Two kinds of prompt, two retrieval paths
"What is [your brand]" and "identity security consultancy for a mid-size bank" look like neighbours. To a retrieval system they are different queries with different candidate sets. The first contains the brand name, so your domain is retrieved almost by definition. The second contains the job, so the retrieved set is whoever has a page that answers it. If that is not you, the model has your name and somebody else's work.
A marketing lead at an identity-security firm described it exactly during discovery in August 2026: "Ask about the brand and we get cited. Ask about the work we do and we don't." That is not an entity problem. The model knows the entity. It is a missing-page problem.
The diagnostic, in two rows
| Prompt type | Example | You are cited? | What it means |
|---|---|---|---|
| Brand | what is [brand]; [brand] pricing; [brand] vs [competitor] | Usually yes | The entity is known and the domain is retrievable |
| Job | tool to see why ChatGPT recommends a competitor; how to measure brand mentions | Often no | No page of yours answers the job in its H1 and first paragraph |
If the first row is yes and the second is no, every dollar spent on brand-building is going to a problem you do not have. The work is a page.
Why comparison and About pages do not fix it
A comparison page answers "X vs Y". A features page answers "what does X do". An About page answers "who is X". All three are brand-adjacent: they are retrieved when the model already has the brand in its query. None of them is retrieved for "how do I solve this problem", because none of them has that question as its title. Amplitude's product-analytics content is a useful model here: its explore pages are titled by the job (a retention-rate definition, how to choose an event-tracking approach) and the product appears after the answer, not in the H1.
What a job page is
- The H1 is the job, in the buyer's words. No product name, no year, no "best".
- The first paragraph answers it. The mechanism, in plain text, inside the first 200 visible characters. ChatGPT's fast path indexes roughly that much of the body.
- One liftable block. A table or a numbered list a model can quote without the prose around it.
- The product appears once, as the method, in a boxed aside near the end. Not in the title, not in the lead.
- A URL that says the job. /guides/how-to-measure-brand-mentions-in-chatgpt, not /blog/post-4812.
Every page in this guides section is built that way, which is deliberate. The section exists to be retrieved for job prompts the brand pages could never win.
The order of work
- List the job prompts. Sales calls are the source. Eight to twelve to begin.
- Run them, record who is cited, and open the cited pages. You are looking at the job pages you do not have.
- Write one. Ship it under a job-shaped URL, link to it from the related brand pages, and leave the brand pages alone.
- Re-run the same prompts weekly. The pass condition is your URL in the cited column on the job prompt, on repeated runs. Then write the next one.