# Why publishing a page doesn’t mean AI has read it

Research notes · By Optiview · Published October 29, 2026

Updated October 31, 2026.

Canonical article: https://optiview.ai/resources/what-openai-browser-code-taught-us/

Sources checked October 29, 2026.

We followed business information through a published AI browsing tool to see where details can get lost.

## Key takeaway

Check what an AI tool actually receives. A published page, a successful download and an AI answer are different pieces of evidence.

## What the published browsing code does

As part of its gpt-oss project, OpenAI published a simple browsing tool. It gives a model actions for searching, opening pages and finding text within them. It also includes connections to Exa and You.com, services that can supply search results and page content.

This is a working open-source example, not a disclosure of the browsing system behind production ChatGPT. Our tests use a fixed version of that example so the processing can be reproduced. They do not prove how ChatGPT, Claude, Gemini or Perplexity browse.

We can inspect what the tool actually does instead of guessing. Its [browser actions](https://github.com/openai/gpt-oss/blob/5681a6b6552635e7b29cde94a61ca6a7aee90ebd/gpt_oss/tools/simple_browser/simple_browser_tool.py), [content processing](https://github.com/openai/gpt-oss/blob/5681a6b6552635e7b29cde94a61ca6a7aee90ebd/gpt_oss/tools/simple_browser/page_contents.py) and [Exa and You.com connections](https://github.com/openai/gpt-oss/blob/5681a6b6552635e7b29cde94a61ca6a7aee90ebd/gpt_oss/tools/simple_browser/backend.py) are available in its published code.

## How information moves from a webpage to the model

Think of it as a research assistant working through text, not someone looking at the finished website design.

1. **Search for information.** The tool sends a query to a search provider. In this version, the Exa search listing uses provider-generated summaries; the You.com listing uses returned snippets. A search result is not the same as reading the whole page.
2. **Retrieve a page.** When asked to open a URL, the tool requests content from the provider. That returned content is the starting point for the next step. The code alone does not tell us how recently the provider obtained it or which website representation it used.
3. **Turn it into readable text.** The processing code converts returned HTML into text with structure and link references. The model does not need the website’s visual design to read that representation.
4. **Read a portion and explore further.** The default display budget in this implementation is 1,024 tokens—pieces of text, not a fixed number of words. The model can request another part, find a phrase or follow a link.
5. **Use the information in an answer.** The browsing tool makes information available. It does not guarantee that the model will consider it relevant, use it or cite it.

That last distinction matters. A fact can exist somewhere in the retrieved text without appearing in the first portion shown to the model. And a page can appear in search results without ever being opened.

## Where Optiview can help

Optiview works on the information a website makes available. People see the normal website. AI tools that support it can get a cleaner, easier-to-read version of the same website content, along with any extra information the business has approved.

The automatic conversion does not require a second website to maintain. The optional Narrative Graph lets a business add explanations: what it offers, what makes it different, who it serves and which limitations or responsibilities customers should understand.

For example, “we introduce customers to independent contractors” means something different from “we perform repairs.” Publishing that distinction clearly gives a system information it can potentially use to avoid a misleading description.

Optiview can also record what it served to an observed request. That verifies a delivery. It does not, by itself, verify what an intermediary passed on or which information ultimately shaped an answer.

We use OpenAI’s code as a research reference. It is not Optiview’s production delivery engine, and using it does not imply an OpenAI partnership, endorsement or certification.

## First, we checked whether the information survived processing

We took captured content from two websites and supplied it to the published text-processing code. When the supplied content included our added business context, all 13 supporting passages we had selected in advance remained in the full processed text.

None of those passages appeared in the initial viewing window. They were still accessible later. This showed why “present in the full text” and “immediately visible to the model” are different observations. We did not test whether a model would choose to navigate to those passages.

We supplied the content ourselves for this local test. That tells us what survived processing, not whether a search provider would retrieve the same version. [See the controlled study and its limitations](https://optiview.ai/proof/methodology/#browser-reference).

## Next, we checked what live retrieval services returned

We tested the pinned browser’s live Exa and You.com paths with 12 fixed requests across those two websites: four direct page requests and eight searches that did not name the target businesses. All 12 requests succeeded without retries.

All four direct requests returned page content. But none of those four responses included a complete added brand passage we were looking for. Separate checks confirmed that the current prepared website versions contained the published context.

Some related information was present in ordinary page content. We therefore do not interpret the result as “the providers understood nothing” or proof that all business meaning was lost. We also cannot tell whether caching, selection of a different version, provider extraction or another difference explains the absent complete passages.

Of the eight searches, one returned a target website among the first 20 results: Exa returned Learnings.org at position 13. That result was below the browser’s initial viewing window but accessible later. This was a search-result appearance, not an assistant recommendation or citation.

These two tests measure different steps. Their counts should not be treated as a before-and-after performance comparison. [Inspect the live-provider findings and records](https://optiview.ai/proof/methodology/#live-provider-method).

## The takeaway: publishing content is not the same as retrieving it

The prepared context survived this particular transformation when we supplied it. The four live provider responses did not supply complete tested passages to that transformation in the first place. We can locate that observable gap without claiming to know its cause.

This is why a visit from an AI-related bot is useful evidence of activity, but not proof of a citation. It is also why converting a website to Markdown cannot, on its own, guarantee inclusion in an answer. A cleaner format may make information easier to process; it cannot make another service retrieve or use it.

For Optiview, the next responsibility is clear: check what we publish, test what supported retrieval services return, and compare those observations separately with what assistants say. These findings do not establish better citation performance or justify replacing the parser or Narrative Graph.

## What we still need to learn

We have observed real AI answers using published business language. Those examples are interesting, but they do not by themselves establish that Optiview caused a recommendation or improved an answer.

We also ran a separate 72-answer comparison of different source conditions. Problems with automated scoring prevent us from treating it as reliable evidence that one condition produced better answers. We retain that study and its limitations rather than turn its scores into an improvement claim.

The bigger question remains: when approved business context demonstrably reaches a model, does it improve the answers people receive? A future controlled study could examine that. This article reports the completed work, not a new experiment or a promise of a particular result.

Businesses need more than facts scattered across the web. They need those facts and their meaning to remain clear as other systems handle them. That is the problem we are exploring with Optiview.

[Explore real AI conversations](https://optiview.ai/proof/#live-evidence), or [read the complete research methods and results](https://optiview.ai/proof/methodology/).
