Answer-first content structure
Most hospitality pages open with atmosphere. The retrieval step does not store atmosphere, and the query it is matching against is full of constraints.
Short answer
Research into ChatGPT's retrieval stack found the snippet it keeps for a page is roughly 200 characters, query-independent, frozen at index time, and taken from the text immediately around the H1. The same extract comes back whether someone asked about your pricing or your location. That makes the paragraph under your heading the highest-value text you own.
Query-independent means you get one shot
Research into ChatGPT's retrieval stack found the stored snippet is roughly 200 characters, frozen at index time, and taken from the text immediately around the H1, which it captured in about 84 percent of the pages that had one.
Query-independent is the part that matters. You do not get a different extract for a different question. One stored paragraph answers every query that reaches your page, so it has to carry the facts rather than the mood.
The opening that works
Compare these two openings for the same restaurant.
Atmosphere: “A love letter to slow food and slower evenings, where the light falls just so across reclaimed oak and every plate tells a story.”
Answer-first: “A twenty-four cover neighbourhood restaurant in Fitzroy, Melbourne, serving a daily-changing seasonal menu. Open Wednesday to Sunday, dinner only. Bar seats are walk-in, tables by reservation.”
The second contains a type, a location, a cuisine, a format, a service window and a booking policy. It survives extraction into a 200-character snippet with most of its meaning intact. The first survives as nothing a query can match.
This is not an argument against good writing. Put the atmosphere in the second paragraph, where a human reading the page will still enjoy it and the retrieval step has already taken what it needs.
Constraint coverage
Conversational queries are unusually constrained. “Quiet courtyard”, “open on a Monday”, “walkable from the station”, “under two hours”, “somewhere the kids will be fine”. Every constraint you fail to state in plain words is a query family you drop out of, whatever else the page does well.
We build the constraint inventory from what customers actually ask you, not from a keyword tool, then check each one is answered somewhere a crawler will reach it. The awkward constraints matter most, because those are the ones competitors leave out.
Question headings, answered immediately
Where a page answers a real question, the question goes in the heading and the answer goes in the first sentence under it. That habit is what helps. FAQPage markup is not: Google retired FAQ rich results on 7 May 2026 and the markup does not win AI citations. Keep it if you have it. It is not the gap in your setup.
What you get
- A constraint inventory for your category, built from real customer questions and the prompts we sample.
- An answer-first rewrite of the paragraph under every H1 that matters.
- A coverage map showing which constraints are now stated, where, and which are still missing.
- Page structure work: question headings, plain-language facts, atmosphere moved below the fold.
Common questions
No. It means the first forty words have to carry facts. Everything after that can be written however you like, and should be, because a human still has to want to book. The mistake is spending the opening paragraph on mood and leaving the facts for a page nobody links to.
Only the parts that carry retrieval weight: the paragraph under each H1, the headings themselves, and wherever a constraint needs stating. Most brands keep the majority of their copy. We are not a rebrand.
That is what the research into ChatGPT’s retrieval stack found, and it is specific to that system. Other assistants have less public documentation. We treat it as the best-evidenced constraint available rather than a universal law, and writing to it costs nothing if it turns out to be more generous elsewhere.
What else an engagement covers
Want to know whether this is your problem?
Book a 30 minute call. We will fetch your site the way the crawlers do and tell you which of the prerequisites is failing. If none of them are, we will say so.