The engine picks a URL, so you should too
A citation is attached to one page. Foglift's benchmark of 375 buyer-intent answers across five engines in August 2026 collected 3,412 cited URLs. Only 1.5% were homepages, and 74.7% sat at least two path segments deep. The page that gets quoted answers the specific question; it does not introduce the company.
"Improve our AI visibility" is not a task anyone can start. "Make /pricing the page an engine quotes for 'how much does X cost'" is. A marketing lead at a rewards platform put it to us in August: "tell me which URL, and what to do on that URL."
The selection rule
For each prompt you want to win, ask three questions in order. Stop at the first URL that passes all three.
- Does a page already answer this job in its H1? A page whose title is the question, not one that mentions the topic. If yes, that is the candidate. If no, you will be writing one, and the rest of this guide is the brief.
- Is the answer in the first 200 visible characters after the H1? ChatGPT's fast path indexes roughly that much body text per page. If the answer is in paragraph four, the model may never see it.
- Is there a block a model can lift? A table, a numbered list, a definition. Nectiv's November 2025 comparison found ChatGPT-cited pages were 2.3 times more likely to contain a table than Google's results for the same queries (30% versus 13%). Tables are the shape an extractable answer takes.
How to audit a candidate page in ten minutes
| Check | How | Pass |
|---|---|---|
| Visible HTML | View source, or curl the URL, and search for the first sentence of your answer | The sentence is in the raw HTML, not injected by JavaScript |
| First 200 characters | Copy the text immediately after the H1 and count | The mechanism is stated inside that span, in plain words |
| One job per page | Read the H1 and every H2 | Every heading serves the same question; nothing about a second product |
| A liftable block | Look for a table, ordered list, or definition | A model could quote it without the surrounding prose |
| Page weight | Check the response size | Well under 4 MB; ChatGPT rejects larger pages outright |
| Status and canonical | Check the response code and the canonical tag | 200, self-canonical, no noindex |
When no page passes
Write one page whose H1 is the buyer's question, whose first paragraph answers it, and which has one table or list a model could quote. Put it at a URL that describes the job. Link to it from the pages you already have on the topic. Then run the prompt weekly and watch the cited URL column.
What not to choose
- The homepage. It answers "who are you", a brand prompt. The benchmark shows engines almost never quote it for a job prompt.
- A page whose answer lives in a JavaScript widget. Calculators, tabs and accordions that render client-side are invisible to a fetcher that does not execute scripts.
- A PDF. Engines can read them but cannot link a heading inside them, and rarely prefer them over an HTML page with the same content.
- Five pages at once. You will not know which one entered the shortlist. Change one, measure, then the next.