ChatGPT decides which brands to recommend by running two systems in parallel. A training-data prior nominates a shortlist of brands the model has seen mentioned often in authoritative contexts. When the user's query benefits from current information, ChatGPT rewrites the prompt, sends it to Bing, retrieves the top results, and either confirms the prior or replaces it with brands surfaced in the retrieved pages. The shortlist that survives both passes is what the user sees. Parse tracks AI visibility across ChatGPT, Google AI Overviews, and Perplexity, which is the source of the platform-level numbers in this piece.
- ChatGPT uses two stages: a training-data prior and a Bing-fed retrieval layer. Brands need to win both to be reliably recommended.
- 87% of SearchGPT citations match Bing's top organic results for the same query (Seer Interactive). Google's index correlates at 56%. Bing is the supply chain.
- ChatGPT and Google AI Overviews disagree on brand recommendations for 61.9% of identical queries (BrightEdge). A single-platform read is a vanity metric.
- Authoritative third-party list mentions, brand search volume, and recently updated content move the needle. Backlink count and on-page keyword density barely register.
- Recently updated content gets cited more: 76.4% of ChatGPT citations come from pages updated in the last 30 days (Profound, 2025–26 datasets).
ChatGPT runs two brains, not one
ChatGPT does not have a single ranking algorithm in the way Google does. Two distinct systems decide which brand names land in an answer.
The first is the parametric model, which is the brands and entities baked into the weights during pretraining. When you ask "what is the best CRM for a 50-person sales team?" without browsing enabled, the model reaches into that prior and produces a list. There is no live lookup. The brands it picks are the ones that appeared most often in authoritative training contexts during the pretraining window, biased by reinforcement learning from human feedback toward "helpful and trustworthy" answers.
The second is the retrieval layer. ChatGPT search auto-triggers when a query likely benefits from fresh information, when the user explicitly invokes search, or when a custom GPT or agentic flow forces a tool call. OpenAI's documentation confirms ChatGPT "rewrites your query into one or more targeted queries" before handing them to a search partner.
Both systems run on most modern ChatGPT requests. The retrieval layer can confirm or override the prior. The brands that survive both stages are the brands the user sees.
How retrieval actually works inside ChatGPT search
When ChatGPT invokes search, it executes a four-step pipeline. None of the steps look like Google's blue-link ranking, and none of them are visible to the user.
Step one, query rewrite. The model paraphrases the user's natural-language prompt into one or more keyword queries. A long, conversational ask becomes two or three short search strings. OpenAI's example: "what's the latest on the development of drugs that target CCR8 for cancer?" becomes "CCR8 immunotherapy drug development 2025."
Step two, Bing fetch. Those rewritten queries hit Bing's search index. Seer Interactive analyzed over 500 SearchGPT citations and found 87% matched Bing's top organic results for the same query, with most citations coming from Bing's top 10. The same study found Google matched only 56%, with a median rank of 17. The supply chain runs through Bing.
Step three, top-k extraction. ChatGPT reads the body of the retrieved pages, not just the titles. Profound's analysis of ChatGPT search behavior found the model only cites about 15% of the pages it retrieves; the other 85% are read but not surfaced.
Step four, answer synthesis. The model assembles the answer using a mix of the retrieved evidence and its training-data prior. Brand mentions inside retrieved third-party content are weighted more heavily than brand-owned pages.
What signals push a brand into the retrieval set
Getting fetched is the prerequisite for being recommended. Three signals move that needle.
Bing rank for the rewritten query. This is the dominant signal. Authority Tech's 2026 audit showed sites that rank in Bing's top 10 for an intent-matching query appear in ChatGPT roughly 8 in 10 times for that intent, regardless of Google rank. If you optimized exclusively for Google, you have a Bing problem you cannot see in Search Console.
Indexability for crawlers OpenAI's stack reads. Two distinct user agents matter: OAI-SearchBot for ChatGPT search retrieval, and GPTBot for training. Blocking either via robots.txt removes you from a different layer of the stack. We cover the configuration in our robots.txt for AI crawlers guide.
Freshness signals. Profound's analysis of ChatGPT citations found 76.4% of cited pages were updated in the last 30 days, and pages going three months without a substantive update were three times more likely to lose visibility. Bing's index re-crawls active sites quickly; static pages decay out of the retrieval window.
The composite picture: rank in Bing for the rewritten phrasing, stay crawlable, and refresh substantively. Two of those three are owned by your engineering and editorial cadence, not your link-building budget.
How the model picks brands once retrieval lands
Retrieval surfaces a candidate set. The model still has to choose which brand names to actually mention. Three weights dominate that choice.
Authoritative list inclusion. Onely's analysis of ChatGPT recommendations found roughly 41% trace back to "best of" list articles, expert roundups, or category leader compilations on third-party sites. When a model sees a brand named alongside competitors in an editorial list, it treats that as strong evidence the brand belongs in a recommendation. A single G2 category leader badge or an industry "top 10" piece often outweighs hundreds of brand-owned pages.
Brand search volume as a prior. Wellows analyzed 7,000 LLM citations and found brand search volume correlated with citation rate at 0.334, the strongest single factor in their dataset. ChatGPT mentions brands roughly 3.2× more often than it provides clickable citations, which means in-text recognition outweighs link mechanics. Parse's data on being mentioned versus being the pick shows that being named and being recommended are separate outcomes too.
Third-party over first-party. Independent sources are cited about three times more than company-owned pages. Reddit threads, G2 reviews, expert blogs, and trade publications act as trust filters. The model reads "this brand is named in a non-promotional context" as a stronger signal than "this brand says it is the best at X."
What does not move the needle: backlink count, exact-match anchor text, on-page keyword density, or schema specifically tuned for legacy SEO. Those signals can help indirectly by improving Bing rank, but they do not enter the model's brand-selection logic.
If you want to see how AI engines describe your own brand, run a free brand check — it takes a minute.
Why ChatGPT and Google AI Overviews disagree on 62% of queries
If the brand-selection pipeline is reasonable, why do AI platforms produce such different answers? Because ChatGPT and Google use different retrieval indexes and different selection priorities. BrightEdge ran the same prompt set across ChatGPT, Google AI Overviews, and Google AI Mode and found the brands disagreed on 61.9% of queries. Only 17% of queries returned the same brand set across all three.
The behavioral split is consistent:
-
Index: Bing-fed retrieval plus parametric prior
-
Tone: brand recommender with minimal sourcing
-
Citations: 15% of retrieved pages
-
Mentions per answer: lower volume, higher trust threshold
-
Best for: action-driven queries where one or two brands suffice
-
Index: Google's web index plus knowledge graph
-
Tone: citation-first, source-heavy
-
Mentions: 2.5× more brands per query than ChatGPT
-
Coverage: wider, includes more long-tail brands
-
Best for: informational queries with multiple valid answers
-
Index: Google web index plus deeper agentic retrieval
-
Tone: selective recommender, more conservative than Overviews
-
Coverage: narrower than Overviews on commercial queries
-
Best for: research-mode queries with follow-up reasoning
The implication for an operating team: tracking brand visibility on a single platform tells you about that platform. It does not tell you about AI visibility as a category. A defensible AI visibility number is a multi-platform number, which is why we map Share of Model across ChatGPT, AI Overviews, and Perplexity rather than reporting one of the three.
The signals traditional SEO teams underestimate
A team that has been running SEO for ten years will misweight the AI visibility playbook on first contact. Three corrections matter most.
Brand search volume is now a ranking input, not just a downstream metric. Search teams treat branded search as something they earn through the rest of the program. For AI visibility, branded search is itself a feature the model uses to decide which brands to mention. Branded volume therefore becomes a primary KPI, not a vanity stat.
Third-party placements outweigh owned pages. A press feature in a tier-2 trade publication can be worth more than 50 well-optimized owned pages, because the model treats independent endorsement as a credibility signal. PR and earned media stop being soft initiatives and start carrying measurable AI weight. Adobe's October 2025 disclosure that one Acrobat content adjustment drove a 200% LLM-visibility lift on adobe.com hints at how leveraged single, well-placed pieces can be, but the heavier lift typically lives off-domain.
Bing matters again. For most B2B teams, Bing has been a rounding error. With ChatGPT search using Bing as its primary index, Bing visibility is now upstream of ChatGPT visibility. That means Bing Webmaster Tools, the IndexNow protocol, and Bing's specific ranking quirks deserve a quarter or two of focused attention.
A short diagnostic for "why does ChatGPT not recommend us?"
Before assuming the model is broken, run the four checks below in order. Most "ChatGPT does not know us" diagnoses end here.
Step 1: Bing rank for the rewritten query
Take the user query you care about, generate the two or three search-engine paraphrases ChatGPT would likely produce, and run each in Bing. If you are not on page one, the retrieval layer is not seeing you.
Step 2: Indexability for OAI-SearchBot and GPTBot
Pull your robots.txt and confirm OAI-SearchBot and GPTBot are not blocked. Many security teams blocked AI crawlers in 2024 and forgot. Reverse the policy on your priority pages.
Step 3: Authoritative third-party mentions
Search the same query and read the top five results. If your brand is not mentioned in any of them, especially in any "best X" or category-leader piece, you are missing the editorial-list signal that drives ~41% of ChatGPT recommendations.
Step 4: Freshness on the pages that mention you
Check Last-Modified or visible publish dates on the pages that mention you. If everything is older than three months, the retrieval layer is decaying you out. Either refresh those pages or earn newer placements.
If all four pass and you are still invisible, the answer is usually that brand search volume and entity recognition are too low for the parametric prior to surface you. That is solved with category-level PR, not with on-page changes.
How to track this end-to-end
The reason a real-time view matters is that all four signals shift week to week. A press feature decays. A competitor publishes a "best of" piece you are not in. Bing recrawls and your rank moves. Tracking AI visibility once a quarter is the equivalent of tracking SEO once a year, which is directionally accurate, operationally useless.
Parse measures the brands that ChatGPT, Google AI Overviews, and Perplexity surface for a fixed prompt set on a recurring cadence. The Citations tab shows which third-party sources are being pulled in for the queries that matter to your category, which is where the ~41% list-inclusion signal becomes actionable. Pair that with a weekly review cadence and the diagnostic above stops being a one-time audit and becomes an operating rhythm.
No. ChatGPT search uses Bing as its primary retrieval partner. Seer Interactive's analysis of 500+ citations found 87% matched Bing's top organic results, while only 56% matched Google's. Optimizing exclusively for Google leaves a Bing-shaped gap in your ChatGPT visibility.
Fewer than Google AI Overviews. BrightEdge found Google AI Overviews mentions roughly 2.5× more brands per query than ChatGPT. ChatGPT favors a short, high-confidence list of typically two to four brands and is more likely to consolidate around category leaders. Parse's own count of how many brands a typical AI answer names puts the shortlist at about five.
Indirectly at best. Backlinks help your Bing rank, which feeds into ChatGPT retrieval. But the model itself does not use backlink count as a brand-selection signal. Authoritative list inclusions, brand search volume, and third-party content mentions correlate far more strongly with citations than backlink profiles do.
The retrieval layer updates as fast as Bing's index, typically within days for a recrawled page. The parametric prior only changes when OpenAI ships a new model. That is why freshness signals matter on the retrieval side and why long-running brand search volume matters for the prior. The two layers move on different clocks.
Usually one of three reasons: the competitor outranks you in Bing for the rewritten query, the competitor appears in the "best of" lists ChatGPT is reading and you do not, or the competitor has higher branded search volume so the parametric prior pulls them in by default. The diagnostic in this piece narrows down which of the three is binding for your category.