the keyword rules stay. The switch day is marked on trend charts and earlier data isn’t recalculated. If a level looks wrong, use “Judged wrong?” in the evidence panel.

AI agents choose too

More and more people ask AI agents to find a price, compare products or fill in an inquiry form for them. Agents read only the page HTML; they run no scripts and do not log in. If your price appears only through scripts, forms have no labels or key pages sit behind a login, the agent gets stuck and picks someone else. AI Traffic → Agents checks this step; see Agent readiness.

How long changes take

Retrieval-based engines can reflect a new page within days once it is indexed; Citadex pings Bing IndexNow and Google Search Console on publish to speed that up. Changes that depend on training data take months and depend on being covered by sources the next model version learns from.