Do LLMs always use real-time browsing?

No. ChatGPT (with browsing), Perplexity, and Google AI Overviews are browsing-grounded. Claude relies more on training plus optional browsing. Gemini blends both. Recommendations from non-browsing modes reflect training data, which lags by months.

Can I influence LLM recommendations without ads?

Yes. The signals are public — page copy, schema, comparison content, and outside corroboration — and they can be edited or published. There is no LLM ad placement; recommendations come from retrieval over public content.

Why does each LLM recommend different brands?

Each LLM has different training data, retrieval pipelines, ranking weights, and prompt-handling. Coverage on ChatGPT does not predict coverage on Claude. Running the same prompts across all 7 engines reveals the engine-specific gaps.

Are LLM recommendations affiliate-driven?

No major consumer LLM has disclosed affiliate-driven recommendations. The output reflects retrieval and ranking over public content. If a brand is consistently absent, the cause is signal-level, not commercial.

WhoCites

How Do LLMs Decide What to Recommend?

Q: How do LLMs decide what to recommend?

LLMs retrieve relevant chunks from training data and, when browsing, from live indexes, then rank candidates by category fit, factual density, schema clarity, and outside corroboration. The recommended brand is whichever passes the retrieval and ranking pipeline.

LLM product recommendations follow a measurable retrieval-and-ranking process. WhoCites measures the outcome across 7 engines and reports the signals competitors use to win the slot.

Last updated 2026-05-21

Run an AI visibility scan

Diagnostic answer

How do LLMs decide what to recommend? is not a mystery problem. It is usually a retrieval-surface problem: public pages either give AI systems enough clear evidence to quote, or they leave the engine to choose a better-documented competitor.

Symptom

The same prompt produces different recommended brands across ChatGPT, Claude, Gemini, Copilot, Perplexity, Grok, and Google AI Overviews.

Likely cause

Each engine has different training data, browsing access, freshness, source ranking, and confidence thresholds. A brand that is retrievable in one engine can be invisible in another.

What to check

Check the public evidence layer before changing the product. The highest-signal checks are crawl access, answer-shaped copy, schema, citations, and whether the same buyer question is answered on one canonical URL.

Whether the answer is browsing-grounded or training-grounded.
Which source pages are cited when the engine exposes citations.
Whether the brand has one canonical page for the exact recommendation intent.
Whether competitors have clearer category pages, comparison pages, or third-party mentions.
Whether Search Console and Bing can see the changed pages.

What WhoCites measures

WhoCites does not guess from metadata alone. It runs the category prompts against live AI and search sources, then ties the result back to the pages and signals that explain the miss.

Per-engine brand mention rate and rank.
Competitor displacement by prompt and engine.
Citation source patterns where links are exposed.
Prompt-fit weighting so adjacent/off-category answers do not distort the score.

What to fix next

The next fix should improve the source material AI can retrieve, not just the words on a landing page.

Build one stronger public source per high-value recommendation prompt.
Use methodology and benchmark pages as internal proof sources.
Add comparison or fit language only where it is factually true.
Wait for discovery before interpreting a post-fix re-scan.

How do LLMs decide what to recommend?

LLMs decide what to recommend through a two-step process: retrieve candidate chunks (from training data, from live browsing, or both) and rank them by category fit, factual density, schema clarity, and outside corroboration. WhoCites measures the retrieval and ranking outcome across ChatGPT, Claude, Gemini, Grok, Copilot, Perplexity, and Google AI Overviews for $49.

Retrieval is signal-driven, not random

Retrieval prefers chunks that are self-contained, factually dense, semantically aligned with the prompt, and supported by recognizable schema. Pages that are vague, unstructured, or contradictory get skipped even when they sit on high-authority domains.

Browsing-grounded versus training-grounded

Browsing-grounded LLMs (ChatGPT with web, Perplexity, Google AI Overviews) retrieve from live indexes; new content can appear in days. Training-grounded LLMs (Claude default, Gemini without retrieval) reflect data from the last training cut, so changes take longer to propagate.

Ranking weights vary by engine

Each engine weights signals differently. Perplexity is recency-biased. Google AI Overviews favors high-authority sources cited in Search. ChatGPT browsing favors chunk-shape and schema. There is no single ranking formula across all 7 engines — which is why cross-engine measurement matters.

Why competitors win the recommendation slot

Competitors win when their pages give the LLM a higher-confidence chunk to quote and a higher-authority source to cite. The fix is to publish equivalent or stronger signals on the brand's own pages — not to game the LLM but to match what the pipeline actually rewards.

What WhoCites reports on recommendation mechanics

WhoCites runs buyer-intent prompts across 7 engines and records which brands the LLMs recommend, brand rank when mentioned, competitor displacement, and citation sources. The report ties the outcome back to specific page-level signals so the fix is concrete.