How should I optimize my website content for GPT-5 search? Our client playbook

Clients ask us how to optimize for GPT-5 search as if it were a new algorithm with a secret recipe. In practice, ChatGPT's search feature does something familiar: it breaks a question into smaller searches, pulls pages, reads them and writes an answer with sources. The model name changes. The workflow we use does not.

We should be upfront about one thing. Nobody outside OpenAI can give you a ranking formula for ChatGPT search. What we offer is the order of operations that has made sense across client work, and a way to check results rather than trust opinion.

Start with access, not copy

The most common finding when we audit a new client is not weak content. It is a crawler that cannot get in. OpenAI documents separate bots: OAI-SearchBot for surfacing sites in ChatGPT search results, GPTBot associated with training, and ChatGPT-User for fetches triggered by a user's request. Read OpenAI's own bot documentation for the current wording before you edit robots.txt. A rule written to block training can also block search if it is too broad.

We check this in Promptwatch Agent Analytics, which shows IP-verified AI crawler logs. Spoofed user agents are filtered out, so the hits are real. We look at which bots arrive, which pages they ask for, and whether they get errors. The crawl-to-citation path then shows which fetched pages ended up cited. If a key page returns errors to OAI-SearchBot, we fix that before writing a word. Crawler logs start on the Professional plan.

Map the sub-questions

ChatGPT search tends to run several narrower queries per prompt. We start each content project by listing those sub-questions. Promptwatch tracks query fan-outs per model, and the free GEO Inspector Chrome extension shows the fan-outs ChatGPT ran and which sources were fetched versus cited. That is a better brief than a keyword list, because it reflects what the engine actually searched for.

We then check whether each sub-question has a page that answers it directly. Where it does not, that is a content gap.

Rewrite answer-first

Our rewrite habits for pages that should be cited:

• The direct answer sits in the first one or two sentences of each section.
• Headings are phrased as the questions buyers ask.
• Comparisons go in tables.
• Claims are specific and sourced. Vague copy is hard to quote.
• Structured data matches what is visible on the page. We do not add markup for things the page does not say.
• Pages render their text without needing scripts.
• Dates change when content changes, not before.

None of this is mystical. It is the clarity that helps any reader, including a model.

Look beyond your own domain

Engines cite third-party sources heavily in many categories. Where a client's brand is missing from an answer, the winning pages are often a Reddit thread, a YouTube video or a roundup on someone else's site. Promptwatch citation analytics separates page, domain, Reddit, YouTube and LinkedIn citations and lists offsite mentions, where a brand is named on a page without a link. We use that list as an outreach and participation plan. It is slow work, and it is often where the biggest gaps close.

Refresh before you create

New pages take time to be crawled and cited. Updating a page that already ranks or already gets crawled is usually faster. Promptwatch's Content Refresh suggests diff-style updates for existing pages, and content gap analysis scores coverage against prompts. For clients on Business, the Content Agent can plan and publish against the gaps on a schedule. We use the review inbox, not the hands-off mode, because our clients want a human read before anything goes live.

Measure on the engine itself

We choose 25 to 50 prompts and track them daily on ChatGPT Search. Promptwatch monitors it through the real interface, so we see what a buyer sees. We watch Visibility Score, citation rate per page, and whether movement follows our changes. Visitor analytics then shows AI-referred visits and conversions, so we can tell the client whether visibility is turning into business.

We wait two weeks before judging any change. ChatGPT answers vary run to run, and a single day tells you almost nothing.

What we tell clients not to do

Do not stuff keywords for a model. Do not mass-produce thin pages. Do not promise yourself a guaranteed placement, because no tool or tactic can deliver one. And do not change twenty things at once, or you will not know what worked.

Our order of operations

1. Confirm bots can reach the site and are not hitting errors.
2. List the sub-questions behind the top prompts.
3. Rewrite the most important pages answer-first.
4. Close the biggest citation gaps, onsite and offsite.
5. Track daily, review in two weeks, and adjust.

That is the whole playbook. It is plain, and it works because it starts with the part that is easy to overlook: whether the engine can read you at all.