For years, Google has guarded the workings of its search engine as if they were state secrets - and we've learned to accept this as being "normal". Sure, we can measure outputs and infer patterns, but we're rarely told how ranking decisions are made. That opacity has shaped entire industries.
OpenAI’s GPT-5 is a very different beast. GPT-5 is not a simple, ranked search index - it's an autonomous, tool-using agent. It reasons about a request, decides what outside information is needed, and then reaches for tools such as web search and site visits in much the same way a human researcher would.
Asking GPT-5 About Its Own Behavior
You can ask GPT-5 directly how it knows when to search, how it searches, when to open a page, how many results it typically considers...
I know what you're thinking. "How do we know it's not hallucinating?" It's unlikely - you're asking the question of the same model that's doing the thinking and using the tools. If its knowledge of process was that unreliable, that process wouldn't work.
You may also be thinking "It might have been told to hide it's processes." Its system prompt however, the underlying instruction set that governs its behavior, has been exposed recently by a variety of sources and it includes no such directives.
GPT-5 can, and will, describe its own process with refreshing candour. In response to my probing, GPT-5 explained that it:
- Uses search when freshness, detail, and accuracy require it.
- Opens pages when snippets from the search results are not enough to answer with confidence.
- Usually considers 5 to 10 results per query, sometimes as many as 20 for broad discovery tasks - but will query again until it's satisfied with what it has to work with
- May reformulate a user’s query into multiple related searches to ensure coverage. This rewrite is driven by what it "remembers" about the user, the ambiguity of the query, usual conceptual relationships and other factors.
This is a striking level of transparency compared to Google, where search mechanics remain hidden behind layers of secrecy and patents.
A Human-like Process
This behavior is recognisably human. If you were asked a complex question, you would:
- Run a quick search.
- Scan the top few results.
- Open one or two pages for depth.
- Reformulate your query if the first search was not good enough.
GPT-5 does the same thing, only faster and with more pages. It is not following rigid workflows or being driven by obscure algorithms. It is reasoning about when and how to use its tools.
Practical Implications
For organisations, the implications are clear:
- Structure content for snippets. If your page titles, meta descriptions, and opening paragraphs are concise and informative, GPT-5 may never need to open the page to cite you for simpler queries.
- Make your site retrievable. Clear structure, accessibility, and trustworthy signals increase the chance that an AI agent will open and use the content it can't find in its index of snippets.
- Think beyond SEO. Traditional optimisation aimed at Google rankings is no longer enough. It will only get you into the top of the reasoning funnel.
Looking Ahead
AI agents like GPT-5 are becoming the new intermediaries between businesses and customers. They are not search engines. They are tool-using agents that search, open, and reason in ways that echo human behaviour. And unlike Google, they are willing to explain how they do it.
For those of us building digital strategies, this is an opportunity to stay ahead. By understanding how AI agents use our websites, we can adapt faster than competitors who are still playing the old SEO game. OpenAI has lifted the veil, and businesses that pay attention will reap the benefits.
