What ChatGPT actually consults
ChatGPT, as deployed by OpenAI, answers commercial questions through a combination of model knowledge and real-time retrieval. The retrieval layer — present in the browse and search modes that have been the default for most users since late 2024 — pulls candidate sources from a search index, reranks them, and supplies a subset to the model as context. The model then synthesises an answer and, increasingly, cites the sources it found dispositive.
The implication is straightforward but consistently misunderstood. ChatGPT does not consult your website; it consults a graph of sources that contains your website only if your website has been indexed, deemed credible, and judged relevant to the question being asked. Three filters sit between any business and ChatGPT's answer.
The first filter is index inclusion. The second is authority weighting. The third is relevance to the specific prompt. A business with poor performance against any one filter is invisible regardless of its performance against the other two.
How the answer is assembled
Across the engagements the institution leads, we observe a consistent assembly pattern in ChatGPT's commercial answers. Five components recur in approximately this order.
- The base entity. ChatGPT establishes who or what is being asked about, typically anchoring to a Wikipedia article, an authoritative directory, or a regulator's register.
- The substantive claim. The model draws its dispositive content from the highest-authority sources retrieved — typically professional bodies, sector publications, and recognised editorial outlets.
- The candidate list. Where a question asks for businesses or providers, the model assembles candidates from a combination of directory-style sources and editorial mentions.
- The reasoning layer. The model overlays its own reasoning — comparison, qualification, sometimes hedging — drawn from training and from the retrieved context.
- The citation set. The model surfaces the sources it considered most load-bearing. Typically three to seven citations are shown.
A business that is missing from the candidate list is missing from the answer. A business present on the list but absent from the citation set is present but unrecommended. A business present in the citation set but contradicted by the substantive claim is present but adversely characterised. Each of these is a different problem and each is solved by a different intervention.
The most common cause of absence from ChatGPT's candidate list is not absence from the web. It is presence on the web under three different identities — Google Business Profile, professional register, trading name — that the retrieval layer cannot reconcile to a single subject. The fix is not more content. The fix is entity reconciliation.
The three moves that work
Across the institution's engagements, three moves account for the substantial majority of observable improvement in ChatGPT discoverability. They are unfashionable, slow, and largely unrelated to the activities most "AI SEO" vendors recommend.
One — entity reconciliation
The retrieval layer ChatGPT consults is an entity-resolution system before it is anything else. If the system cannot determine that "Smith & Co.," "Smith and Company Pty Ltd," and "smithco.com.au" describe the same subject, it cannot aggregate authority across them, and it cannot return the subject as the answer. Reconciliation work — name, address, phone, identifiers (ABN, registration numbers), trading variations, historical aliases — is the unglamorous prerequisite to every later move.
Two — authority graph distance
ChatGPT weights sources by approximate distance from a small set of trusted authorities. Wikipedia, recognised professional bodies, accredited registers, established editorial publications, university and government domains. A business that is one or two hops from these authorities — cited by them, listed by them, mentioned in their editorial — appears with materially higher frequency than one that is three or four hops away. The Index measures this distance directly. Closing it is the second move.
Three — passage-level citability
ChatGPT's citation set tends to favour passages that are self-contained, attributable, and directly responsive to the prompt's substance. A business whose website carries content of this kind — short, evidence-bearing, attributed to a named author with credentials visible — is more frequently cited than one whose content is long, unattributed marketing copy. The work here is editorial, not technical: writing pages that stand up under quotation.
What does not work
The institution observes a consistent set of activities that occupy budget and attention without moving discoverability. They are listed without commentary because the commentary would be unkind.
- Stuffing pages with the phrase "as recommended by ChatGPT" or analogous language.
- Building "AI-friendly" landing pages that duplicate existing pages with rephrased headings.
- Acquiring backlinks from domains the model's training data already deprecates.
- Producing high-volume, low-substance "thought leadership" content under no named author.
- Submitting content to ChatGPT through the chat interface in the belief that this trains the model.
- Buying inclusion in directories whose editorial standards do not survive disclosure.
How the institution measures it
The AI Discovery Score computes a subject's standing across the eight signals the Index measures. ChatGPT-specific behaviour is not a separate signal; it is reflected in the same signals, weighted for ChatGPT's particular source-selection logic. Subjects engaged on a ChatGPT-led brief are typically reviewed quarterly with prompt-level instrumentation: the institution observes how a stable set of representative prompts surfaces the subject across the quarter, and reports the deltas.
Where a subject moves from invisibility to consistent appearance in ChatGPT's commercial answers — the most common engagement outcome — the move is rarely the consequence of a single intervention. It is the consequence of the three moves above, conducted patiently and concurrently, over a period typically measured in quarters rather than weeks.
A closing observation
The work of becoming discoverable inside ChatGPT is the work of becoming the kind of business the open citation graph already wants to recommend. The model is downstream of the graph. The graph is not downstream of the model.
Most efforts toward ChatGPT visibility invert this relationship — they treat ChatGPT as a venue to win, rather than as a reader of a graph the business has not yet earned its place in. The institution treats the second framing as the correct one. It is also the slower one, and the one that compounds.