How brands appear in ChatGPT responses

Tarun Kapoor Updated July 25, 2026 12 min read

Brands appear in ChatGPT responses through two independent filters: retrieval, which pulls candidate pages from a search index, and selection, which grounds the answer in a small subset of passages. Roughly 15% of retrieved pages are ever cited, and answers name only three to five brands — a shortlist decided far more by entity-level signals across the wider web than by anything on your own site. This guide explains the mechanism, the signals that drive it, and how to diagnose which stage you are failing.

The two stages: retrieval, then citation

Most explanations of AI visibility collapse a two-step process into one, which is why so much advice about it misfires. Appearing in a ChatGPT response requires clearing two independent filters, and the tactics that clear the first do almost nothing for the second.

Stage one is retrieval. When a question requires current information, the model issues searches against a web index and pulls back a set of candidate pages. This stage behaves much like classical search: your page has to be indexed, crawlable, and topically matched to a query the model actually issues. If you fail here, nothing else matters — you are not in the room.

Stage two is selection. The retrieved pages are chunked into passages, evaluated, and a small number are chosen to ground the generated answer. This stage behaves nothing like classical search. It is not ranking ten blue links; it is assembling a short, coherent, defensible answer from the smallest set of passages that will do the job.

The practical consequence is that "we rank well, so we should be cited" is a false inference. Ranking is evidence you clear stage one. It says very little about stage two, which is where most brands are quietly eliminated.

Query fan-out: one question becomes many

A single user question rarely produces a single search. The model decomposes it into multiple sub-queries — a pattern usually called query fan-out — and retrieves against each.

Someone asking "what's the best scheduling tool for a small clinic?" may generate sub-queries closer to: best appointment scheduling software small medical practice, clinic scheduling software pricing comparison, HIPAA compliant scheduling tools, patient no-show reduction software reviews. Each returns its own candidate set, and the answer is assembled across all of them.

Two consequences follow, and both cut against conventional keyword thinking:

• You cannot target the visible question. The question the user typed may never be issued as a search. You are competing for the sub-queries the model invents, which are usually more specific and more comparison-shaped than the original.
• Coverage beats optimisation. A brand that has credible content addressing pricing, compliance, comparison, and outcomes can be pulled in on several sub-queries at once. A brand with one heavily-optimised page competes on one.

This is the mechanical reason topic clusters outperform single pages in AI search. It is not that clusters are a nicer content model; it is that fan-out gives a cluster more surfaces to be caught on.

Why 85% of retrieved pages are discarded

The most useful published number on this comes from an AirOps analysis of 548,534 pages across 15,000 prompts: ChatGPT cites roughly 15% of the pages it retrieves. The other 85% are pulled in, read, evaluated, and dropped without ever appearing in the answer.

That ratio should reframe how you think about visibility work. The bottleneck for most brands is not discoverability — plenty of pages get retrieved. The bottleneck is being usable once retrieved. A page can be perfectly relevant and still be discarded because a competing passage stated the same thing more directly, more recently, or more verifiably.

It also explains a frustrating and common pattern: a brand invests in content, sees no change in AI visibility, and concludes AI search is unpredictable. In reality the pages are very likely being retrieved and then losing the selection round — an invisible failure, because nothing in your analytics records a page that was read by a model and rejected.

Brands are entities, not pages

Here is the shift that most confuses people arriving from SEO. When a model answers "what are the best tools for X," it is not ranking pages and reading brand names off the top results. It is naming three to five brands it holds as credible entities in that category, then grounding claims about them in retrieved passages.

Two distinct things are therefore being judged:

This is why a brand can hold the best-written page in its category and still not be recommended: it is winning the citability contest and losing the entity contest. It also explains the inverse — a competitor with thinner content getting named repeatedly because the wider web treats them as a real option in the category.

The shortlist framing matters too. Because answers name a handful of brands rather than a ranked list, AI visibility is closer to consideration-set membership than to ranking position. There is no position four that quietly collects traffic. You are in the answer or you are absent.

The 6.5× third-party effect

If one finding should change how budget is allocated, it is this: analysis of brand citations has found that 85% of brand mentions came from external domains, and that brands were 6.5× more likely to be mentioned through third-party sources than through their own.

The logic is not mysterious. A model assembling a commercial recommendation weights independent evidence above self-description for the same reason a buyer trusts a review over a brochure. Your own site tells the model what you claim to be. Third-party sources tell it whether that claim is corroborated.

In practice the sources that carry disproportionate weight for commercial questions are:

• Review platforms — structured, comparative, and explicitly evaluative, which is exactly the shape of a buying question.
• Community discussion — forums and communities where practitioners describe real experience, including the failure modes marketing copy omits.
• Independent comparisons and listicles — "best X for Y" content, which maps almost one-to-one onto the sub-queries fan-out generates.
• Earned editorial coverage — trade and industry press, which supplies both credibility and consistent entity description.

None of these are things you publish. All of them are things you influence. That is an uncomfortable reallocation for a content team, and it is the single largest gap between how brands spend on AI visibility and what actually moves it.

What actually influences selection

Consolidating what published analysis converges on, the factors that separate cited passages from discarded ones fall into four groups.

Notice how few of these are traditional SEO levers, and how many are either editorial or reputational. Backlinks still matter, but mainly as one input to domain credibility — not as the dominant factor they remain in classical ranking.

Why structure beats prose

Selection operates on passages, not pages. A model looking for a groundable claim wants a chunk of text that stands on its own — that answers a specific question completely without needing the three paragraphs above it for context.

That preference has a measurable effect. Analysis of citation patterns has reported that pages carrying FAQ structured data are weighted meaningfully higher in ChatGPT source selection and receive substantially more citations than equivalent plain prose. The mechanism is straightforward: a question-and-answer pair is already the shape of a groundable passage, with the question supplying explicit intent matching.

Practical patterns that improve extractability, in rough order of effect:

1. Answer in the first two sentences under each heading. Do not build to a conclusion; state it, then support it.
2. Write headings as questions or explicit claims, matching how a sub-query would be phrased.
3. Use a real FAQ section with proper markup — not a decorative accordion, but marked-up question-answer pairs.
4. Prefer tables and lists for comparative or enumerable facts. They chunk cleanly; a comparison buried in a paragraph does not.
5. Keep each section self-contained. Avoid "as discussed above," which makes a passage unusable in isolation.
6. Attach numbers and dates to claims. Specificity is both more citable and more verifiable.

Information gain: the originality filter

There is one filter that no amount of formatting survives. If your page says what the ten pages around it already say, a model assembling an answer has no reason to reach for yours — it already has that claim, grounded elsewhere, probably on a more established domain.

This concept, usually called information gain, is the most under-appreciated factor in AI visibility. It also inverts a lot of standard content practice. The common playbook — survey the top-ranking pages, cover everything they cover, add a bit more — produces content with near-zero information gain by construction. It is optimised to look comprehensive to a human editor and to be redundant to a machine.

What actually produces gain:

• Original data. Your own benchmarks, survey results, or aggregate account data — anything that exists nowhere else.
• Named specifics. Exact figures, dates, field limits, thresholds, and version numbers rather than ranges and generalities.
• Documented experience. What happened when you actually did the thing, including what failed.
• Resolved contradictions. When two credible sources disagree, explaining why is high-gain content almost nobody writes.
• Honest negatives. Who a product is wrong for, when a tactic fails. Models grounding balanced answers need this material and there is very little of it.

Note that the last two are the hardest for marketing teams to approve and the most valuable for citation. A page that says "this works well for A, poorly for B, and here is the threshold where it flips" is far more useful to an answer engine than a page that says everything is excellent.

What changes during buyer evaluation

Selection behaviour shifts depending on where the question sits in a buying process, and the shift is worth planning around.

Early, exploratory questions — "what should I consider when choosing X" — pull broad, educational sources. Category-explainer content and neutral guides do well here, and vendor content can be cited if it is genuinely instructive rather than promotional.

Evaluation-stage questions — "best X for Y," "X versus Z," "is X worth it" — behave differently. Here the model is assembling a recommendation with implicit stakes, and the weighting toward independent sources becomes much more pronounced. Reviews, comparisons, and community discussion dominate; vendor self-description is discounted precisely when the commercial value of being named is highest.

Late, specific questions — "how do I configure X," "what does X cost" — swing back toward official documentation, because the model needs authoritative, precise, current facts and the vendor is the best source for them.

The strategic read: your own content can win the early and late stages; the middle stage is won off your domain. That middle stage is where purchase decisions are actually formed, which is why brands with excellent documentation and no third-party presence see AI visibility that looks fine on informational queries and collapses on commercial ones. We cover the commercial-stage dynamics in buyer-evaluation visibility.

Diagnosing why you're absent

Absence has several distinct causes with different fixes, and treating them as one problem wastes most of the effort. Work through them in order.

The third row is the most common and the most misdiagnosed. Teams see informational citations, conclude their AI visibility is healthy, and never check the commercial queries where the money is. Measuring by query stage rather than in aggregate is what surfaces it — see how to track ChatGPT visibility.

Five myths about getting named

Myth 1: "Rank first on Google and ChatGPT will cite you." Ranking helps you get retrieved. Only about 15% of retrieved pages get cited, and the filter that decides the rest weighs credibility and extractability, not position.

Myth 2: "More content means more citations." Volume without information gain adds redundant pages. If your new page says what your existing pages and your competitors' pages already say, it contributes nothing to a selection process that is looking for the smallest sufficient set of sources.

Myth 3: "Longer pages get cited more." Selection operates on passages. A 6,000-word page competes with its own length for chunk clarity; a focused 2,000-word page that answers one question completely often extracts better. Length is neutral at best — what matters is whether any single passage is self-contained and specific.

Myth 4: "This is just SEO with a new name." Classical SEO optimises a page for a ranking position. AI visibility optimises an entity for inclusion in a shortlist, using signals that are mostly not on your website. The overlap is real but partial, and treating them as identical is why so many technically excellent sites are invisible in AI answers.

Myth 5: "You can't influence it." The mechanism is not fully documented, but the direction of the levers is well evidenced: coverage across sub-queries, extractable structure, genuine information gain, and third-party corroboration. That is a workable programme even without a published algorithm.

Frequently asked questions

Sources and further reading

• AirOps — Tracking LLM brand citations (548,534 pages across 15,000 prompts; the ~15% retrieval-to-citation rate).
• The Digital Bloom — LLM ranking factors in 2026 (entity signals, the three-to-five brand shortlist, third-party weighting).
• Context Hints — how to track ChatGPT visibility, why ChatGPT recommends your competitor, and appearing in ChatGPT answers.

Tarun Kapoor

Founder & CEO, Context Hints

Twelve years of media buying across GroupM, WPP, Ogilvy & Mather, and Neil Patel Digital. Has personally owned media for Nestlé, Sage, Qualcomm, Aetna, Weight Watchers, Chubb and Novotel.