AI assistants do not “decide” which brands to recommend — they assemble recommendations from whatever their retrieval layer happens to surface in the seconds after you ask, which is why a brand’s visibility can collapse by 50 percentage points in a month without the brand changing anything. That is not a hypothetical. Between early August and mid-September 2025, ChatGPT’s citation rate for Reddit fell from roughly 60% of responses to about 10%, per a Semrush study of 230,000 prompts and 100M+ citations (published November 10, 2025). Understanding the machinery behind that volatility is the point of this page — the hub for our AI visibility coverage.
The three-layer pipeline behind every brand recommendation
When someone asks “what’s the best CRM for a 10-person team,” the answer is produced by three distinct layers, and each one can make or break a brand’s presence.
Layer 1: Parametric memory. The model’s training data encodes brand associations frozen at the training cutoff. If a query triggers no web search, recommendations come entirely from this layer — which is months to years stale, and where documented failure modes like fabricated product details live (we track those in the Limitations Ledger).
Layer 2: Retrieval. For most commercial queries, assistants now search the live web. OpenAI’s help documentation confirms ChatGPT sends queries to third-party search providers such as Bing, and OpenAI operates its own crawler (OAI-SearchBot) for a supplementary index. Google’s Gemini API grounding documentation describes the same pattern: the model generates one or more search queries, executes them against Google Search, and synthesizes from the results. Critically, Google’s docs state that grounding excludes pages that have disallowed Google-Extended via robots.txt — meaning a single line in a text file can remove a brand’s site from Gemini’s evidence pool.
Layer 3: Synthesis. The model reads the retrieved candidates and composes an answer, citing some sources and silently absorbing others. The query it searched is rarely the query you typed — assistants rewrite prompts into narrower sub-queries, so the page that best answers a sub-query wins the citation, not the page that matches the original keywords.
The practical consequence: a brand can rank well in Google, be absent from Bing, and therefore be invisible in ChatGPT while dominating Gemini. There is no single “AI ranking.”
What large-scale measurement actually shows
The Semrush study cited above is the most transparent public dataset to date: 230K prompts, weekly snapshots over 13 weeks (July 14 – October 12, 2025), across ChatGPT search, Google AI Mode, and Perplexity. Its headline findings:
- ChatGPT cited Reddit in ~60% of responses in early August 2025, collapsing to ~10% by mid-September. Wikipedia fell from ~55% to under 20% in the same window.
- The collapse was ChatGPT-specific. On Google AI Mode, Wikipedia held near 3% of responses; on Perplexity, 0.8%.
- Google AI Mode’s top-cited domains — LinkedIn (~15% of responses), YouTube, Reddit, Google, Google Blog — are all properties Google owns or partners with.
- Perplexity’s top sources were Reddit, LinkedIn, NIH, Microsoft, and Google.
Two competing explanations exist for the September shock, and the disagreement is instructive. Many SEOs attributed it to Google removing the num=100 search parameter in mid-September 2025. Semrush’s own analysis pushes back: only ~34% of Reddit’s Google rankings sat in positions 21–100, which cannot explain a six-fold citation drop; their head of organic visibility argues OpenAI deliberately rebalanced to avoid over-citing a handful of domains. Nobody outside OpenAI knows. That opacity is itself a finding: citation share can be repriced overnight by an unannounced platform decision.
What the peer-reviewed and preprint evidence says moves recommendations
The academic record is younger than the vendor-blog record but far more careful.
The original Generative Engine Optimization paper (Aggarwal et al., arXiv:2311.09735, presented at KDD 2024) ran controlled experiments on 10,000 queries and found that adding citations, quotations, and statistics to a page boosted its visibility in generative-engine responses by up to 40% — while keyword stuffing, the classic SEO move, did nothing or backfired.
A June 2026 preprint, “Generative Engine Optimization at Scale” (arXiv:2606.20065), measured brand visibility across AI search engines and reported that roughly 78% of citations go to corporate websites, the single most-cited content format is the ranked “best-of” listicle (~21% of all citations), and sentiment — whether a brand is framed positively or negatively — flips about 6.7× more often than whether the brand is mentioned at all. If those numbers replicate, the volatile variable isn’t presence; it’s framing.
Related preprints — “Incumbent Advantage” (arXiv:2606.17443) on brand bias in LLM recommendations and GEO-Bench (arXiv:2605.29107) on ranking manipulation — document that models exhibit measurable preference for well-known incumbents and that recommendation rankings can be manipulated by third-party content the brand doesn’t control. None of these are settled science; all are more rigorous than any vendor case study.
Platform comparison
| Platform | Retrieval backbone | Top-cited domains (Semrush, Oct 2025) | Documented quirk |
|---|---|---|---|
| ChatGPT search | Bing API + OAI-SearchBot index (per OpenAI docs) | Reddit, Wikipedia, Medium, Forbes, LinkedIn | Sept 2025: Reddit/Wikipedia citation share cut ~5–6× in weeks, unannounced |
| Google AI Mode / Gemini | Google Search grounding; respects Google-Extended robots.txt | LinkedIn, YouTube, Reddit, Google, Google Blog | Top sources skew heavily toward Google-owned or partnered properties |
| Perplexity | Own crawler + live search per query | Reddit, LinkedIn, NIH, Microsoft, Google | Most stable citation mix of the three across the 13-week study |
How to think about this as a measurement problem
Brand visibility in AI answers is a benchmark like any other, and it inherits every methodological trap we cover in How to Evaluate LLMs: sampling variance, prompt sensitivity, and non-stationarity. A single “does ChatGPT recommend us?” check is one sample from a distribution that shifts with model version, retrieval index, time of day, and phrasing. The Semrush data shows the underlying distribution itself gets repriced without notice. Any vendor selling a single-number “AI visibility score” without disclosing prompt count, sampling window, and platform version is selling noise.
The retrieval layer is also where classic search infrastructure quietly re-enters the picture — index coverage, crawlability, canonical URLs. The mechanics we described years ago in how a search engine works didn’t become obsolete; they became the invisible substrate of AI answers.
Three testable takeaways from the current evidence: first, being retrievable on multiple indexes (Google and Bing) matters more than ranking #1 on either. Second, third-party corroboration — the listicles, comparison pages, and community threads models actually cite — moves recommendations more than owned content does. Third, monitor sentiment, not just mentions, because the arXiv:2606.20065 data suggests framing is the unstable variable.
FAQ
Do AI assistants have a ranking algorithm for brands?
Not in the traditional sense. There is no brand-level index. Recommendations emerge from retrieval (which pages surface) plus synthesis (what the model does with them). The controllable inputs are your presence in the underlying search indexes and the third-party pages that describe you.
Can a brand pay to be recommended?
As of this writing, no major assistant sells recommendation placement in organic answers, though OpenAI has been publicly rolling out advertising formats around ChatGPT. Watch this space — the incentive structure is obvious, and disclosure norms are unsettled.
Why does my brand appear in Gemini but not ChatGPT?
Different retrieval backbones. ChatGPT leans on Bing plus OpenAI’s own crawler; Gemini grounds on Google Search. A Semrush-scale study found only a minority of domains are cited by multiple platforms — cross-platform presence has to be earned per-index.
Last updated July 14, 2026. This page is refreshed as benchmarks and scores move.