# Ten questions we still need to answer about AI and websites

Research design · By Optiview · Published October 30, 2026

Updated October 31, 2026.

Canonical article: https://optiview.ai/resources/ai-readable-web-research-agenda/

Sources checked October 30, 2026.

Separate promising ideas from proven results. These are questions for future tests, not findings from studies already completed.

## Key takeaway

Turn uncertain claims into questions you can test. Keep the plans separate from completed findings, including earlier research that is already closed.

Serving Markdown, adding approved context and improving source accuracy may help software interpret a website. But each changes something different, and the outcome may depend on the source, retrieval service, question and evaluation method.

Optiview's [published methodology](https://optiview.ai/proof/methodology/) separates observed delivery, retrieval accessibility and downstream interpretation. That distinction should carry into future research.

These ten questions form a proposed research agenda. They are not newly executed tests, results or commitments to buy API access. Earlier findings keep their original scope and limitations.

## Questions 1–3: Does important information survive?

| Question | What must be observed or controlled | What remains unknown |
| --- | --- | --- |
| 1. Does equivalent Markdown preserve more answer-relevant information than HTML in a defined extraction workflow? | Freeze equivalent facts and links; compare extracted content under identical processing conditions. | Whether differences generalize across tools and page types. |
| 2. Which page structures cause material information loss during preparation? | Use controlled fixtures containing headings, tables, lists, qualifications and linked content; retain outputs. | Which observed failures occur in other conversion pipelines. |
| 3. Does adding approved contextual information improve answers to otherwise ambiguous business questions? | Hold the base source and question fixed; vary only the approved supplemental context; score repeated answers. | Whether findings transfer to external browsing and retrieval systems. |

Question 1 is a representation study. Question 3 changes available information and therefore cannot establish a format-only effect. The conditions must be reported separately.

## Questions 4–6: What reaches retrieval systems and answers?

| Question | What must be observed or controlled | What remains unknown |
| --- | --- | --- |
| 4. Under which documented conditions do retrieval clients request alternative source representations? | Use authorized request traces, current provider documentation and fixed public fixtures. | Behavior of providers whose acquisition paths are not observable. |
| 5. How quickly do different observed retrieval paths reflect an authorized source revision? | Record immutable versions, publication times, provider requests and returned content under defined conditions. | Universal refresh intervals or unobserved intermediary behavior. |
| 6. When a URL is cited, how often does the cited source support the associated claim in the observed sample? | Retain questions, answers, citation links, source captures and independent support classifications. | The complete internal source-use path or a causal relationship. |

A visible citation and verified source retrieval are different observations. Neither should be inferred automatically from the other.

## Questions 7–8: Do review and scope choices help?

| Question | What must be observed or controlled | What remains unknown |
| --- | --- | --- |
| 7. Does clearly scoped page-specific context reduce ambiguity compared with broadly applied context in a controlled evaluation? | Create approved fixtures, defined scope conditions and a fixed rubric for factual qualifications. | Whether results apply to different external systems or broader content classes. |
| 8. Do documented source-review controls reduce material publication errors in a bounded workflow? | Define error categories, compare comparable review procedures and retain approval and correction records. | Whether differences persist across organizations, workloads or time. |

The [W3C provenance overview](https://www.w3.org/TR/prov-overview/) offers concepts for documenting the origins and activities associated with information. It does not validate these hypotheses or certify any proposed governance workflow.

## Questions 9–10: Can we establish answer or business effects?

| Question | What must be observed or controlled | What remains unknown |
| --- | --- | --- |
| 9. Which source-information defects are most associated with incorrect answers in a defined question set? | Classify missing facts, contradictions, outdated terms and omitted qualifications against independently scored answers. | Whether association establishes causation across systems. |
| 10. Under what conditions can a publisher-controlled source improvement be linked to a meaningful downstream business outcome? | Define interventions, comparable observation populations, appropriate controls and authorized business-action measurements. | Whether the full attribution chain can be isolated or measured reliably. |

Question 10 is intentionally the hardest. A corrected source, a citation and a qualified business action are separate events. A credible causal conclusion requires more than observing them in chronological order.

## Set evidence standards before running a study

- Specify a narrow hypothesis before collecting observations.
- Use authorized public or explicitly approved source fixtures.
- Freeze relevant source versions and retain the transformation procedure.
- Define scoring rules and expected exclusions in advance.
- Preserve complete observations, including failures and abstentions.
- Distinguish directly observed delivery from inferred retrieval or answer use.
- Report sample sizes, denominators, controls and confounders.
- Publish null and inconclusive results without inventing a positive interpretation.

The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) provides voluntary principles relevant to trustworthy AI evaluation. These proposed study standards are independent research recommendations, not NIST certification requirements.

## Prioritize only research that can inform a decision

Prioritize studies that could change a real decision, have authorized evidence available and can reasonably isolate the variable being tested. A small, repeatable source-preservation study may be more useful than a broad experiment whose retrieval path or commercial data cannot be inspected.

The agenda does not authorize new paid experiments, provider requests, customer-data processing or the reopening of previously closed studies. Each proposed experiment would require a separate decision about scope, resources, permissions and publication.

The finished work product is a ten-question research backlog, each with a hypothesis, required observations, controls, limitations and decision owner. The objective is to make future claims more defensible—not to assume the outcome before the work begins.

Use the [reproducible retrieval study protocol](https://optiview.ai/resources/ai-retrieval-experiment-design-template/) to turn a selected question into a bounded study plan before collecting observations.
